Question of FID score - Githubissues

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

Question of FID score #70

Open cuikaiwen18 opened 3 years ago

cuikaiwen18 commented 3 years ago

Hi,

It seems like the reported baseline FID results are slightly different from the value reported in ADA (e.g. FID for FFHQ1k is 62.16 in your paper and 100.16 in ADA. ) depict the same architecture and hyperparameter. I am afraid I might miss some important details, please advise the causes.

zsyzzsoft commented 3 years ago

In our paper, we also augment the real images with random horizontal flips when training the baseline as a fair comparison.

thusinh1969 commented 3 years ago

Hi,

How is it possible that my pre-trained score FiD was 59 in original StyleGanV2-ADA, and in the first 200 kimg with your new codes, it went down to 10 ?!? However the quality did not change much. Something is not right :) ...!