Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[BUG REPORT] Pitch extraction silently fails. #22

Closed
Ph0rk0z opened this issue Jun 13, 2023 · 3 comments
Closed

[BUG REPORT] Pitch extraction silently fails. #22

Ph0rk0z opened this issue Jun 13, 2023 · 3 comments
Assignees
Labels
bug Something isn't working

Comments

@Ph0rk0z
Copy link

Ph0rk0z commented Jun 13, 2023

I have 4gb of data that I am processing into RVC numpy's. The audio is cut fine but the pitch extraction is failing silently. There are about 8000 files it makes per folder and keeps failing on the last or second to last, regardless of whether I use chrome or firefox.

Since it can't pick up where it left off, I always have to start again and it's not failing on the same file.

@Ph0rk0z Ph0rk0z added the bug Something isn't working label Jun 13, 2023
@gitmylo
Copy link
Owner

gitmylo commented Jun 14, 2023

4 gigabytes might be overkill, that would probably overtrain the model before even one epoch, unless your audio data is somehow extremely large file-size,
~400 minutes of audio, estimating from the amount of files, that's a lot considering you'd normally use 10-60 minutes, i've rarely seen people use more than 30 minutes too.

for the pitch extraction, I'll make it continue then, but you'll have to delete the folders if you want to switch pitch extraction method

@gitmylo gitmylo self-assigned this Jun 14, 2023
gitmylo added a commit that referenced this issue Jun 14, 2023
@gitmylo gitmylo closed this as completed Jun 14, 2023
@Ph0rk0z
Copy link
Author

Ph0rk0z commented Jun 14, 2023

I found what it was. The logs in the status box were crashing the browser. I stopped outputting "processing" and "extracting pitch" then it successfully completed.

As for over training, I think it came out that way. This audio is only 3 videos of someone streaming. Previous model was one and I think that came out better. Seems 1000 steps is sort of a sweet spot. The previous estimator used too few, about 10 epochs was good. Current one I need to re-run at about 4-5 vs the 1 it recommends.

Ironically, sometimes the over trained models do better on certain samples. This is just talking so singing might be a different story.

2.0 is a good loss here? I'm used to LLMs where 1.5-1.0 was the zone before it got too much of the material. I never found any best practices so I'm winging it and trying things.

@gitmylo
Copy link
Owner

gitmylo commented Jun 14, 2023

yeah, i'm still trying to find the sweet spot for the amount of training, it can still depend a lot on the audio.

the exact number in the loss is not important, just the loss relative to the previous losses is important, if it becomes unstable, you're overtraining, and need to take a previous checkpoint

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
bug Something isn't working
Projects
None yet
Development

No branches or pull requests

2 participants