kan-bayashi / ParallelWaveGAN

Unofficial Parallel WaveGAN (+ MelGAN & Multi-band MelGAN & HiFi-GAN & StyleMelGAN) with Pytorch
https://kan-bayashi.github.io/ParallelWaveGAN/
MIT License
1.57k stars 343 forks source link

fix norm issue #409

Closed kan-bayashi closed 1 year ago

kan-bayashi commented 1 year ago

The norm for scale discriminator is not applied correctly. This causes mismatch of parameters with configuration. As a result, we cannot load the pretrained models.

To solve this issue, when parameter mismatch happens in loading, we remove the norm at first, load the parameters, and then apply the norm in post-hook functions.

It seems that applying weight norm and spectral norm for pretrained parameters causes unexpected behavior. Therefore, I changed to just show the message about the training error when fine-tuning and show the instruction to use pretrained parameters.

Fix the following issues: