Closed innnky closed 1 year ago
Yes, thanks for the catch!
Should this also be fixed in train_ms.py? @p0p4k
Should this also be fixed in train_ms.py? @p0p4k
Yes, thanks for reminding me! Another thought on this matter, is then why did earlier training experiments by other users work even if the discrimator was kind of wrong? I guess it ended up like a coin flip, 50-50 (?)
Should this also be fixed in train_ms.py? @p0p4k
Yes, thanks for reminding me! Another thought on this matter, is then why did earlier training experiments by other users work even if the discrimator was kind of wrong? I guess it ended up like a coin flip, 50-50 (?)
seeing some weird speed up in ms/vctk training, so kinda seeing it
@AWAS666 speed up after changing the discriminator bug?
What about the following code from the generator part(which comes after)? It is passing the same sequence again.
y_dur_hat_r, y_dur_hat_g = net_dur_disc(hidden_x, x_mask, logw, logw_)
https://github.com/p0p4k/vits2_pytorch/commit/dba480542a54fc9c50b2117c3288675ddfc1bfa7#diff-ed183d67207df065a11e1289f19d34cc2abbc5448dea952683cfe9728c342b95R290
@FENRlR right, I am being so careless with these corrections. Thanks for spotting it. Fixed now.
@AWAS666 speed up after changing the discriminator bug?
no using it before the changes, just wanted to mention it as an example
https://github.com/p0p4k/vits2_pytorch/blob/dbdf9362ff9b8033d93e4acaa96ef2bc7a6b7646/train_ms.py#L289C54-L289C54 The last two parameters of the discriminator should be dur_real and dur_hat, respectively. https://github.com/p0p4k/vits2_pytorch/blob/dbdf9362ff9b8033d93e4acaa96ef2bc7a6b7646/models.py#L182 However, the passed parameters logw and logw_
https://github.com/p0p4k/vits2_pytorch/blob/dbdf9362ff9b8033d93e4acaa96ef2bc7a6b7646/models.py#L935 logw is predicted duration, logw_ is real duration
so it should be net_dur_disc(hidden_x, xmask, logw, logw) ?