search

starrytong / SCNet

MIT License
60 stars 7 forks source link

Please add these options #22

Closed dillfrescott closed 1 week ago

dillfrescott commented 1 week ago

I really sloppily added (or attempted to add) the prodigy optimizer and also auraloss here.

Would you mind taking a look at the changes and adding the option to use prodigy and auraloss a better way in the official repo? I apologize as my coding knowledge is limited and I am not much help.

I would love to have these two things officially added as config options!

starrytong commented 1 week ago

Do prodigy and auraloss help with training? In your experiments, were they able to effectively improve performance?

dillfrescott commented 1 week ago

Prodigy basically handles the learning rate for you, and I've heard that it greatly improves performance. Auraloss doesnt necessarily improve training performance but it is more akin to human hearing and can give much better results than MSE or L1 loss functions. It can sort of repair the instrumental after the vocal removal to an extent.

dillfrescott commented 1 week ago

never mind they are already added in zfturbos training repo which has scnet so Ill close this.