LAST-ViT for ViT-L/14

[jojo-git-3518]

Dear authors,

Thanks for your great work on LAST-ViT! We are working on new research and need LAST-ViT pre-trained weights for ViT-L/14 (CLIP/DINO/Label-supervised). The current README only provides ViT-B/16 weights, so we wonder:

Do you have ViT-L/14 weights available to share?

Thanks a lot!

[jojo-git-3518]

All existing pre-trained weights are for the ViT-B/16 architecture and cannot be directly migrated to the ViT-L/14 architecture. Direct adaptation (of these weights) would fail to demonstrate the outstanding efficacy of your method, while retraining from scratch requires massive computational resources that are unaffordable in student‘s academic settings.

[ChengShiest]

Hi, thanks for your interest!

Unfortunately we don't have ViT-L/14 weights ready to release at the moment. Could you let me know which specific setting you need (CLIP, DINO, or label-supervised)? Once I have more GPU resources available I'd be happy to re-train and release one.

[jojo-git-3518]

Thanks for your great support!

We need the CLIP version of ViT-L/14, which corresponds to the link below:
https://openaipublic.azureedge.net/clip/models/b8cca3fd41ae0c99ba7e8951adf17

[wangjunyi-1]

亲爱的作者们，
可以公布ViT-L/14版本的DINO权重吗？
非常感谢！