RepLKNet-31L 21k weights seems to be wrong about replknet-pytorch HOT 3 CLOSED

The weights file may not have been well trained, but at least it worked reasonably on our downstream tasks... I guess it just happens to make a mistake on this picture?

from replknet-pytorch.

I think this is not just a mistake on particular picture.
Could you check predictions of this pytorch checkpoint with your eval pipeline?

from replknet-pytorch.

I think this is not just a mistake on particular picture. Could you check predictions of this pytorch checkpoint with your eval pipeline?

The TF model might (see timm vision_transformer) be trained on 21843 classes, while ours are pretrained on 21841 classes following FAIR and microsoft (see BeiT). It seems google's imagenet22k is slightly different from ours, and a proper class mapping is likely to resolve this issue.

I think this discussion may help you find the 2 missing classes: google-research/big_transfer#18

from replknet-pytorch.