FID from the mainline code is different from https://github.com/mosaicml/diffusion/tree/ejyuen-patch-1

[viyjy]

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.

[coryMosaicML]

Apologies for the large delay in getting a response to you on this. Can you share some more information about the model + training setup you're training and seeing lower FID with? We've matched/improved on the FID of the pre-trained models with this codebase on our training stack, but if the training data distribution differs, or if the training setup/config otherwise changes one could easily get different results.