YoussefMoNader/ink-8um-v8-patchpack

ink-8um v8 patch pack

This is the complete training and validation corpus of the v8-in ink-detection model, frozen to disk as 64 × 64 × 24 uint8 patches. It holds 508,006 training and 43,249 validation patches from 19 surface segments at ~8 µm: Scroll 1, Scroll 5, PHerc. 1667, 0139, 0814, 0500P2 and 0009B, plus fragments 1, 3 and 4.

The patches were extracted with the training code's own loaders and admission rules, so training from the pack is equivalent to training from the full segment renders. The pack needs no segment downloads and loads through memory mapping.

Files

File	Shape	dtype	Content
train_images.npy	(508006, 64, 64, 24)	uint8	surface-volume patches (H, W, depth), intensities clipped to [0, 200]
train_labels.npy	(508006, 64, 64)	uint8	ink labels, 0–255 (the training target is label / 255)
train_xyxys.npy	(508006, 4)	int32	patch window [x1, y1, x2, y2] in its segment's canvas
train_seg_idx.npy	(508006,)	uint16	index into segments.json["train"]
val_images.npy … val_seg_idx.npy	43,249 rows		same layout for the validation segments
segments.json			per split: segment id, group, layer_range, reverse_layers, canvas_shape, patch_start / patch_end
val_gt/			full-resolution ink labels and masks of the validation segments, for stitched (whole-segment) metrics
SHA256SUMS			checksums of every file (sha256sum -c SHA256SUMS)

Patches of one segment are contiguous (patch_start:patch_end). Windows are 64 px on a 32 px grid, taken from 256 px super-tiles that lie entirely inside the segment's mask. For training, the super-tile must also contain ink labels. Duplicate windows are removed. Validation patches are not label-filtered. The 24 layers are the window given by layer_range in the segment's render.

Segments

Split	Segment	Group	Layers	Canvas (H × W)	Patches
train	20231005123336	Scroll1	20–44	10939 × 29872	108,688
train	20231012184423	Scroll1	20–44	15798 × 23350	120,357
train	20241030152031	Scroll5	20–44	10562 × 21987	56,447
train	20241108111522	Scroll5	20–44	7394 × 4799	21,064
train	20241108115232	Scroll5	20–44	7349 × 4379	7,854
train	20241108120732	Scroll5	20–44	7414 × 4771	22,375
train	20241113080880	Scroll5	20–44	1579 × 11367	11,474
train	20241113090990	Scroll5	20–44	1583 × 11445	9,407
train	1667_20240304144031	PHerc1667	3–27	10595 × 24525	138,818
train	0139_w033v5	PHerc0139	2–26	7100 × 7680	1,907
train	0139_w035v8	PHerc0139	2–26	5820 × 5240	1,567
train	0139_w044v8	PHerc0139	2–26	6040 × 8160	1,115
train	0139_w041v8	PHerc0139	2–26	6200 × 8020	1,034
train	0814_46527n9	PHerc0814	2–26	2180 × 3560	699
train	0500P2_front	PHerc0500P2	2–26	6280 × 3580	1,312
train	Frag1	Frags	0–24	3351 × 2592	3,888
val	0009B_ag132115	Eval0009B	0–24	7319 × 6479	32,768
val	Frag3	EvalFrags	0–24	3115 × 2150	3,698
val	Frag4	EvalFrags	0–24	4010 × 2483	6,783

The segments are Vesuvius Challenge surface-volume renders at native (~8 µm) resolution. Fragment labels (Frag1, Frag3, Frag4) derive from infrared photographs. Scroll-segment labels are annotated ink labels.

Usage

import json, numpy as np

images = np.load("train_images.npy", mmap_mode="r")    # (N, 64, 64, 24) uint8
labels = np.load("train_labels.npy", mmap_mode="r")    # (N, 64, 64) uint8
seg_idx = np.load("train_seg_idx.npy", mmap_mode="r")
segments = json.load(open("segments.json"))

x = images[i].astype(np.float32) / 255.0                # model input, depth last
y = labels[i].astype(np.float32) / 255.0
group = segments["train"][seg_idx[i]]["group"]          # scroll / fragment of the patch

For stitched validation, place each patch prediction at its val_xyxys window in a canvas of the segment's canvas_shape, and compare against val_gt/<id>_inklabels.png inside val_gt/<id>_mask.png.

To train v8-in from the pack:

hf download YoussefMoNader/ink-8um-v8-patchpack --repo-type dataset --local-dir v8_patchpack
hf download YoussefMoNader/ink-8um-v8in --local-dir ink-8um-v8in
python ink-8um-v8in/training/train_v8in.py --patch-pack v8_patchpack --output-dir runs/v8in --devices 4

The full recipe is in the v8-in model card.

License

CC BY-NC 4.0, following the Vesuvius Challenge data license of the underlying scans.