Open softrime opened 6 years ago
Hi, I've uploaded some samples in 'results/' directory. As you can listen, neither could I get the good result as the author's own demo. I'm not sure is this mainly because I did my experiments on a different dataset(VCTK) compared to the author(VCC2016), or the hyperparameters shoud be tuned more carefully.
I've listened your samples. Well I used VCC1 dataset, still not good. I saw you using L1loss for cycle-consistent loss. Have you tried using L2loss for cycle-consistent loss ? When I used L1loss as the same as paper, I got bad result and I can not train a good lower loss. But if I use L2loss instead, it was much better , although not good as author's demo. So if you have time you can have a try:)
I am re-produce this paper now, and I haven't gotten the good result as demos shown here: http://www.kecl.ntt.co.jp/people/kaneko.takuhiro/projects/cyclegan-vc/ What's your performance and may you share them? Thanks!