We have an Inference dataloader and a simple collator for inference. This supports passing on the name so we can easily find the originals for benchmarking.
fixed zero_padding, potentially. But it is still unused during generation as code in LitDiffWaveModel.forward() is not batchable right now.
The generation of 650-ish audiofiles will take 14 hours on GPU.
In this PR:
zero_padding, potentially. But it is still unused during generation as code inLitDiffWaveModel.forward()is not batchable right now.The generation of 650-ish audiofiles will take 14 hours on GPU.