search

mit-han-lab / data-efficient-gans

[NeurIPS 2020] Differentiable Augmentation for Data-Efficient GAN Training
https://arxiv.org/abs/2006.10738
BSD 2-Clause "Simplified" License
1.27k stars 175 forks source link

Model collapse problems #76

Closed RayGuo-C closed 3 years ago

RayGuo-C commented 3 years ago

Dear,

I have trained the stylegan2 + DiffAugment (DiffAugment-stylegan2-pytorch) with a simple model (64*64pixel), like following. 1628431987

But the model collapse occurred. Like following, 1628431987

I think the main reason is that the network is too strong(Filters in conv2D keep same). But there are no parameters to adjust that. As I want to generated high quantity images with limited data, the wgan-gp cannot produce attractive results although with DiffAugment. So can you give me some suggestions on this? image

zsyzzsoft commented 3 years ago

You can try adjusting fmap-base, which is an overall multiplier of the number of channels.

RayGuo-C commented 3 years ago

Thank you.