Congrats on getting this insightful paper out, and I will definitely use it as a backbone in my research, really nice work.
However I do have one minor feedback on how to load MetaCLIP checkpoints at this stage. Right now's process(hopefully I did not do something wrong) was a bit tedious, as I cannot strictly follow below line
model, _, preprocess = open_clip.create_model_and_transforms('ViT-B-32-quickgelu', pretrained='metaclip/b32_400m.pt')
to download weights and load models. I'm wondering whether the team is considering to add few simple lines of code to include the model checkpoints for a little bit better user experience. Thanks.