Fid score is different from the paper

[hdjsjyl]

Hi Author,
Thanks for your great work.
For the three checkpoints you provided "best_with_denoise", "best_without_denoise" and "checkpoint_300000.pth", I got fid scores around 39 by specifying the argument "--fast_fid". But the paper shows that the Fid score is around 10. Am I missing something? Any help would be appreciated, thanks!

[yang-song]

Fast fid uses only about 1000 images (if I remember correctly) to compute the FID. You should use 50000 instead.

[shileims]

Hi @yang-song , thanks for your reply.
1. Yes, fast fid gets fid score by using only 1000 images; I will use 50000 to test again.
2. In the fast id function, it has a choice for "fast_ensemble_fid", which means it will use multiple models to comput fid score. The corresponding code is:

For reproducing the results in the paper, should I use the ensemble tech or not? Really appreciate your help. Thank you.

[yang-song]

The "fast_ensemble_fid" function should not be used. It was there for early experiments.

[shileims]

Hi @yang-song , thanks for your reply.
Another question is about the experiment on CelebA data.
Actually, the original setting of batch size for CelebA is 128. In terms of my gpu memory, I reduce the batch size from 128 to 32 and keep other settings, the checkpoint could get any image when I do sampling. Does it mean, if I reduce the batch size 4 times, I need to increase the number of iterations 4 times as well? Thank you!