scrollprize/dinovol_v2_ps8_supcon3class_step362500

dinovol — v2 backbone, patch size 8, supervised-contrastive fiber fine-tune (step 362500)

A further fine-tuned checkpoint of scrollprize/dinovol_v2_ps8_with_paris4_352500, continuing training from that backbone's step 352500 with an added supervised-contrastive ("supcon") objective on top of the ongoing DINO/iBOT losses. This is a representation/feature-extraction model — there is no task-specific head for direct inference; downstream models consume its dense patch embeddings. It is the frozen encoder used by scrollprize/fiber_dinoguided_2class_step010000.

Training code: dinovol (base pretraining). This fine-tuning stage's own training script was not located in the available repository snapshot at the time of writing, so the description below is reconstructed from the checkpoint's embedded config and its Weights & Biases run, not re-verified against source code.

Model details

Same backbone architecture as the base checkpoint (unchanged by this fine-tuning stage, confirmed by inspecting both checkpoints directly):

Backbone family	DINOv2/EVA ViT, 3D, with 3D RoPE (DINOv3-style)
model_type	v2
Embedding dim	864
Depth	24 blocks
Attention heads	16
MLP	SwiGLU, mlp_ratio ≈ 2.667
Register tokens	4
Patch size	8 × 8 × 8
Global crop size (train)	128 × 128 × 128
Input channels	1 (grayscale CT)
Positional encoding	RoPE mixed (base 100, normalize_coords=separate, rescale=2.0, shift=0.05, jitter=1.05)
Continues from	step 352500 of dinovol_v2_ps8_with_paris4_352500
This checkpoint's step	362500
W&B run (this fine-tune)	fiber_supcon3__paris4_0332__from352500__t0p1__w1p0__warm2k__20260522 (58oj0suf, project dinov2_pretrain, state finished)

Verified directly from the checkpoint file: student and teacher sub-payloads each carry backbone (463 tensors) + dino_head (8) + ibot_head (8) + a new contrastive_head (4 tensors: a 2-layer MLP projecting 864 → 512 → 128) not present in the base checkpoint. The field contrastive_warmup_start_step is recorded as 352500 inside this checkpoint, matching the base backbone's step exactly, so this is a genuine continuation rather than a restart.

What changed vs. the base backbone: supervised-contrastive fine-tuning

On top of the continuing DINO + iBOT + KoLeo objectives, this stage adds a supervised contrastive loss over patch tokens, driven by a per-anchor label with num_labeled_classes: 3. The checkpoint's own config logs per-batch label counts under the names contrastive_labels_air, contrastive_labels_fiber, and contrastive_labels_ignore — this strongly suggests the 3 label IDs are {air, fiber, ignore/unlabeled} rather than a fiber-orientation split, but we did not locate the exact patch-labeling source code, so treat this as a well-supported inference rather than a confirmed fact.

Other contrastive settings read from the checkpoint's config: temperature 0.1, projection dim 128, 2-layer head (hidden dim 512), loss weight 1.0, a 2000-step warmup starting at step 352500, teacher targets used for positives/negatives, cross-rank gathering enabled, plus a variance-regularization term (variance_loss_weight=0.5, variance_target_similarity=-0.5). Patch/anchor labeling used heuristic thresholds including air_threshold_hu=70.0, patch_air_threshold=0.7, patch_fiber_threshold=0.0, patch_surface_threshold=0.1, and tube_thickness_voxels=1.5. The recurrence of a "70" raw-intensity threshold here echoes the same convention seen in otsu_light_threshold=70 in the downstream fiber_dinoguided_2class training config, and in the upstream teacher UNet's dark70 filename fragment (see scrollprize/fiber_selftrain_teacher_epoch30) — though we can't confirm all three uses are the literal same threshold serving an identical purpose, just that the value recurs across this lineage.

Training data for this stage: crops from PHercParis4 and PHerc0332 (2.399 µm scan), drawn 75% / 25% from a "fiber-aware" sampler vs. a generic one. The fiber-aware sampler reads from a fiber_manifest / fiber_cache_dir under an autoreg_fiber path — plausibly connected to (not proven identical to) villa PR #825's cross-frame fiber-registration infrastructure, which added matching functionality under similarly-named autoreg-fiber-* branches in the same repository. Gram anchoring was disabled for this stage (gram.enabled: false).

Metrics (W&B run 58oj0suf, near end of training, train / val)

No accuracy metric applies to self-supervised pretraining — these are training losses only, logged at a point where the run's internal step counter had reached ~364821 (state: finished; the last saved checkpoint remained step 362500, since the next scheduled save at 365000 was never reached):

metric	train	val
total loss	24.26	23.94
contrastive loss (supcon + variance)	8.55	8.59
— supcon component	8.51	8.54
— variance component	0.087	0.106
DINO global loss	1.56	1.53
DINO local loss	9.46	9.26
iBOT loss	4.76	4.65
KoLeo loss	-0.76	-0.87
Gram loss	0 (disabled)	0 (disabled)
learning rate	1.87e-05	—

Files

File	Size	Use
frozen_dino_backbone__3class_supcon__step_362500.pt	~4.7 GB	Full training-checkpoint payload: student, teacher (each with backbone + dino_head + ibot_head + contrastive_head), optimizer, scaler, loss-centering buffers, RNG state, and the embedded config. For downstream feature extraction, use the teacher sub-dict's backbone.* weights (EMA teacher, matching the base backbone repo's recommendation).
avg_fiber_embedding__864d_backbone.npz	~3.7 KB	Supplementary. Single array avg_embedding, shape (864,), float32, L2-normalized (verified: norm = 1.0000). A reference/prototype "fiber" embedding in this backbone's feature space, used as a cosine-similarity lookup vector by the dynamic pseudo-labeling pipeline of fiber_dinoguided_2class_step010000 (its ref_embedding config field points at a file matching this one). The exact averaging/extraction procedure that produced it is not independently confirmed by us.

Unlike the base backbone repo, this repo does not include a separate slim "backbone-only" file — only the full training-checkpoint payload above.

Usage

This checkpoint has the same top-level shape as the base repo's full training checkpoint (student, teacher, optimizer, ...), plus the additional contrastive_head. Loading the backbone for feature extraction is expected to work with the same loader used for the base checkpoint:

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_supcon3class_step362500",
    "frozen_dino_backbone__3class_supcon__step_362500.pt",
)

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

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)

We inspected this checkpoint's structure directly (torch.load(..., weights_only=False)) to write the description above, but did not re-run this exact loading path end-to-end ourselves before publishing. If load_backbone_from_checkpoint rejects the extra contrastive_head keys, load teacher["backbone"] (or student["backbone"]) directly into the base architecture instead.

Related models

• Base backbone: scrollprize/dinovol_v2_ps8_with_paris4_352500 (step 352500, pre-contrastive-fine-tune)
• Direct downstream consumer: scrollprize/fiber_dinoguided_2class_step010000
• Frozen fiber teacher used alongside this backbone in that downstream run: scrollprize/fiber_selftrain_teacher_epoch30

Links

• Code: https://github.com/ScrollPrize/villa (fiber pipeline: PRs #825, #985) · dinovol (base pretraining)
• W&B run (this fine-tune): https://wandb.ai/vesuvius-challenge/dinov2_pretrain/runs/58oj0suf
• Data: https://scrollprize.org/data_browser
• Vesuvius Challenge: https://scrollprize.org

Caveats

• The training script for this specific supervised-contrastive fine-tuning stage was not located in the available repository snapshot; the description above is reconstructed from the checkpoint's own embedded config and its W&B run, not from reading the training source directly (contrast with the downstream 4-class fiber/ink pipeline, whose source we did verify directly).
• The exact semantics of the "3-class" scheme (we infer {air, fiber, ignore} from logged metric names) and the precise procedure behind avg_fiber_embedding__864d_backbone.npz are not confirmed with certainty.
• Trained on single-channel Herculaneum micro-CT; behavior on other modalities is untested. This is a pretraining/fine-tuning checkpoint only — no segmentation head is included.

License

MIT.