search

yl4579 / StyleTTS2

StyleTTS 2: Towards Human-Level Text-to-Speech through Style Diffusion and Adversarial Training with Large Speech Language Models
MIT License
4.78k stars 391 forks source link

Do we need lr scheduler? #274

Open Dforgeek opened 2 months ago

Dforgeek commented 2 months ago

In all repository, I didn't find any place where scheduler.step() was used in training. Should it be this way?

martinambrus commented 4 weeks ago

I've found step() function of scheduler being used on these lines: https://github.com/yl4579/StyleTTS2/blob/main/optimizers.py#L43-L45

Also, the lr_scheduler is being used here: https://github.com/yl4579/StyleTTS2/blob/main/optimizers.py#L54

All of that is used in imports of all training scripts as from optimizers import build_optimizer.

Dforgeek commented 4 weeks ago

It is true that there's single place where .step() is being called, иut the method .scheduler() in which .step() is called has never been used for any of the training scripts. So I think it's not used.

ethan-digi commented 6 days ago

I can confirm, at least for train_finetune.py, that the lrs are not adjusted for any of the models