Exploding loss about audio-diffusion-pytorch

[alexrodi]

Turns out this was a good-old "too high a learn-rate" problem...
I was using default optimizer settings for lr, betas, eps and weight_decay
Using the base configuration on audio-diffusion-pytorch-trainer solved this issue:

self.optimizer = torch.optim.AdamW(
    params = list(self.model.parameters()),
    lr = 1e-4,
    betas= (0.95, 0.999),
    eps= 1e-6,
    weight_decay= 1e-3)

[flavioschneider]

I didn't thoroughly test vk diffusion (usually I go with v) but also never had exploding problems. Check that your dataset is distributed equally, i.e. if there are suddenly in a batch multiple silent samples that can mess up the model. For that, you might want to use the WAVDataset in https://github.com/archinetai/audio-data-pytorch with the check_silence set to true, or do some similar checks.

[alexrodi]

I am actually, using WAVDataset with check_silence set to True - which is the default, also, the datasets I'm using are taken from one-shot sound packs and another was a set of wavetables, so that's not the problem apparently...
Thanks for the help btw! :)