Maciej Korzepa

Sure. https://drive.google.com/drive/folders/0BwyHNzMwVsM-SXdpdlduTlVTT2c?usp=sharing WER should be around 12%.

I didn't make a single change in the architecture, so I am using the default sizes.

Actually, I changed some parameters and code. I used for training everything apart from test-clean and test-other. 12% error rate was achieved on test-clean. I got -inf costs almost in...

@SeanNaren Sure, go ahead! I think I might be able to use 4xK80 for the training soon, I might then try reducing the batch size and see if WER can...

@nn-learner Maybe your input spectrograms were processed with different parameters? I used: -windowSize 0.02 -stride 0.01 -sampleRate 16000 -processes 8 -audioExtension flac I actually didn't manage to run Test.lua due...

@suhaspillai To be honest, I set it to 100 after reading posts from issue #51 and I haven't tried other values since then.

I can confirm the same problem in the latest release 1.5.1. For now my workaround is delegating the message to the topic-filter-specific callback within the on_message callback that receives the...

Thanks for the quick reply! Here's a gist to reproduce the issue: https://gist.github.com/maciejkorzepa/23bbb4443c2fcb93927eb36d9ef2091c I can also share the PyTorch code if that would help but I would need to edit...

Here's PyTorch code that computes the kernel for 80 examples below 1 second on 12GB Titan X GPU: https://gist.github.com/maciejkorzepa/af3b5ef3676d982f8c4298f224960c53

I found out some new things: 1) non-batched version works very fast as long as it fits in the memory (and the input has always the same shape) 2) batched_version...