Hello, I have largely gotten my WinPython/Theano/CUDA orchestra running, I think, but I ran into a problem when I tried to compile the theanets mnist example at http://theanets.readthedocs.io/en/stable/examples/mnist-classifier.html.
First, let me indicate what works correctly. The check_blas.py test runs correctly and is linking against the gpu. The specific test program on http://deeplearning.net/software/theano/install_windows.html to determine whether the GPU is being used runs without problems and indicates the GPU is running. I have also been able to run the lasagne mnist program [https://github.com/Lasagne/Lasagne/blob/master/examples/mnist.py] on the GPU. So I think the above means that lots of things are working properly... ...but when I tried to run the theanets mnist example at http://theanets.readthedocs.io/en/stable/examples/mnist-classifier.html or one of the other theanets examples I get errors being thrown by File "C:\Winpython\WinPython-64bit-3.4.4.2\python-3.4.4.amd64\lib\site-packages\theano-0.7.0-py3.4.egg\theano\sandbox\cuda\nvcc_compiler.py" [Note, I've tried the bleeding edge version as well, same errors.] At first I got the error that crtdefs.h was missing. I fixed this by modifying my INCLUDE statement on my nvcc.profile to include C:\Program Files (x86)\Microsoft Visual Studio 12.0\VC\include Then I got an error saying that basetsd.h was missing, so I added C:\Program Files (x86)\Microsoft SDKs\Windows\v7.1A\Include to the include file as well. Now all hell breaks loose when I try to run the program. First I get some warnings like: C:\Winpython\WinPython-64bit-3.4.4.2\python-3.4.4.amd64\lib\site-packages\theano-0.7.0-py3.4.egg\theano\sandbox\cuda\cuda_ndarray.cuh(17) : warning C4005: 'PyString_Check' : macro redefinition C:\Winpython\WinPython-64bit-3.4.4.2\python-3.4.4.amd64\lib\site-packages\numpy\core\include\numpy/npy_3kcompat.h(63) : see previous definition of 'PyString_Check' Later I see this: C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v7.5\include\crtdefs.h(496): error: invalid redeclaration of type name "size_t" And then there are a bunch of errors all related to size_t. The base call in the traceback is File "C:\Winpython\WinPython-64bit-3.4.4.2\python-3.4.4.amd64\lib\site-packages\theano-0.7.0-py3.4.egg\theano\sandbox\cuda\nvcc_compiler.py", line 423, in compile_str 'for cmd', ' '.join(cmd)) This appears to have something to do with unsigned integers. And I get this note: Exception: ('The following error happened while compiling the node', Shape_i{1}(out.w), '\n', 'nvcc return status', 2, 'for cmd', 'nvcc -shared -O3 --use-local-env --cl-version=2013 -arch=sm_52 -Xlinker /DEBUG -D HAVE_ROUND -m64 -Xcompiler -DCUDA_NDARRAY_CUH=m11b90075e2397c684f9dc0f7276eab8f,-D NPY_NO_DEPRECATED_API=NPY_1_7_API_VERSION,/Zi,/MD -IC:\\Users\\David\\AppData\\Local\\Theano\\compiledir_Windows-10-10.0.10586-Intel64_Family_6_Model_94_Stepping_3_GenuineIntel-3.4.4-64\\cuda_ndarray -INone\\include -IC:\\Winpython\\WinPython-64bit-3.4.4.2\\python-3.4.4.amd64\\lib\\site-packages\\numpy\\core\\include -IC:\\Winpython\\WinPython-64bit-3.4.4.2\\python-3.4.4.amd64\\include -IC:\\Winpython\\WinPython-64bit-3.4.4.2\\python-3.4.4.amd64\\lib\\site-packages\\theano-0.7.0-py3.4.egg\\theano\\sandbox\\cuda -o C:\\Users\\David\\AppData\\Local\\Theano\\compiledir_Windows-10-10.0.10586-Intel64_Family_6_Model_94_Stepping_3_GenuineIntel-3.4.4-64\\tmp4rrj8mgl\\me2fb8274f75ca25753e818dfaeeb725b.pyd mod.cu -LC:\\Winpython\\WinPython-64bit-3.4.4.2\\python-3.4.4.amd64\\libs -LC:\\Users\\David\\AppData\\Local\\Theano\\compiledir_Windows-10-10.0.10586-Intel64_Family_6_Model_94_Stepping_3_GenuineIntel-3.4.4-64\\cuda_ndarray -LC:\\Winpython\\WinPython-64bit-3.4.4.2\\python-3.4.4.amd64 -lpython34 -lcudart -lcublas -lcuda_ndarray', '[Shape_i{1}(out.w)]') Any help is appreciated! My theanorc file is [global] device = gpu floatX = float32 [nvcc] flags = --use-local-env --cl-version=2013 -- --- You received this message because you are subscribed to the Google Groups "theano-users" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. For more options, visit https://groups.google.com/d/optout.
