pei2748 / linear-svm-with-pycuda Goto Github PK

Date:
2018-06-01

Author:
WRF

Description:
This project demonstrates how PyCUDA can be used to build a linear support vector machine using the Huberized squared hinge loss.

Setup (other than installing packages denoted in the scripts):
- Access to GPU
- Install NVIDIA GPU Computing Toolkit with CUDA v9.2: 
	https://developer.nvidia.com/cuda-downloads
- Install MS Visual Studio 2017 Community version (v14.14.26428 worked for me):
	https://www.visualstudio.com/downloads/

Notes: You may need to do the following:
1) In the file host_config.h found here,
	C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v9.2\include\crt
change line 131 from:
	#if _MSC_VER < 1600 || _MSC_VER > 1913
to:
	#if _MSC_VER < 1600	
This prevents an error being thrown by PyCUDA. Although this has let me use PyCUDA, I do not know it's other ramifications!

2) In %PYTHON%\Lib\site-packages\skcuda\cublas.py, you may need to manually set the location of the relevant cublas.dll file. E.g.,

	dirname_DLL = 'C:\\Anaconda3\\DLLs'
	_libcublas = None
	for _libcublas_libname in _libcublas_libname_list:
	    try:
		if sys.platform == 'win32':
		    _libcublas = ctypes.windll.LoadLibrary(dirname_DLL + "\\" +_libcublas_libname)

-------------------------------------------------------------------------------------------------------------------

DISCLAIMER: 
For anyone that wants to use this code, be aware that it's limited to the following:
i) Subsetted models in one-vs-one fitting cannot be larger than 1024 records (i.e., to fit on a GPU block). 
E.g., your training design matrix could have shape (10M, 100), so long as each one-vs-one label pair has at most 1024 observations.
ii) A small number of response categories, because of an unknown memory issue. In other words, 
too many response categories will likely eat up your RAM and crash your python kernel.

I strongly suspect these issues can be resolved, but have been unable to do so with my limited python/GPU experience. 
If anyone has any recommendations, feel free to post them!