FID evaluated on the pretrained model pkl has large difference with the reported FIDs

autonomousvision / projected-gan

[NeurIPS'21] Projected GANs Converge Faster

MIT License

883 stars 96 forks source link

FID evaluated on the pretrained model pkl has large difference with the reported FIDs #75

Closed Zhendong-Wang closed 2 years ago

Zhendong-Wang commented 2 years ago

Hi,

I tried to use the calc_metrics.py to compute the FID score of the provided pretrianed model to reproduce the reported values. However, I got large mismatchings, for example, I got 2.27 for LSUN-Bedroom and 3.41 for LSUN-Church from the pretrained model pkls. Are they the correct values that I should get?

BTW, will you release the pretrained models for AFHQ-Cat, AFHQ-Dog and AFHQ-Wild datasets.

Thanks!

xl-sr commented 2 years ago

These should be the correct pkls, the reported FIDs are calculated with this repo. Maybe there is a minor difference in how you prepared the data.