DINOv3¶

Meta's general-domain DINOv3 ViT-Base — facebook/dinov3-vitb16-pretrain-lvd1689m. Not medical-specific, but a strong general-domain baseline for comparison.

Embed dim	Input size	Returns	Extra
768	224×224	`(transform, encoder)`	`dinov3`

ViT-B/16 (16-px patches) with 4 register tokens in addition to the CLS token. Register tokens are an architectural choice from DINOv3 — they participate in attention but don't correspond to spatial positions, and are skipped in segmentation mode (patches only).

Install¶

uv pip install -e ".[dinov3]"

Weights are auto-downloaded from HuggingFace on first call.

Usage¶

from radharmony.evaluator.backbones import make_dinov3
from radharmony.dataset import VinDrCXRTrainDataset

transform, encoder = make_dinov3(device="cuda:0", output_keys={"img", "cls"})

ds = VinDrCXRTrainDataset(
    base_image_dir="/data/vindr/train/",
    transform=transform,
    cache_dir="/tmp/cache/dinov3/",
    output_cls=True,
)

Segmentation mode¶

from radharmony.dataset import SIIMACRPTXTrainDataset

transform, encoder = make_dinov3(device="cuda:0", output_keys={"img", "mask"})
# forward(x) -> Tensor[B, 768, 14, 14]   (224 / 16 = 14)

ds = SIIMACRPTXTrainDataset(
    base_image_dir="/data/siim-acr-ptx/dicom-images-train/",
    csv_path="/data/siim-acr-ptx/train-rle.csv",
    transform=transform,
    mask_output_dir="/tmp/cache/siim_ptx_masks/",
    output_mask=True,
    cache_dir="/tmp/cache/dinov3_seg/",
)

The recipe slices off CLS + 4 register tokens from last_hidden_state before reshaping into the spatial map.