If you are the one who struggle to install Python Theano (tensor based numerical computation library) with CUDA-based computation capability on modern Microsoft Windows operating system, this very article might helped you. Here is some steps I already tried in my freshly installed Windows 10 x64. The procedure here are based from this poorly written documentation.
First, I would assume you already install the Python library itself. To make it more convinient, I recommend to install pre-packaged Python library which already included some basic necesities (numpy, scipy, etc) like Anaconda. The next procedure will assume you already installed the Anaconda.
The second thing you need to install are:
- LibPython MinGW library (use this command: conda install mingw libpython),
- Visual C++ no later than version 2013 (I would recommend instaling Visual Studio Express 2012 which still can be grabbed freely)
CUDA toolkit from nVIDIA. In my case I use version 7.5.
After all of three software installed, then create a new file on
%USERPROFILE%/.theanorc.txt with content as explained in http://deeplearning.net/software/theano/install_windows.html#configure-theano-for-gpu-use (change the –cl-version according to what version of Visual C++ you already installed).
The last step before compile a Theano code is to include “cl.exe” into your PATH environment system variable by adding “C:\Program Files (x86)\Microsoft Visual Studio 11.0\VC\bin\x86_amd64″ in it (Google for how to do this).
To compile a Thenao code, make sure you run
c:\Program Files (x86)\Microsoft Visual Studio 11.0\VC\vcvarsall.bat x86_amd64 script PRIOR to opening Python intrepreter. For the testing purpose, you can try a code given from the original tutorial: http://deeplearning.net/software/theano/install_windows.html#configuring-theano.
Hope this article will help you. If any case of error still ocurring after following those step above, please read carefully the error given by C++ compiler during code compilation process (The way Theano works is by compiling the Python code into C++ before doing the calculation).