File size: 5,653 Bytes
e3a728c 30ebf9a e3a728c 30ebf9a | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 | ---
license: other
license_name: same-as-upstream-reader-v3
library_name: pytorch
pipeline_tag: image-segmentation
tags:
- vesuvius-challenge
- ink-detection
- herculaneum
- reader-v3
- depth-head
---
# 3D Reader V3: reader-v3 with a sheet-localised depth head
**In one sentence:** the stock reader-v3 ink model plus a 16-plane depth head that says on which plane (sheet) each stroke sits, fine-tuned without any 3-D human labels, at about the same cost as reader-v3 (+1.3 % time, measured).
This is an experimental fine-tune of the public reader-v3 checkpoint (34.55 M parameters, 3-D depth-fusion stem + 2-D U-Net, 17 input planes at ~9.6 um). The extra head is a 1x1 convolution (528 parameters) on the finest decoder stage that outputs 16 logits per pixel, one per plane of the central 16 input planes. Code, tests and the full write-up: https://github.com/StewartSethA/villa/tree/progress-2026-11/sheet-localized-3d-ink-head

*Columns: CT middle plane | refined label | stock reader-v3 | this model | this model's depth hue (blue = top plane, red = mesh plane, green = bottom plane) | Hecate depth hue. Rows: Scroll 4 20251212185248 and Scroll 1 20230929220926 windows in the held-out right 10 % band of segments the model was fine-tuned on (so the sharper map partly reflects fine-tuning to the refined labels; the transfer evidence is the table below).*
## Files
| file | content |
|---|---|
| `reader_v3_depth_head.safetensors` | net + head (net keys keep the `model.` prefix; alias keys `.all_modules.` are omitted because they are the same tensors as `.conv.` / `.norm.`; head keys start with `head.`). Loading it into a freshly built network reproduces the original checkpoint's logits exactly (max abs difference 0.0). |
| `reader_v3_depth_head_last.pt` | the same weights as a plain tensor `.pt` (`state_dict`, `head`, `step`; loads with `weights_only=True`) |
| `datasets.json` | the data record of the run (see below) |
| `MD5SUMS` | checksums |
## Training data (all of it)
- **Initialisation:** the stock reader-v3 checkpoint.
- **Additional data:** 13 open-data segments, 8 of Scroll 1 (PHercParis4: 20230702185753, 20230929220926, 20231007101619, 20231012184424, 20231016151002, 20231031143852, 20231106155351, 20231210121321) and 5 of Scroll 4 (PHerc1667: 20240304144031, 20240304161941, 20251208130119, 20251212185248, 20251223230000), as 17-plane 9.6 um surface renders.
Labels: "refined core" = human label AND the registered upstream 2.4 um prediction >= 128; "maybe" pixels are never trained on; the right 10 % of every segment is held out.
Depth supervision: Hecate 9.6 um 3-D probability volumes (https://huggingface.co/scrollprize/hecate) on those segments; the 2-D label is snapped onto the ink-bearing sheet (the teacher's depth profile restricted to +-1 plane around its peak). Hecate's own training segment list is not published, so overlap of Hecate with the test segments cannot be ruled out.
- **Store mix:** every batch also contains ~9 windows of the reader-v3 training store (the data the stock model was trained on), x in [0, 0.88) of each representation. Without it the net specialises to Scroll 1/4 and loses transfer (0.75-0.79 upstream-test AUC in our first attempts).
- **Recipe:** SGD momentum 0.99 nesterov, lr 2e-4, 200-step warm-up, per-window robust-MAD normalisation, flips + rot90 on the GPU, masked BCE (2-D) + masked BCE (16 depth planes), batch 6 refined windows + ~9 store windows at 320 px, seed 0, stopped by a held-out plateau rule (step 7500).
- No data from any other scroll was used.
## Results (upstream test: 8 segments, human-label AUC, equal weights, same harness for every row, one seed each)
| model | P384 | P448 |
|---|---|---|
| stock reader-v3 (positive control; README reference 0.892) | 0.8893 | 0.892 |
| single-task control (same recipe, depth head not trained) | 0.8911 | 0.8941 |
| **this model (film +-1)** | **0.8968** | **0.8994** |
| film +-1, other seeds / 13-segment run | 0.8947 / 0.8968 | 0.8957 / 0.8991 |
| film +-3 | 0.8933 | 0.8953 |
| any sheet (no film restriction) | 0.8841 | 0.8871 |
The gain over stock (+0.0054 to +0.0075 at P384) is small and single-seed; replicate seeds are still training. The 3 of the 8 test segments that also appear in the reader-v3 training list (see the leakage-audit branch) are in-sample for stock and fine-tuned models alike.
## Use
```python
import torch
ck = torch.load("reader_v3_depth_head_last.pt", weights_only=True) # keys: state_dict, head, step
net = build_reader_v3() # the reader-v3 network of villa's vesuvius.ink_detection
net.load_state_dict(ck["state_dict"]) # strip the "model." prefix if your build uses bare keys
head = torch.nn.Conv2d(32, 16, 1); head.load_state_dict({k.removeprefix("head."): v for k, v in ck["head"].items()})
# hook the finest decoder stage (net.model.network.shared_decoder.stages[-1]) and apply `head` to its output:
# sigmoid(logits) per plane = probability that the ink sits on that plane (planes 0-15 of the 17-plane window).
# Normalise each window with the reader-v3 robust-MAD recipe; plane order must be the one the checkpoint expects
# (see the render-direction-and-intensity-check branch).
```
Reference implementation: `attach_depth_head` in the GitHub branch above.
## Limitations
Single training seed per arm; no confidence intervals; the upstream-test gain is within what seed noise could produce until replicates finish; not run through the `vesuvius` package loaders (adapter point documented in the branch); depth probabilities are soft targets, not calibrated.
|