search

odegeasslbc / FastGAN-pytorch

Official implementation of the paper "Towards Faster and Stabilized GAN Training for High-fidelity Few-shot Image Synthesis" in ICLR 2021
GNU General Public License v3.0
600 stars 100 forks source link

questions for BatchNorm #19

Closed LonglongaaaGo closed 3 years ago

LonglongaaaGo commented 3 years ago

Hello! Thanks for your beautiful work. I noticed that in your network, there are BatchNorms in the whole network. So I confused why you do not use the InstanceNorm since InstanceNorm is better than BN in the image synthesis task?

odegeasslbc commented 3 years ago

Hi I'm not sure about this statement "InstanceNorm is better than BN in the image synthesis task". Please feel free to replace BN with IN and see how it turns out.

LonglongaaaGo commented 3 years ago

Thank you! I found that the hinge loss of the Wgan incorporates the random noise like "torch.rand_like(pred) * 0.2". So can you share some reason for it? Thank you!

odegeasslbc commented 3 years ago

It's about relaxing the boundary between real and fake labels. Think about it as a label smoothing trick.