Improvements

[ChrisChiasson]

I noticed the "Wavenet" part of the code from Graph-Wavenet has been updated in a different repo so that it only refers to Conv1d instead of 2D convolutions. By itself, that would only make the code cleaner, but switching to Conv1d helps with another issue below.

https://github.com/f90/Seq-U-Net/blob/master/raw_audio/wavenet_model.py

In this post, one of the PyTorch developers showed how to wrap Conv1d to do causal convolution, like the actual Wavenet paper.

https://discuss.pytorch.org/t/causal-convolution/3456/4

In the same repo that has the "wavenet" modifications, the top level code for their own implementation actually does show some causal convolution padding that is similar to what the PyTorch developer stated on the forum:

https://github.com/f90/Seq-U-Net/blob/d2e1c8daad890c87694ddffb5956d117240b46e5/sequnet.py#L5

If I could figure out what changes would be needed to incorporate causal convolution into your code, would you be willing to show me how to benchmark it on the dataset in your paper?