Closed avihu111 closed 10 months ago
@avihu111 thanks Avihu! indeed i forgot to clip the main discriminator! thank you 🙏
Thanks @lucidrains ! I was also wondering if I could submit pull requests. I already added to my code:
Not sure what's the common practice :) Thanks again for this awesome resource!
yeah absolutely! open source means you can always send up your changes for the common public good
thank you!
Awesome repo! Thanks a lot! I have a minor fix, which might make adversarial training more stable.
clip_grad_norm_
only clips the stft discriminator gradients. I guess the intention was clipping the gradient from all discriminators?I couldn't push a new branch to create a PR, so I'll just post the change here: Line: https://github.com/lucidrains/audiolm-pytorch/blob/898f6479aee10e5b0604a6a5a00b8a3fa6359521/audiolm_pytorch/trainer.py#L539
should change from:
to:
Thanks!!