Audio-to-Audio
audio
speech
voice-conversion

Want to make checkpoints 1000 each.

#11
by SkyLaptor - opened

I have tried to train various sources, and generally, no matter which source, there are two points where loss_fm drops the most.
However , both of them are separated by 1000 (for example, 15000, 23000, etc.)

The following is the Tensorboard when training a certain inaccurate source.
2025-02-23_12h54_48.png

I tried to take checkpoints at 1000 separations, so I changed lines 3573 and 3674 of main.py as follows, but is there anything else I should change?

50000 if h.n_steps > 200000 else 1000

It seems to be working for the moment.

Sign up or log in to comment