Open
Description
Hi, I found that the current mainline code can generate reasonable FID score for pre-trained models, but generate very high FID score for the model that is pre-trained using this codebase.
For example, I have a checkpoint that is pre-trained on the LAION dataset and get the following FID scores by using fid-clip-evaluation.py
:
Mainline -> 18.46875
ejyuen-patch-1 -> 14.32812
For anther checkpoint, I get the following result:
Mainline -> 21.46875
ejyuen-patch-1 -> 15.89062
Note that for all those FID calculation, I use the same COCO2014-10K dataset.
Metadata
Metadata
Assignees
Labels
No labels