Closed fire closed 3 years ago
Hi, we've tried WGAN GP but there was not a significant improvement, at least on Mixamo dataset. Training WGANGP will introduce second-order derivative and there will be a considerable overhead. Personally I feel it is not necessary.
Thanks.
I was wondering if the effort to replace LSGAN with WGAN GP is worthy.
The literature says that the Generate loss becomes a useful metric in WGAN GP. That could be useful.