search

github-pengge / PyTorch-progressive_growing_of_gans

PyTorch implementation of Progressive Growing of GANs for Improved Quality, Stability, and Variation.
580 stars 98 forks source link

WGAN-GP Loss #8

Open dougsouza opened 6 years ago

dougsouza commented 6 years ago

Have you tried WGAN-GP loss? I am having trouble while fading in a new layer using WGAN-GP loss, it causes the loss to go grazy and results in mode collapse.

zhangqianhui commented 6 years ago

Recently, I am trying Pg-gan using WGAN-gp loss. I think the differences from ls-gan is (1)The most important: setting lower learnrate, in my experiments, lr = 0.0001 (2)n{critic} = 1, but you should keep the same variables numbers between g and d.

Hope to help you.