Reproduce the clip score
Hi, thanks for your great work. I wonder which exactly the CLIP model you applied to generate the clip_l14_similarity and clip_b32_similarity. Since I found that I can not generate the same similarity score. Take the following data sample as a example, the reported clip_l14_similarity is about 0.122009, while I reproduce the similarity score are 0.1393 and with openai/clip-vit-large-patch14, which is not aligned with the reported similarity score.
"055394c2036fe38e8d42aa1ff0a28594"
"https://blog-imgs-96-origin.fc2.com/t/o/k/tokinas/DSC04818.jpg"
"DSC04818.jpg"
700
467
0.241943
0.122009
[]
"ccad213724707560a1259be8c795b4f59551b2bcd6ceeee6be
Hi @zhangjc404 ,
Sorry for the late response! You can have a look here for some pointers on this: https://github.com/mlfoundations/datacomp/issues/20
Let us know if there are still issues!