To install MatConvNet on Windows

1. Visual Studio
   • Download and install Visual Studio Community 2013 (under Visual Studio 2013 in left panel), better use .iso version.
2. CUDA
   • Download and install Cuda Toolkit 7.5
   • To test, compile Cuda samples in Visual Studio (build output for me: 135 succeeded, 10 failed, 0 up-to-date, 0 skipped), and run compiled deviceQuery.exe
     (Possible path: C:\ProgramData\NVIDIA Corporation\CUDA Samples)
3. MATLAB
   • Configure mex (“mex -setup c++”, “mex -setup c”) for VS2013
4. cuDNN rc4
   • Download cuDNN rc4 and extract somewhere (need an Nvidia username).
   • Add to “windows PATH” the full path to extracted \bin directory.
   • If MATLAB is open, close it and re-open before continuing.
5. MatConvNet
   • Download and extract MatConvNet
   • Test with the quickstart script
   • Test MatConvNet with GPU support, e.g.:
     • vl_compilenn(‘enableGpu’, true, …
       ‘cudaMethod’, ‘nvcc’, …
       ‘cudaRoot’, <ROOT OF INSTALLED CUDA 7.5>, …
       ‘enableCudnn’, true, …
       ‘cudnnRoot’, <DIR OF EXTRACTED CUDNN rc4>) ;