Anders Boesen Lindbo Larsen

Ah, great job finding it finally! I admire your persistence. :) I will try to look into `kernel_win2img()` at a later point. It seems like this is the problem.

If you install with `sudo` you need to export the environment variables as well. This is done with `sudo -E`.

@mirzman: what is the output of `ldd `? I think you need to update the environment variable `LD_LIBRARY_PATH` to point to the correct libraries.

I would prefer not to check array dimensionality in convnet layers for every call to fprop(). How about raising an error in _setup() if the array is not 4D?

Hey, thank you for the proposal! I have intentionally left out `cudarray.nnet.tanh()` because NumPy exposes it at the root level. Therefore, you can find it as `cudarray.tanh()`. The implementation is...

Hey, I have just pushed a commit. In this new version of CUDArray you can check the backend by printing `cudarray._backend`. If this variable is `'numpy'` instead of `'cuda'`, something...

Good question! FP16 is not supported at the moment. I expect it would be fairly straightforward (but tedious) to add FP16 to all the CUDA kernels. Maybe it's easier to...

I think all kernels have types as template parameters. Thus, you can add float16 support by instantiating template functions with float16. Alternatively you can try to change all occurences of...

For fully-connected architectures, you should make sure that MKL parallelizes the matrix multiplications across multiple cores. Maybe it isn't configured correctly? For convolutional architectures you are out of luck. The...

The CPU implementation of the convolution operation is too slow for any practical use. You would have to implement something faster first.