[FEATURE] Add ViT weights: RADIO

[seefun]

https://github.com/NVlabs/RADIO

The code and model weights of paper [CVPR 2024] AM-RADIO: Agglomerative Vision Foundation Model - Reduce All Domains Into One has been released by Nvidia

RADIO , a new vision foundation model (actually a new vit pretrained weight), excels across visual domains, serving as a superior replacement for vision backbones. Integrating CLIP variants, DINOv2, and SAM through distillation, it preserves unique features like text grounding and segmentation correspondence.

[NightMachinery]

Does RADIO have ImageNet-1k heads?

[seefun]

Does RADIO have ImageNet-1k heads?

I haven't seen it yet. But I notice the new RADIOv2.5 model is released, which merged knowledge from DFN CLIP, DINOv2, SigLIP, and SAM through multi-teacher distillation. It looks very practical in downstream task.
https://github.com/NVlabs/RADIO/blob/main/RADIOv2.5_tech_report.md