vits icon indicating copy to clipboard operation
vits copied to clipboard

why not use multi-scale D and loss NAN

Open hdmjdp opened this issue 3 years ago • 0 comments

https://github.com/jaywalnut310/vits/blob/2e561ba58618d021b5b8323d3765880f7e0ecfdb/models.py#L369

such as and another question, sometimes train used FP16 may cause loss=NaN? How can I fix it? thanks

hdmjdp avatar Feb 18 '22 02:02 hdmjdp