WGAN-GP does not work!!!

Question

WGAN-GP does not work!!!

LynnHo opened this issue 5 years ago · 3 comments

I have updated the code from TensorFlow 2.0-alpha to TensorFlow 2.0, everything works well except for WGAN-GP (it works in tf2.0-alpha). In tf2.0, The gradient penalty seems very unstable, but I cannot find out the problem. Does anybody help? I will be grateful.

Answer 1 · 2020-02-24T14:21:08.000Z

Try removing training=True at 111, 112, 125, 127, 128 lines of 'train.py'
(e.g. x_fake = G(z) instead of x_fake = G(z, training=True) etc.)
I got it working with that trick.

Answer 2 · 2020-02-25T01:59:36.000Z

@jaeseongyou Thanks a lot! I will try to figure out what 's wrong with training=True.

Answer 3 · 2020-08-10T14:19:34.000Z

After updating to TF2.2, this problem disappears and WGAN-GP works great.