search

NVIDIA / mellotron

Mellotron: a multispeaker voice synthesis model based on Tacotron 2 GST that can make a voice emote and sing without emotive or singing training data
BSD 3-Clause "New" or "Revised" License
855 stars 183 forks source link

why ignore speaker embedding? #72

Open zzw922cn opened 4 years ago

zzw922cn commented 4 years ago

HI, thank you for nice code. I wonder why the ignore layers in hparams.py contains speaker embedding.weight? Don't we restore speaker embedding?

rafaelvalle commented 4 years ago

It is to be used when training on speakers not seen by the pre-trained model.

LeeYangseung commented 4 years ago

@rafaelvalle Hello, thanks for your great work.

You said ignore speaker embedding layer is to be used when training on speakers not seen by the pre-trained model. Then I don't need to ignore embedding layer when training with my custom data. Am I right??

Like this in hparams.py

ignore_layers=[''],

Thanks.

It's my mistake. I reviewed code and found ignore_layers option is activated only --warm_start when transfer learning from pre-trained model.

Thank you!