REGION-DINO β teacher checkpoints (RATS)
Self-supervised REGION-DINO backbones whose register tokens latch onto coherent image regions, enabling RATS (recursive register decomposition + grouping) for label-free, part-to-whole segmentation.
π¬ Demo: https://huggingface.co/spaces/TimingYang/RATS
These are teacher-only weights, saved every 20 epochs plus the final epoch.
Models
| variant | arch | embed | register tokens | regs/head | epochs | per-ckpt |
|---|---|---|---|---|---|---|
| base | region_vit_base |
768 | 192 | 16 | 400 (final 388) | ~436 MB |
| small | region_vit_small |
384 | 96 | 16 | 300 | ~176 MB |
Both use patch size 16, cross-attention, num_reg_for_loss_per_head=16.
Files
base/ checkpoint_teacher_ep0020.pth β¦ ep0380.pth, ep0388.pth (20 ckpts)
small/ checkpoint_teacher_ep0020.pth β¦ ep0300.pth (15 ckpts)
Each file is a dict {"teacher": state_dict} (the EMA teacher backbone + head).
Usage
import torch
from huggingface_hub import hf_hub_download
import vision_transformer as vits # from the RATS repo / Space
ckpt = hf_hub_download("TimingYang/RATS", "base/checkpoint_teacher_ep0388.pth")
state = torch.load(ckpt, map_location="cpu")["teacher"]
state = {k.replace("module.", "").replace("backbone.", ""): v for k, v in state.items()}
model = vits.region_vit_base(patch_size=16, num_register_tokens=192,
use_cross_attn=True, num_reg_for_loss_per_head=16)
model.load_state_dict(state, strict=False) # head.* keys are unused for features
model.eval()
For small, use vits.region_vit_small(num_register_tokens=96, ...).
License
Apache-2.0. Built on facebookresearch/dino.
Inference Providers NEW
This model isn't deployed by any Inference Provider. π Ask for provider support