Have the parameters of ResNet-50 been changed in three years?
It's hard to reproduce the results of the original paper with the downloaded parameters on ResNet-50 now, but it's amazing that they are reproduced with the ViT backbone.
Using the latest version of the CLIP code now, it's not clear what's causing this. The results in the ViT architecture are almost comparable to the original paper, but ResNet50 tested more than 10 datasets and they were much worse!
Have the parameters of ResNet-50 been changed in three years?
It's hard to reproduce the results of the original paper with the downloaded parameters on ResNet-50 now, but it's amazing that they are reproduced with the ViT backbone.
Using the latest version of the CLIP code now, it's not clear what's causing this. The results in the ViT architecture are almost comparable to the original paper, but ResNet50 tested more than 10 datasets and they were much worse!