lockie / cnn-performance Goto Github PK

This is but a simple bunch of code to test performance of widely known convolutional neural networks on several publicly available image datasets.

Just build cuda-convnet included as git submodule & run run.sh. Dont forget to fix its CMakeLists.txt:52 by uncommenting arch=compute_20,code=sm_21 for older video cards (like GTX 450 of mine 😒) & to pass something like

-DBLAS_LIBRARIES=/usr/lib/atlas-base/libcblas.so

to cmake when building.

Also, several additional modifications to python code of cuda-convnet would be required. I'll describe them later.