search

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

in StyleMelGAN, the input of block not match #310

Closed yt605155624 closed 2 years ago

yt605155624 commented 2 years ago

in inference, the shape of c and x matches: https://github.com/kan-bayashi/ParallelWaveGAN/blob/6d4411b65f9487de5ec49dabf029dc107f23192d/parallel_wavegan/models/style_melgan.py#L238

but in forward, the shape of c and x not match https://github.com/kan-bayashi/ParallelWaveGAN/blob/6d4411b65f9487de5ec49dabf029dc107f23192d/parallel_wavegan/models/style_melgan.py#L142

Does this the design for GAN TTS in espnet,? https://github.com/espnet/espnet/issues/3536#issuecomment-917306178 but here is no segment_size, so , it this a bug?

yt605155624 commented 2 years ago

solved , the batch_max_steps is actually 24000(80*300) , which is different with 25500 in pwgan,

kan-bayashi commented 2 years ago

Right, you need to match batch_max_steps.