scrollprize/ink_canonical_2um

Canonical 2 µm ink-detection model (ResNet-152, 3D)

⚠️ This is a canonical/production ink-detection checkpoint from the Vesuvius Challenge model registry, released here for completeness. The ink detector described in the paper is the ResNet3D-50 family in the pherc1667-ink-detection-ablation collection — this ResNet-152 model is a separate, larger variant.

Model details

Backbone	ResNet-152 3D encoder + 3D decoder ("i3d")
Recipe	new_canon_autoresearch_recipe
Input	single-channel CT at ~2 µm
Framework	PyTorch-Lightning (v2.0.9) checkpoint
Checkpoint	epoch 13 · global step 113,246
Weights	under state_dict (backbone.* + decoder)
Hyperparameters	in config.json (hyper_parameters: enc, size, pred_shape, with_norm, total_steps)

Files

• r152_3ddec_v2_l5_epoch13.ckpt — PyTorch-Lightning checkpoint (renamed from …i3depoch=13.ckpt; contents unchanged).
• config.json — the checkpoint's hyper_parameters.

How to load

import torch
ck    = torch.load("r152_3ddec_v2_l5_epoch13.ckpt", map_location="cpu", weights_only=False)
state = ck["state_dict"]
hp    = ck["hyper_parameters"]

Inference code is in https://github.com/ScrollPrize/villa.

Links

• Vesuvius Challenge: https://scrollprize.org
• Code: https://github.com/ScrollPrize/villa
• Data: https://scrollprize.org/data_browser
• Paper's ink detector (different model): pherc1667-ink-detection-ablation

License

MIT — released by the Vesuvius Challenge. Underlying tomographic data are distributed under CC BY-NC 4.0.