The original great cuda-convnet package is here: http://code.google.com/p/cuda-convnet/
To build the convnet using CUDA 5:
build.sh to use correct pathes. For cuda SDK path, use the path contains file helper_cuda.h.dummyinclude folder, and put the cutil_inline.h into that folder.Try ./build.sh and see whether that can compiles.
The functionality is not fully tested and this can just compiles the package.
| /* | |
| Dummy Includes for cutil_inline.h | |
| Wei LI <kuantkid <at> gmail <dot> com> | |
| */ | |
| #include <helper_cuda.h> | |
| #define cutilCheckMsg(a) getLastCudaError(a) | |
| #define cutGetMaxGflopsDeviceId() gpuGetMaxGflopsDeviceId() | |
| #define MIN(a,b) (a) < (b) ? (a) : (b) |
| # Change the include line in Makefile | |
| INCLUDES := -DCUDA_5 -I$(PYTHON_INCLUDE_PATH) -I$(NUMPY_INCLUDE_PATH) -I$(CUDA_SDK_PATH) -I./include -I./include/common -I./include/cudaconv2 -I./include/nvmatrix -I./dummyinclude/ |
There is some further comments in email from David Reichert and hope this will benefit others:
Thanks a lot! It looks like you also added this to the Makefile
-LGLEW
as well as commented out the lines below makedirectories, which is what caused me trouble
(I don't really know how makefiles work in detail). Maybe these details could be added to
your explanation. Also, I also had to comment out those cutil lines to make it work. It does
seem to work now though, so thanks!
In addition to this I also followed instructions from here to modify common-gcc-cuda-4.0.mk
Thank you very much！
This is great, but can anyone verify that it is in fact functional? I am unable to recreate the sample specified here: https://code.google.com/p/cuda-convnet/wiki/TrainingNet with no error output...
did you get it installed with cuda 6 on osx?
This is awesome, thank you! I also had to comment out line 332 in common-gcc-cuda-4.0.mk to stop it from linking to cutil and shrutil. That made it compile; I don't know if it actually works yet.