[FEATURE] Weights of DINO fine-tuned on Imagenet?

[NightMachinery]

Currently, only headless weights are available for DINO ViTs. It'd be good if the weights of a version fine-tuned for Imagenet was also added. I don't know if Facebook actually published these weights, but they do have experiments on Imagenet DINO in their paper.

[leng-yue]

Meta released a linear projection head for the DINOv2. We will be able to load it if we modify the safetensor weights.

[NightMachinery]

if we modify the safetensor weights.

What do you mean?

[leng-yue]

The official DINOv2 ships the weight of backbone and head separately. We need to concat them into one weights.

[rwightman]

It also uses a non-trivial head, needs features from more than one spot, not too crazy, but needs design for a custom head + backbone that I've thought of, but not implemented...

[mrT23]

Similar Issues:

1. [FEATURE] DINOv2 pretrained vit weights

2. Are there weights for ViT-B that was only trained on ImageNet1k?

3. [FEATURE] Model requests on dino_vit

4. ImageNet 1K pretrained weights