pscamillo commited on
Commit
ffbbdb4
·
verified ·
1 Parent(s): 65cb8d2

Upload folder using huggingface_hub

Browse files
README.md CHANGED
@@ -1,5 +1,82 @@
1
  ---
2
  license: other
3
  license_name: vesuvius-challenge-derived
4
- license_link: LICENSE
 
 
 
5
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
  license: other
3
  license_name: vesuvius-challenge-derived
4
+ tags:
5
+ - vesuvius-challenge
6
+ - ink-detection
7
+ base_model: scrollprize/ink_9um
8
  ---
9
+
10
+ # ink9-depth-window — checkpoints
11
+
12
+ Five arms measuring how much depth the 9 µm ink model actually needs, and what
13
+ happens when the z-jitter does not shrink with the window.
14
+
15
+ Full method, blind judgement protocol, sealed preregistrations and results:
16
+ **https://github.com/pscamillo/ink9-depth-window**
17
+
18
+ ## Which file is which
19
+
20
+ Every arm is the official `aligned21_hybrid_3d2d` recipe with **one** thing
21
+ changed. Same corpus, same seed (42), same 78,125 iterations at batch 64.
22
+
23
+ | file | window | jitter | verdict vs. `m17` |
24
+ |---|---|---|---|
25
+ | `m17-step077500.pth` | 17 slices, 163.1 µm | ±2 | baseline — replica of the official recipe |
26
+ | `b13-step077500.pth` | 13 slices, 124.7 µm | ±2 | indistinguishable, 60/60 ties |
27
+ | `b09-step077500.pth` | 9 slices, 86.4 µm | ±2 | indistinguishable, 60/60 ties |
28
+ | `b05-step077500.pth` | 5 slices, 48.0 µm | ±2 | **breaks**, 58–0, p = 0.0000 |
29
+ | `b05j0-step077500.pth` | 5 slices, 48.0 µm | **0** | indistinguishable, 60/60 ties |
30
+
31
+ **`b05` and `b05j0` have the same window.** The only difference is
32
+ `flat_z_window_jitter.max_offset`, 2 against 0. That pair is the point of the
33
+ whole thing: at 5 slices, ±2 shifts the window by 40% of its own width, and that
34
+ — not the missing depth — is what breaks it.
35
+
36
+ ## Running one
37
+
38
+ The checkpoints embed their training config, so inference rebuilds the model and
39
+ its normalisation contract on its own. Same CLI as the official release:
40
+
41
+ ```bash
42
+ python -m koine_machines.inference.infer \
43
+ <segment_9um_iso.zarr> b05j0-step077500.pth out.tif \
44
+ --overlap 0.5 --blend-mode hann --batch-size 4 --no-compile
45
+ ```
46
+
47
+ Each arm reads the depth it was trained with from the same 21-slice pooled
48
+ input, so nothing about the input has to change between them.
49
+
50
+ **Display:** with `bce_label_smoothing 0.5` the confident background sits near
51
+ 64, not 0. Rescale with `(p − 64) / 128` clipped to [0,1] or the maps look like
52
+ uniform grey even when they carry signal.
53
+
54
+ ## Two things to know before comparing against anything else
55
+
56
+ **Step 77,500, not 78,125.** `num_iterations: 78125` with `save_every: 2500`
57
+ writes no checkpoint at the final step, because the trainer only saves when
58
+ `(step+1) % save_every == 0` and the closing call does not force it. All five
59
+ arms are at the same step, so the comparisons among them are clean. The same
60
+ applies to the official release with `save_every: 5000` — its published
61
+ `step-075000` is the last one, not the final one.
62
+
63
+ **Labels.** All five arms trained on the `ink_9um` labels as they stood before
64
+ the 2026-08-18 correction, when the seam regions from the alignment pipeline
65
+ were unmasked. The defect is common to every arm — each comparison is between
66
+ two models that saw identical supervision — but these weights are **not**
67
+ directly comparable with models trained after the fix. Measured impact of the
68
+ correction, per segment, is in `results/label_correction_impact.json` in the
69
+ GitHub repo: median change 0.0%, mean +2.5% pulled by a single outlier.
70
+
71
+ ## Limits
72
+
73
+ One seed per arm. One judge for all 300 blind comparisons. Region-held-out
74
+ rather than segment-held-out validation, so nothing here speaks to cross-scroll
75
+ generalisation. The full list is in the repository README, and it is worth
76
+ reading before drawing conclusions from these weights.
77
+
78
+ ---
79
+
80
+ Derived from Vesuvius Challenge data and models (`scrollprize/ink_9um`, the
81
+ open-data S3 bucket) and subject to the terms of those sources. Code and
82
+ results in the GitHub repo are MIT.
b05-step077500.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:49d5cc0c0b7cf91c25b8b8ff416174d58d76a79c549ef59077c0f16cded45744
3
+ size 276634039
b05j0-step077500.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fa7a02d60c14d1a97003352095bc316baed143acb80c6a8420ef8f9e912ac875
3
+ size 276634423
b09-step077500.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:62829d4f5aa05f6695ff79617de3023254c5b1587f89eb861ce09948690e1975
3
+ size 276634039
b13-step077500.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fe0aca3408c31a6e6b915f8eaa4d1bbf82611e04e969547e6fce09bca6cba984
3
+ size 276634039
m17-step077500.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2f5fbc9342e030fb8d75ca6ed51000775e47c81dee1accb8d43df143d0914231
3
+ size 276634167