audio Updates necessary to WaveRNN training script

Updates necessary to WaveRNN training script

Open mthrok opened this issue 4 years ago • 2 comments

Get rid of bg_iterator https://github.com/pytorch/audio/blob/a6f9cf8babfb096381e914e23950371478672b3e/examples/pipeline_wavernn/main.py#L13 This is known to slow down the process, and was removed from the library.
Replace DataParallel with DistributedDataParallel. https://github.com/pytorch/audio/blob/a6f9cf8babfb096381e914e23950371478672b3e/examples/pipeline_wavernn/main.py#L325 DataParallel is officially deprecated and slower than DistributedDataParallel.

Oct 13 '21 13:10 mthrok

We can replace the training recipe by using PyTorch Lightning which provides ddp as the backend. Then the two issues will be resolved.

Oct 13 '21 16:10 nateanl

We can replace the training recipe by using PyTorch Lightning which provides ddp as the backend. Then the two issues will be resolved.

That's a good idea. Let's do that.

Oct 13 '21 22:10 mthrok

Addressed via #2645

Oct 19 '23 16:10 mthrok