dinovol_v2_ps8_with_paris4_352500

A 3D, DINOv2/DINOv3-style self-supervised representation model trained on volumetric micro-CT scans of carbonized Herculaneum scrolls. This repository publishes the EMA teacher backbone from the paris4 training run at training step 352500.

Training code: dinovol. This is a representation/feature-extraction model — there is no task-specific head; you take its dense patch embeddings and use them downstream.

Model details

Pretraining objective: DINO + iBOT + KoLeo, with late dense-feature refinement via Gram anchoring (DINOv3-style), trained with AMP.

Training data

Self-supervised on 11 open-data Herculaneum volumes (scale 0) from s3://vesuvius-challenge-open-data/:

PHerc0009B, PHerc0500P2, PHerc0814, PHerc1299, PHerc0343P, PHerc0332, PHerc0139, PHercMAN5, PHerc1451, PHercMANB, PHercParis4.

Files

Both .pt files embed the full training config, so the architecture is rebuilt automatically — no separate config needed at load time.

Usage

Clone the training repo and use its loader (the config travels inside the weights):

import torch
from huggingface_hub import hf_hub_download
from dinovol_2.eval import embedding_utils as eu

path = hf_hub_download(
    "scrollprize/dinovol_v2_ps8_with_paris4_352500",
    "dinovol_v2_ps8_paris4_step352500_teacher_backbone.pt",
)

loaded = eu.load_backbone_from_checkpoint(path, device="cuda")
backbone = loaded.backbone.eval()

# dense patch embeddings for a 1-channel volume, dims multiples of patch_size (8)
vol = torch.randn(1, 1, 128, 128, 128, device="cuda")
with torch.no_grad():
    out = backbone.forward_features(vol, masks=None, view_kind="global")
patch_tokens = out["x_norm_patchtokens"]   # (B, num_patches, 864)

For a normalized, windowed embedding grid over a real OME-Zarr volume, use eu.compute_patch_embedding_grid(...), which applies the checkpoint's normalization scheme and tiles the volume. The repo also ships a napari inspector at dinovol_2/eval/napari_visualizer.py.

License

MIT. See LICENSE.

Caveats

• Trained on single-channel Herculaneum micro-CT; behavior on other modalities is untested.
• Pretraining only — no finetuning / segmentation head is included.