iris-dinov3-vits16-ssl-pilot

IRIS DINOv3 ViT-S/16 SSL Pilot

Built with DINOv3.

This private repository contains the EMA teacher backbone from a small-data, DINO-style self-distillation continuation of Meta's pretrained DINOv3 ViT-S/16. It is a development artifact for nearest-neighbour retrieval of ophthalmic surgical instruments.

This is not Meta's full DINOv3 training recipe. It is not clinically validated and must not be the sole authority for instrument identity, grasping, sorting, or robot motion.

Training integrity

• Base model revision: 114c1379950215c8b35dfcd4e90a5c251dde0d32
• Controlled allowlisted images: 496
• Epochs: 20
• Optimizer steps: 2480
• Precision: BF16
• Global crop: 224 x 224
• Local crops: 4 at 96 x 96
• Output embedding width: 384
• First epoch mean loss: 8.9329
• Final epoch mean loss: 4.9198
• Minimum recorded feature spread: 0.0227

These checks establish training and serialization integrity only. The 496 captures are adaptation-only and are not separated into independent gallery, calibration, known-query, and unknown-query sessions.

Loading

from transformers import AutoImageProcessor, AutoModel

repo_id = "Jyang1206/iris-dinov3-vits16-ssl-pilot"
processor = AutoImageProcessor.from_pretrained(repo_id)
model = AutoModel.from_pretrained(repo_id).eval()

Use outputs.pooler_output as the 384-dimensional image embedding and L2-normalize it before cosine-similarity search.

Promotion gate

Compare this adapted checkpoint with the untouched frozen DINOv3 and DINOv2 backbones using one locked, session-separated gallery/calibration/test protocol. Calibrate explicit unknown rejection and YOLO-family disagreement. If the frozen model performs better, keep the frozen model.

License

This derivative retains the upstream DINOv3 license. Review that license and keep the repository private unless redistribution and the IRIS images are both approved for public release.