Closed zzgg1 closed 2 years ago
Most of the training code should be available here: fid.py Basically, a denoising autoencoder is first trained with reconstruction loss, which is then used as features for FID. The computation of FID follows the standard FID formula.
Thanks for your amazing code! I am wondering if it could be possible that you provide pretrained model for FID, and elaborate on the detailed training procedures of FID