[GH-ISSUE #911] Tensorboard information #585

New issue

Closed

opened 2026-10-20 10:00:55 +00:00 by bsakaguchi · 2 comments

bsakaguchi commented

2026-10-20 10:00:55 +00:00

Owner

Originally created by @jhchase on GitHub (Jul 31, 2023).
Original GitHub issue: https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/issues/911

There was a question about the overtraining of the model. Here is the training data of the model (screen[1..3].png). Is the underlined place (screen1.png) an overtraining point or did I misunderstand something? So the question arises if smoothing is set to 0.999 (screen2.png) - it is not clear at all that the model is initially incorrectly trained. Also in the last screenshot (screen3.png), the strange behavior of the training is broken, is this a mistake?

Originally created by @jhchase on GitHub (Jul 31, 2023). Original GitHub issue: https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/issues/911 There was a question about **the overtraining of the model**. Here is the training data of the model (_screen[1..3].png_). Is the underlined place (_screen1.png_) an overtraining point or did I misunderstand something? So the question arises if smoothing is set to 0.999 (_screen2.png_) - it is not clear at all that the model is initially incorrectly trained. Also in the last screenshot (_screen3.png_), the strange behavior of the training is broken, is this a mistake? ![screen1](https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/assets/140259609/cc2f87eb-2cc3-4962-ade8-de1c1aea1317) ![screen2](https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/assets/140259609/d890a168-cf7e-439a-b451-7c4fe7ff9b41) ![screen3](https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/assets/140259609/b249acd3-0e72-4840-8547-96d5af01272c)

bsakaguchi closed this issue

2026-10-20 10:00:56 +00:00

bsakaguchi commented

2026-10-20 10:00:56 +00:00

Author

Owner

@ms903x1 commented on GitHub (Jul 31, 2023):

If you terminate training before reaching the save epoch of the next checkpoint, then when you continue training, the program can only continue training from the previous saved checkpoint, and the loss curve recorded by Tensorboard will experience this kind of backtracking, which generally has no impact on the training.

@ms903x1 commented on GitHub (Jul 31, 2023): If you terminate training before reaching the save epoch of the next checkpoint, then when you continue training, the program can only continue training from the previous saved checkpoint, and the loss curve recorded by Tensorboard will experience this kind of backtracking, which generally has no impact on the training.

bsakaguchi commented

2026-10-20 10:00:56 +00:00

Author

Owner

@ms903x1 commented on GitHub (Jul 31, 2023):

For the issue of incomplete display of Tensorboard's loss curve, it is only a matter of scaling level. You can try setting the smoothness to 0, double-click the loss curve box you want to view, reset the scaling level, and then adjust the smoothness to 0.6, which may be more convenient to view.

@ms903x1 commented on GitHub (Jul 31, 2023): For the issue of incomplete display of Tensorboard's loss curve, it is only a matter of scaling level. You can try setting the smoothness to 0, double-click the loss curve box you want to view, reset the scaling level, and then adjust the smoothness to 0.6, which may be more convenient to view.