Query: WGAN-GP FID SCORE (PyTorch)
KomputerMaster64 opened this issue · comments
Thank you for sharing the implementations of the GAN based models on popular datasets like CelebA.
I have implemented the WGAN-GP model (in PyTorch), the samples are looking closer the reported work (please refer to the attached image). When I try to evaluate the Fréchet inception distance (FID score), I am not able to comprehend the high values of 100+ (best value 113.4). Others have reported lower FID Scores. Authors in the Quality Aware Generative Adversarial Networks compared the vairous FID scores for various GAN models and there Ishaan Gulrajani's official implementation of WGAN-GP got FID score of 12.89.
I request you to guide me.
Regards
Prabhav
I have used the following repositories for reference while implementing the WGAN-GP model and evaluating the FID scores:
LynnHo/DCGAN-LSGAN-WGAN-GP-DRAGAN-Pytorch
CharlesNord/WGAN-GP-DRAGAN-Celeba-Pytorch
eriklindernoren/PyTorch-GAN
hukkelas/pytorch-frechet-inception-distance
mseitzer/pytorch-fid
This may have various causes unrelated to this library, and I can not answer general support requests.
As general advise, FID is sensitive to details related to how images are created and stored. Thus in order to compare with FID scores reported in other works, you need to make sure to reproduce the way they computed FID.
Some ideas for you to try for the case of CelebA:
- Try storing the generated images as png
- Make sure that you use 64x64 images
- Use the 'official' pre-computed statistics for CelebA: http://bioinf.jku.at/research/ttur/ttur_stats/fid_stats_celeba.npz