train the vocoder with the song data

[lhc991025]

When I go to train the vocoder with the song data, I get a situation of nan, is there a recommended configuration for training the song vocoder.

[lhc991025]

I modified these parameters, but there are many shape mismatches during training. Do you have a code modification suggestion?