Natalia Gimelshein

Looks like it's breaking on DIRECTORY, perhaps your CMake is

Preferred method to use a subset of GPUs is setting CUDA_VISIBLE_DEVICES, otherwise torch will try to create context on all the GPUs, and with memory on your "busy" GPUs already...

Not directly, but you can use torch-rnnlib for that https://github.com/facebookresearch/torch-rnnlib or look in torch-rnnlib how the passing of hidden states is done https://github.com/facebookresearch/torch-rnnlib/blob/master/rnnlib/nn/WrappedCudnnRnn.lua#L186-L225

I don't see why you would want to set it to predefined values. Remembering states between sequences might (or might not) be a reasonable thing to do, and there is...

It's a cudnn bug in cudnnGetConvolution_Algorithm that will be fixed in the next version. As a workaround, you can set cudnn.benchmark to true (that would call cudnnFind_ instead of cudnnGet*).

@jpuigcerver Just to be clear, MaxPooling is nondeterministic only when pooling stride is less than pooling window.

No, but you can repurpose average pooling for that. Pooling backward would correspond to upsampling forward, and vice versa. PRs welcome.

No, it does not. You should remove your explicit padding before passing data to batchNormalization. Try to keep batchNorm input contiguous - performance is much better this way, so this...

Can you try applying the following patch and see if it fixes your problem. +cc @SeanNaren ``` diff --git a/test/test_rnn.lua b/test/test_rnn.lua index e7ee3de..d8e83e6 100644 --- a/test/test_rnn.lua +++ b/test/test_rnn.lua @@ -233,7...

@SeanNaren, please take a look at the patch I posted, the cast that is there now is wrong. On some OSs it still works fine, as your and mine examples...