thjashin / multires-conv Goto Github PK

It was mentioned in the paper "Our model resembles WaveNet (Oord et al., 2016a) in the use of tree-structured dilated convolutions. However, our principle-guided design has distinct skip-connection structures and filter sharing patterns, resulting in significantly better parameter efficiency and performance...Additionally, the link we establish between wavelets and tree-structured dilated causal convolutions offers the first principled justification for the effectiveness of WaveNet in modeling raw audio waveforms, an exemplary case of lengthy sequences with multiscale structure."

Do you have any ablations on the difference in performance in any specific tasks or tests? Also any specific audio samples? Overall very interesting paper!

First of all, congratulations on what looks like very exciting work!

I wanted to check your generative modelling application by running autoregressive.py, however, I had to change a couple of things to get it running:

1. I got an error that MultiresLayer doesn't take the mixing argument. I figured, since it was being given False as an input, I could probably just comment it out, which does run.
2. I also got an error when the forward() function calls out = self.output_mapping(x), since that function doesn't seem to exist. I guessed this should probably be out = self.decoder(x), which seems to be running.... and converging. :)
   However, if I'm wrong about these "fixes" please let me know!

More generally, I'm very curious about the application of this model in generative contexts. I'm specifically in the music and audio field, where powerful sequence models are (obviously) essential.

So, a few things I'm wondering about:

1. Is there a relatively painless way to get images from the autoregressive_eval.py script, so I can see the output? (I can obviously dig in and figure this out, but if there's a quick mod you can suggest that would be great).
2. Do you see opportunities for conditional generation?
3. Since autoregressive generation can be somewhat limited, do you see opportunities using different training methods—for example, "infilling" generation via masking?

Again, thanks for your work.

Thanks for sharing this interesting work!
Can the MultiresConv be utilized as a drop-in replacement for conv1d operating on audio signals?

Also, is it possible to build an inverse or transpose MultiresConv and use it as a drop-in replacement for transposed conv1d?

Shared filters across timescales don’t necessarily sound like an advantage, any intuition on why that is better? do you have an ablation on this? Thanks for any insights!