search

auspicious3000 / autovc

AutoVC: Zero-Shot Voice Style Transfer with Only Autoencoder Loss
https://arxiv.org/abs/1905.05879
MIT License
983 stars 207 forks source link

reconstruction loss won't decrease #54

Open billy800413 opened 3 years ago

billy800413 commented 3 years ago

Hi, thanks for your code. But after testing it with your test data(p225, p226, p227, p228), reconstruction loss will be stuck around 0.002.(both G/loss_id and G/loss_id_psnt), num_iters is 50000, others keep original. can you give me any suggestions?

Trebolium commented 3 years ago

@billy800413 did you figure this out in the end? Vaguely recall that when i trained at 100k iterations on original test data as described in the paper, it does actually converge to 0.001. However upon training a model using an adaption of the recently uploaded training code, trained on 20 speakers, it lingers around the 0.005 mark. Have you managed to improve your sany further since writing this?