Introduction

It seems these instruction are deprecated as the preferred method to install Theano is using Anaconda, as described on the page Converting to the new gpu back end(gpuarray).

In a previous article, I described a procedure to install Keras and Theano on Windows 10. The main downside of this procedure is that you had to downgrade to Python 3.4. This tutorial is an improved version which allows you to make Theano and Keras work with Python 3.5 on Windows 10.

Unless stated otherwise, all install options are left to their default/recommended values. In case you change installation paths, make sure you also change them in all subsequent commands.

Installing CUDA

This step is optional but highly recommended for NVidia card owners as it massively speeds up Theano by allowing it to use the GPU:

1. Download and install Visual Studio Community 2013 update 5. This is a direct link which may not work in the future as your are now required to create a Microsoft account to download old versions of Visual Studio. Do not install Visual Studio 2015 as latest versions are not compatible with CUDA 8. Open VS2013 to see if it’s working.
2. Add C:\Program Files (x86)\Microsoft Visual Studio 12.0\VC\bin to the PATH.
3. Download and install CUDA 8.0, you have to create a free NVidia account to download the file. Open the samples directory C:\ProgramData\NVIDIA Corporation\CUDA Samples\v8.0 and then the Samples_vs2013.sln file. Run any project to see if CUDA is working, for instance, 5_Simulations/oceanFFT. Right click on it, then click on Debug>Start new instance. Accept to build the project, then you should see a simulated ocean surface.
4. The CUDA installation may have downgraded your NVidia drivers, you can upgrade them again.

Installing Theano

1. Download and install Anaconda 64 bits with Python 3.5 for Windows. Open a command line and check that Python is working and that it is indeed the Anaconda python 3.5 version.
2. Download and extract MinGW x86_64-5.4.0-release-posix-seh-rt_v5-rev0.7z. Make sure you’re not using a newer version as it won’t work with Theano.
3. Add MinGW bin directory to the PATH. For instance, if you extracted MinGW on C: , then add C:\x86_64-5.4.0-release-posix-seh-rt_v5-rev0\mingw64\bin to the PATH.
4. Open a new command prompt (to take into account the modification to the PATH) and type:
   
   MS DOS

   cd %USERPROFILE%\Anaconda3\libs gendef ..\python35.dll dlltool --as-flags=--64 -m i386:x86-64 -k --output-lib libpython35.a --input-def python35.def del python35.def

   1 2 3 4

   cd %USERPROFILE%\Anaconda3\libs gendef ..\python35.dll dlltool --as-flags=--64 -m i386:x86-64 -k --output-lib libpython35.a --input-def python35.def del python35.def

5. Download and install Git for Windows.
6. In the command prompt type:
   
   MS DOS

   cd %USERPROFILE%\Anaconda3\Lib\site-packages git clone git://github.com/Theano/Theano.git cd Theano python setup.py develop cd %USERPROFILE% type NUL > .theanorc

   1 2 3 4 5 6

   cd %USERPROFILE%\Anaconda3\Lib\site-packages git clone git://github.com/Theano/Theano.git cd Theano python setup.py develop cd %USERPROFILE% type NUL > .theanorc

7. Open the explorer and type in the address bar %USERPROFILE%, open the .theanorc file and add the following lines:
   
   .theanorc

   [global] floatX = float32 device = gpu [nvcc] compiler_bindir=C:\Program Files (x86)\Microsoft Visual Studio 12.0\VC\bin

   1 2 3 4 5 6

   [global] floatX = float32 device = gpu   [nvcc] compiler_bindir=C:\Program Files (x86)\Microsoft Visual Studio 12.0\VC\bin

   
   I assume you installed CUDA, if it is not the case, you have to replace device = gpu by device = cpu . If you have more than one GPU, you can also specify which one you want to use using gpu1 or gpu2 etc.
8. If your Windows username contains a space or any unusual character or if you want Theano to put its temporary files to a specific place (they can use several Gb of disk space) then add base_compiledir=C:\theano_compiledir under the [global] section of your .theanorc file. Your .theanorc file should now contain:
   
   .theanorc

   [global] floatX = float32 device = gpu base_compiledir=C:\theano_compiledir [nvcc] compiler_bindir=C:\Program Files (x86)\Microsoft Visual Studio 12.0\VC\bin

   1 2 3 4 5 6 7

   [global] floatX = float32 device = gpu base_compiledir=C:\theano_compiledir   [nvcc] compiler_bindir=C:\Program Files (x86)\Microsoft Visual Studio 12.0\VC\bin

9. Now it s time to test if Theano is working properly. Copy the following code into a file and run it in the command line.
   
   testTheano.py

   Python

   from theano import function, config, shared, sandbox import theano.tensor as T import numpy import time vlen = 10 * 30 * 768 # 10 x #cores x # threads per core iters = 1000 rng = numpy.random.RandomState(22) x = shared(numpy.asarray(rng.rand(vlen), config.floatX)) f = function([], T.exp(x)) print(f.maker.fgraph.toposort()) t0 = time.time() for i in range(iters): r = f() t1 = time.time() print("Looping %d times took %f seconds" % (iters, t1 - t0)) print("Result is %s" % (r,)) if numpy.any([isinstance(x.op, T.Elemwise) for x in f.maker.fgraph.toposort()]): print('Used the cpu') else: print('Used the gpu')

   1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22

   from theano import function, config, shared, sandbox import theano.tensor as T import numpy import time   vlen = 10 * 30 * 768  # 10 x #cores x # threads per core iters = 1000   rng = numpy.random.RandomState(22) x = shared(numpy.asarray(rng.rand(vlen), config.floatX)) f = function([], T.exp(x)) print(f.maker.fgraph.toposort()) t0 = time.time() for i in range(iters):     r = f() t1 = time.time() print("Looping %d times took %f seconds" % (iters, t1 - t0)) print("Result is %s" % (r,)) if numpy.any([isinstance(x.op, T.Elemwise) for x in f.maker.fgraph.toposort()]):     print('Used the cpu') else:     print('Used the gpu')

   
   If everything s OK, this code will display something like that:
   
   Using gpu device 0: GeForce GTX 960 (CNMeM is disabled, CuDNN not available) [GpuElemwise{exp,no_inplace}(<CudaNdarrayType(float32, vector)>), HostFromGpu(GpuElemwise{exp,no_inplace}.0)] Looping 1000 times took 0.5765759944915771 seconds Result is [ 1.23178029 1.61879349 1.52278066 ..., 2.20771813 2.29967761 1.62323296] Used the gpu

   1 2 3 4 5 6

   Using gpu device 0: GeForce GTX 960 (CNMeM is disabled, CuDNN not available) [GpuElemwise{exp,no_inplace}(<CudaNdarrayType(float32, vector)>), HostFromGpu(GpuElemwise{exp,no_inplace}.0)] Looping 1000 times took 0.5765759944915771 seconds Result is [ 1.23178029 1.61879349 1.52278066 ..., 2.20771813 2.29967761 1.62323296] Used the gpu

Installing Keras

1. In a command prompt, type:
   
   MS DOS

   cd "%USERPROFILE%\Anaconda3\Lib\site-packages" git clone git://github.com/fchollet/keras.git cd keras python setup.py develop

   1 2 3 4

   cd "%USERPROFILE%\Anaconda3\Lib\site-packages" git clone git://github.com/fchollet/keras.git cd keras python setup.py develop

2. As Keras uses TensorFlow as the default backend, we have to change that for Theano. Open a command line and type:
   
   MS DOS

   cd %USERPROFILE% md .keras

   1 2	cd %USERPROFILE% md .keras

3. Create a file named keras.json in %USERPROFILE%/.keras and add the following lines:
   
   keras.json

   { "floatx": "float32", "epsilon": 1e-07, "image_dim_ordering": "th", "backend": "theano" }

   1 2 3 4 5 6

   {     "floatx": "float32",     "epsilon": 1e-07,     "image_dim_ordering": "th",     "backend": "theano" }

4. Now it s time to test if Keras is working properly. Copy the following code into a file and run it in the command line.
   
   testKeras.py

   Python

   from keras.models import Sequential from keras.layers import Dense, Dropout, Activation from keras.optimizers import SGD import numpy as np data_dim = 20 nb_classes = 4 model = Sequential() # Dense(64) is a fully-connected layer with 64 hidden units. # in the first layer, you must specify the expected input data shape: # here, 20-dimensional vectors. model.add(Dense(64, input_dim=data_dim, init='uniform')) model.add(Activation('tanh')) model.add(Dropout(0.5)) model.add(Dense(64, init='uniform')) model.add(Activation('tanh')) model.add(Dropout(0.5)) model.add(Dense(nb_classes, init='uniform')) model.add(Activation('softmax')) model.compile(loss='categorical_crossentropy', optimizer='sgd', metrics=["accuracy"]) # generate dummy training data x_train = np.random.random((1000, data_dim)) y_train = np.random.random((1000, nb_classes)) # generate dummy test data x_test = np.random.random((100, data_dim)) y_test = np.random.random((100, nb_classes)) model.fit(x_train, y_train, nb_epoch=20, batch_size=16) score = model.evaluate(x_test, y_test, batch_size=16)

   1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39

   from keras.models import Sequential from keras.layers import Dense, Dropout, Activation from keras.optimizers import SGD import numpy as np   data_dim = 20 nb_classes = 4   model = Sequential()   # Dense(64) is a fully-connected layer with 64 hidden units. # in the first layer, you must specify the expected input data shape: # here, 20-dimensional vectors. model.add(Dense(64, input_dim=data_dim, init='uniform')) model.add(Activation('tanh')) model.add(Dropout(0.5)) model.add(Dense(64, init='uniform')) model.add(Activation('tanh')) model.add(Dropout(0.5)) model.add(Dense(nb_classes, init='uniform')) model.add(Activation('softmax'))   model.compile(loss='categorical_crossentropy',               optimizer='sgd',                 metrics=["accuracy"])   # generate dummy training data x_train = np.random.random((1000, data_dim)) y_train = np.random.random((1000, nb_classes))   # generate dummy test data x_test = np.random.random((100, data_dim)) y_test = np.random.random((100, nb_classes))   model.fit(x_train, y_train,           nb_epoch=20,           batch_size=16)   score = model.evaluate(x_test, y_test, batch_size=16)

   
   If everything is working properly you should see (it can take several minutes on the first run):
   
   Using Theano backend. Using gpu device 0: GeForce GTX 960 (CNMeM is disabled, CuDNN not available) [Lots of Epoch N/20 with loss and accuracy measures]

   1 2 3 4

   Using Theano backend. Using gpu device 0: GeForce GTX 960 (CNMeM is disabled, CuDNN not available)   [Lots of Epoch N/20 with loss and accuracy measures]

5. To be able to draw graph representation of models, there are two more dependencies to install. In the command prompt, type:
   
   MS DOS

   conda install graphviz pip install git+https://github.com/nlhepler/pydot.git

   1 2	conda install graphviz pip install git+https://github.com/nlhepler/pydot.git

6. Pydot may have trouble finding the path to GraphViz. Open the file __init__.py in %USERPROFILE%\Anaconda3\Lib\site-packages\pydot and add
   
   Python

   elif os.path.exists(os.path.join(path, prg + '.bat')): if was_quoted: progs[prg] = '"' + os.path.join(path, prg + '.bat') + '"' else: progs[prg] = os.path.join(path, prg + '.bat') success = True

   1 2 3 4 5 6 7

   elif os.path.exists(os.path.join(path, prg + '.bat')):                 if was_quoted:                     progs[prg] = '"' + os.path.join(path, prg + '.bat') + '"'                 else:                     progs[prg] = os.path.join(path, prg + '.bat')                   success = True

   
   just after
   
   Python

   elif os.path.exists(os.path.join(path, prg + '.exe')): if was_quoted: progs[prg] = '"' + os.path.join(path, prg + '.exe') + '"' else: progs[prg] = os.path.join(path, prg + '.exe') success = True

   1 2 3 4 5 6 7

   elif os.path.exists(os.path.join(path, prg + '.exe')):                 if was_quoted:                     progs[prg] = '"' + os.path.join(path, prg + '.exe') + '"'                 else:                     progs[prg] = os.path.join(path, prg + '.exe')                   success = True

Troubleshooting

If you encounter any error message during the installation and tests, check that your PATH is clean. Open a new command prompt and type:

1. This command should return the path where you installed Anaconda:
   
   MS DOS

   where python

   1	where python

2. All these command should return the path where you extracted MinGW:
   
   MS DOS

   where gcc where g++ where gendef where dlltool

   1 2 3 4

   where gcc where g++ where gendef where dlltool

3. This command should return the path of the bin folder of your VS2013 installation:
   
   MS DOS

   where cl

   1

   where cl

4. This command should return the path where you installed CUDA (if you use it):
   
   where nvcc

   1	where nvcc

5. Older versions of Theano were not supposed to work with Python 3.5. If you encounter any problem, you have to bypass this limitation (without any know problem) by editing the file __init__.py located in %USERPROFILE%\Anaconda3\lib\site-packages\theano\theano\ and delete the lines:
   
   To be deleted

   Python

   if sys.platform == 'win32' and sys.version_info[0:2] == (3, 5): raise RuntimeError( "Theano do not support Python 3.5 on Windows. Use Python 2.7 or 3.4.")

   1 2 3

   if sys.platform == 'win32' and sys.version_info[0:2] == (3, 5):     raise RuntimeError(         "Theano do not support Python 3.5 on Windows. Use Python 2.7 or 3.4.")

Conclusion

Now you should be ready to do amazing stuff with Keras and Theano. If you want to make your networks run faster, you should check my article on how to use CuDNN and CNMeM with Theano.

References

• Theano on Windows with Python 3.5 probably don’t work
• Installing Theano on Windows 10 using Python 3.5