SoundStream overfitting

[Dannynis]

Hey, I'm trying to overfit the SoundStream model on 10 examples using the most basic setup:

no attentions
only reconstruction loss
https://colab.research.google.com/drive/1pEA1vQwxNnrj7g6B_zOi4laEc82VMFB8?usp=sharing
But still couldn't managed to make it work....
Any suggestions on why is it happening ?
(p.s disabling the discriminator causes error at the trainer - .item() should be discarded)
Thank you !

[lucidrains]

@Dannynis oh oops, disabling the discriminators should work now

[Dannynis]

Thank you, much appreciated :)

[amitaie]

hey @Dannynis I have a follow-up question:
I'm trying to train SoundStream only with reconstruction loss (for debugging) and it stopped going down after a while (around 900, when using norm stft). did it happen to you to? any thoughts about it?