BigGAN: why only use translation?

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

BigGAN: why only use translation? #59

Closed AlexiaJM closed 3 years ago

AlexiaJM commented 3 years ago

It's not clear from the paper why only translation is used on BigGAN. Have color and cutout been tried? Do they not improve results or do they make them worse?

Alexia

AlexiaJM commented 3 years ago

One the same vein, was CR tried on ImageNet?

zsyzzsoft commented 3 years ago

Color and Cutout make the ImageNet results worse. However, I found that they may improve the results when EMA (exponential moving average of generator's weights) is not used. I think maybe ImageNet experiments need more careful tuning to make things work.

AlexiaJM commented 3 years ago

Thanks, I'll have to experiment a little bit then.