Wrong MoCo-v3 pretrained weights?

[tsly123]

Hi,

Thank you for the greate work.
This repo (VPT) provides the link to download the MoCo-v3 pretrained weights, named as mocov3_linear-vit-b-300ep.pth.tar. According to the MoCo-v3 github page (at the last table), it's the linear tuning ckpt not the SSL pretrained which is named as vit-b-300ep.pth.tar.

If I understand correctly, the Moco-based results in the VPT paper are got from fine-tuning experiments on the linear classification ckpt not the SSL pretrained one. So the question is: Have you tried to experiment with SSL ckpt? and Is there any different between finetuning with the SSL ckpt and linear classification ckpt?
Or am I missing something here?

Thank you