audio-diffusion-pytorch
audio-diffusion-pytorch copied to clipboard
Adding dilation args for use with updated a-unet PR
This PR adds default args to the construction of the UNet to include dilation and dropout. Details on changes to underlying repo for this to work can be found here:
- https://github.com/archinetai/a-unet/pull/4
Hi @jmoso13,
I'm curious about this PR. Beyond taking inspiration from the refinements in SoundStream and EnCodec, are there any specific improvements you've noticed with this addition? I'm intuitively guessing that it might be quicker/easier to get high-end detail and keep the audio clean, based the better coverage offered by dilation, but that's a super hand-wavey intuition...
I'm also not super clear on how I'd try this out. Right now I'm using the UNetV0, as in the ADP readme examples.
Any thoughts appreciated!