File size: 4,580 Bytes
b70c5a8
fabf840
b70c5a8
 
 
 
 
 
 
 
 
 
fabf840
b70c5a8
 
 
fabf840
b70c5a8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
fabf840
b70c5a8
 
fabf840
 
c05d896
 
 
 
6049be1
 
 
 
 
 
 
fabf840
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
---
license: cc-by-nc-4.0
pretty_name: Expanded9 original training inputs and labels
task_categories:
- image-segmentation
tags:
- vesuvius
- computed-tomography
---

# Expanded9 training data and labels

Training dataset for `expanded9-resunet128-v2-scratch-20260917`. It is also part of the training data of the Studio ink model behind the PHerc0211 First Letters submission ([code](https://github.com/priyankaranke/vesuvius-studio-multiscroll-cpt)).
113 representations: **112 training and one validation**, across 12 parent-scroll groups.
**No PHerc0211 data or labels were used for training, fine-tuning, or validation.**

- [Training code and usage](https://github.com/priyankaranke/vesuvius-expanded9-training) (private)
- [Expanded9 model (private)](https://huggingface.co/paranke/vesuvius-expanded9-resunet)
- [Submission documents and image](https://drive.google.com/drive/folders/1bvKolL_TIE3k5Ixi1qlOiAmofQ6r2ocH)
- [Methodology](https://drive.google.com/file/d/1LJaEokRQKgDLcSSWQGUBdPqEI2rjWnWe/view)

`data.json` is the original run manifest and split assignment. `prepared/<id>/`
contains the original unthinned `label.png`, sample coordinates, tile coordinates,
source metadata, and all CT tiles required by that representation. The original
raw-label collections are in `labels/` and `repository-labels/`.

The original CT tile cache was deleted after training. These are the complete
CT chunks restored in October using the frozen Expanded9 preparation functions
and source records, with the original sample/tile coordinates and label frame.
They are not claimed to have historical CT file hashes from September.
Labels here are from the original Expanded9 run, **not** the subsequent thin25 run.

To avoid 130,645 tiny Hub files, each `ct-tiles.blosc` concatenates the unchanged
Blosc chunk bytes in `tiles.npy` order. `ct-tile-offsets.npy` is a uint64 array of
N+1 byte offsets. There is no resampling, re-encoding or ZIP bundle. The training
repository supplies `materialize_dataset.py` to restore the original per-tile
files byte for byte, or inspect chunks directly using the offsets.
Each decompressed chunk is uint8 `(128,128,16)` in Y,X,depth order.

The supplied labels combine annotations and teacher-derived labels. This is not
an independently certified ground-truth collection. The sole validation region
is PHerc0841 with the recorded 2mm training exclusion. The final model is epoch10,
74,160 updates; it was not selected by its validation score.

Download (no login needed), for example:
`hf download paranke/vesuvius-expanded9-training-data --repo-type dataset --local-dir dataset`

Released under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/), the licence of the Vesuvius Challenge scan data it is derived from.
Source URLs and attribution are preserved in the original manifests.

## Submission inference data (excluded from training)

`submission/PHerc0211-3305-r002-c01/` contains the native CT planes 07–23, original saved Expanded9 probabilities and coverage, the pre-normalization geometry source, and the fixed 1cm scale-bar raster. This PHerc0211 region is inference-only; it is absent from the training manifest and loader. The final native tifxyz and final images live only in the linked Drive submission. Code lives only in GitHub. `validation-target.npz`, `validation-reference-label.png`, and `validation-source.json` support the historical validation result.

## Reproduction entry point and storage

Follow the [numbered GitHub guide](https://github.com/priyankaranke/vesuvius-expanded9-training#readme) for download → preparation → scratch training → weight export → inference → exact viewer export. All 113 prepared representations (112 training, one validation) and all 130,645 CT tiles are already here; no training sample needs a new CT render. The corpus is about 30.6 GB: 29.24 GB compressed CT plus approximately 1.4 GB labels, coordinates and source metadata. Allow about 70 GB locally when retaining the download, restored tiles and checkpoints.

The default preparation step restores the original compressed tile files from the byte packs; it does not render meshes or resample CT. Once downloaded, training needs no original workstation, historical cache, source-site login or W&B service. The historical source URLs remain provenance; the full input corpus is supplied here. Human-approved labels are preserved.

Normal Hub file downloads are sufficient; no login is needed. The Expanded9 training repository and model linked above are private. The final mesh and PNGs remain on Drive.