Hello,
First of all, thanks for the very useful library.
I found that the additional epsilon term is missing in auraloss/freq.py line 47-48 (torch.log).
This part sometimes causes nan gradient.
I modified the code by my own in my environment, but it would be helpful if the official code is also fixed.
Hello, First of all, thanks for the very useful library.
I found that the additional epsilon term is missing in auraloss/freq.py line 47-48 (torch.log). This part sometimes causes nan gradient. I modified the code by my own in my environment, but it would be helpful if the official code is also fixed.
Thanks.