Install and Run cuda-convnet on Windows OS with VS2010

Xiong Yuanjun

MMLAB, CUHK

Prerequisite

CUDA and a supported NVIDIA GPU. At the time of writing this doc, the CUDA version is 5.5. All following instructions will be based on CUDA 5.5.
Download cuda-convnet. The project site is on the github. To get the code, you can both download the zip file, or clone from the github using cmd

git clone https://github.com/dnouri/cuda-convnet.git
Python 2.7 and Numpy, I suggest using WinPython on Windows OS. Why? WinPython contains a python console and numpy-mkl module. Besides, it also embeds many useful packages and a scientific computation friendly Python IDE - Spyder. The installation experience is far from painful compared to install python and all packages related. Remember to install the 64-bit version.
PThread, for the threading support.
Intel C++ Composer for its intel mkl module.
The Visual Studio Project, kindly provided by Oriol Vinyals.
Other useful tool: Dependency Walker to look into dll problems.

Put the content of the visual studio project into the source folder of cuda-convnet. However, you may not be able to open the .sln solution file in VS2010. This is resulted from default directories setting.
Open pyconvnet.vcxproj with any text editor or xml parser. Since our main execution platform is on a 64-bit windows OS. We only need to modify setting of the profile <release|x64>.
- Search for this profile label, and modify the directory setting to the actual path on your machine. In my experience, labels needing attention include <AdditionalDependencies>,<AdditionalLibraryDirectories>
- Modify the CUDA version identifier. Search for CUDA 4.0 (the original code was written with cuda 4.0, however we will use cuda 5.5) and modify any occurrence to CUDA 5.5.
- The dll and lib files may change their names over versions. First look into the depencies and modify the project profile to adapt to these changes.
Now you should be able to open the solution in VS2010.

Rebuild the solution in VS2010. You may come across some errors. Following are some of them, for more help, you can browse the github homepage and its wiki.
- First set USE_MKL in preprocessor difinition. This enables the mkl support.
- The compiler may not find the cutil_inline.h. Actually it's missing in newer version of CUDA. Refer to this gist and add the file to the projects' Headers folder.
- Compilation error: pthread_muted_t undefined. Try add #include <pthread.h> in nvmatrix.cuh.
After a successful compilation, you will get a .pyd library called pyconvnet.pyd. This is the python module of cuda-convnet.

Download the cifar-10 dataset. Put the content into ($ProjectDir$)/storage2/tiny/cifar-10-batches-py-colmajor/. The zip file should contains 6 data files and 1 meta-file.
Set PATH in environment variable. You need these 3 dirs.
- (pthread directory)\Pre-built.2\dll\x64
- (Intel composer directory)\redist\intel64\mkl
- (Intel composer directory)\redist\ia32\mkl
However, you may still get dll loading fail. So we can directly copy the related dlls into the project folder (not recommended). As unfortunate as me, I copied mkl_intel_thread.dll, pthreadGC2.dll and pthreadVC2.dll.
Now let's run the code, type in the command line python convnet.py --data-path=./storage2/tiny/cifar-10-batches-py-colmajor/ --save-path=./storage2/tmp --test-range=6 --train-range=1-5 --layer-def=./example-layers/layers-19pct.cfg --layer-params=./example-layers/layer-params-19pct.cfg --data-provider=cifar --test-freq=13 --epoch=200
This commnd will run the convnet for 200 iterations and get you a learned classifier.

More info coming soon...