merge it into vits, but got poor results about repgan HOT 10 CLOSED

yes, i have merged, can u share your samples?

from repgan.

I got all noise vioice, did you use weightnorm for repgan？

from repgan.

i think this is not the problem of WN, can u check your training samples but not inference samples?

from repgan.

The training sample has similar issues, I have called convert_weight_bias during inference. Is there anything else to be modified?

from repgan.

the training is almost the same with vits's vocoder model, that should be no problem.
u do not need to call convert_weight_bias and make sure that the training samples is good

from repgan.

Yech, the gererated training samples using forward mode are normal, but when I called convert_weight_bias and used it for inference mode, the generated samples are all noises.

from repgan.

check the test code:

from repgan.

Thanks. I found that if I inference using the train mode first and then called convert_weight_bias using the eval mode, I can generate normal results. If I directly called convert_weight_bias using the eval mode, then the results are bad. Do you know the reason?

from repgan.

may be caused by weight norm， can u try remove weight norm first and then evaluate

from repgan.

@azraelkuan Thanks a lot.
Anyther question is that what' s your configuration for multi stft discriminator? I directly used the structure and default configuration of univet, but the results were not that good than msd.
Thanks again.

from repgan.