search

cryu854 / CUT

"Contrastive Learning for Unpaired Image-to-Image Translation" in TensorFlow 2
31 stars 7 forks source link

Result FID for horse2zebra #1

Open hjzzju opened 3 years ago

hjzzju commented 3 years ago

Thank you for your reproduce of CUT Do you calculate the FID for horse2zebra? My result is about 65, not as well as shown in the paper.

cryu854 commented 3 years ago

Hi @hjzzju I didn't calculate the FID score for CUT unfortunately, the reason is that when calculating FID, we need more than 2048 samples(because inception_v3's output has 2048 dims) to get a full rank of covariance matrix, but the test set only contains hundreds of samples, which may lead to an inaccurate quality estimation.

I'm not sure how the author of the paper overcame this problem, but they trained CUT with an additional data augmentation technique(random crop), which is not being used in this implementation.