why the loss_feat become larger during trainning？

descriptinc / melgan-neurips

GAN-based Mel-Spectrogram Inversion Network for Text-to-Speech Synthesis

MIT License

978 stars 214 forks source link

Open Liujingxiu23 opened 4 years ago

Liujingxiu23 commented 4 years ago

I guess the loss_feat should get smaller during trianning？

aidiary commented 4 years ago

I got the same results. The feature matching loss will increase during training. I doubt the effect of the feature matching loss.