IAHispano / Applio

A simple, high-quality voice conversion tool focused on ease of use and performance
https://applio.org
MIT License
1.81k stars 292 forks source link

[Bug]: training reseting epochs #879

Open BornSaint opened 2 days ago

BornSaint commented 2 days ago

Project Version

3.2.3

Platform and OS Version

Linux x64

Affected Devices

PC

Existing Issues

No response

What happened?

I don't if this is related to me deleting previous models in the folder, but when training for a while each day, someday it just stop to continue from last epoch and started from epoch 1 again, even if last Discriminator and G model was saved correctly (I guess). I didn't deleted D and G, only small models because there weren't good yet and they were occupying large space in my SSD.

Steps to reproduce

sample_rate = '48000' batch_size = '6' total_epochs = '200' #42 save_every_epoch = '1' save_only_latest = True save_every_weights = True custom_pretrained = True overtraining_detector=True overtraining_threshold=15 d_path = 'pretraineds/D-f048k-TITAN-Medium.pth' g_path = 'pretraineds/G-f048k-TITAN-Medium.pth'

2.

train with this settings stopping and continuing train for a few times, and try to delete small models that are not the best model and not the last model each time you stop training, keeping D and G models. like this: D_233333333.pth G_233333333.pth name_10e_step_best_epoch.pth name_10e_step.pth

  1. Of course, don't delete anything else.

Expected behavior

continue training from last epoch done

Attachments

No response

Screenshots or Videos

No response

Additional Information

No response