PHerc. Paris 4 — self-trained fiber teacher UNet (epoch 30)

Internal pipeline component, not a general-purpose fiber detector. This is the frozen fiber-segmentation UNet used both to (1) warm-start and (2) generate dynamic pseudo-labels for scrollprize/fiber_dinoguided_2class_step010000, which is itself a precursor checkpoint further along the same lineage. It is superseded in quality by that later checkpoint, which adds DINO-embedding guidance on top of this model's own predictions. We are publishing it for reproducibility: anyone replicating the fiber_dinoguided_2class training run needs this exact checkpoint as an input.

Model details


Architecture	`vesuvius` `NetworkFromConfig` 3D UNet (`shared_encoder` / `shared_decoder` / `task_heads`) — verified directly from the checkpoint: 544 encoder tensors, 60 decoder tensors, a single task head
Output	2 channels, head named `task_heads.fibers` — no ink head is present in this checkpoint (verified directly: only one task head exists in the state dict, named `fibers`)
Checkpoint format	nnU-Net-style: top-level keys `model`, `optimizer`, `scheduler`, `epoch`, `normalization_scheme` (`zscore`), `intensity_properties` — no EMA weights are stored in this checkpoint
Epoch	30 (stored as `epoch: 29`, 0-indexed)
Filename fragments	`selftrain` (self-training, no fixed ground truth), `prom08`, `demo06`, `dark70` — likely threshold/hyperparameter values from its self-training curriculum; not independently decoded by us

Per the path recorded inside the downstream fiber_dinoguided_2class run's own config (.../selftrain-prom08-demo06-dark70-from-thr07ep95-cross-frame_epoch30.pth), this checkpoint's own name suggests it descends from an earlier "threshold-0.7, epoch-95, cross-frame"-trained checkpoint — plausibly connected to villa PR #825's cross-frame affine-registration training infrastructure, though we did not independently trace that earlier lineage step ourselves.

Role in the pipeline

Used by W&B run ihoo3tpl (→ scrollprize/fiber_dinoguided_2class_step010000) as:

student_init_ckpt — warm-start weights for that run's new student model.
dynamic_label.unet_ckpt — its own live inference probability map feeds the Otsu-threshold step of that run's per-step dynamic pseudo-labeling.

We found no evidence this checkpoint has any ink-related capability — only the single fibers task head exists in its state dict.

Files

File	Size
`frozen_teacher_unet__selftrain_prom08_demo06_dark70_epoch30.pth`	~1.1 GB

Usage

import torch
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    "scrollprize/fiber_selftrain_teacher_epoch30",
    "frozen_teacher_unet__selftrain_prom08_demo06_dark70_epoch30.pth",
)
ckpt = torch.load(path, map_location="cpu", weights_only=False)
state = ckpt["model"]   # no EMA weights are stored in this checkpoint
# Build with vesuvius' NetworkFromConfig (target "fibers", out_channels=2,
# in_channels=1) then load_state_dict(state).

Related models

Direct downstream consumer: scrollprize/fiber_dinoguided_2class_step010000

License

MIT.

scrollprize/fiber_selftrain_teacher_epoch30