generation pipeline
Browse files- code/README.md +104 -0
- code/features/__init__.py +18 -0
- code/features/collate.py +177 -0
- code/features/fragment_features.py +264 -0
- code/forge_LAR_2mosaic/batch.py +67 -0
- code/forge_LAR_2mosaic/cli.py +45 -0
- code/forge_LAR_2mosaic/mosaic.py +232 -0
- code/forge_LAR_2mosaic/palette.py +103 -0
- code/forge_LAR_2mosaic/remono.py +84 -0
- code/forge_LAR_2mosaic/wikiart.py +87 -0
- code/mosaic2fragments/batch.py +90 -0
- code/mosaic2fragments/curriculum.py +217 -0
- code/mosaic2fragments/forge_dataset.py +1161 -0
- code/mosaic2fragments/visualize.py +126 -0
- code/requirements.txt +9 -0
- code/tools/hf_export_tars.py +103 -0
- code/tools/hf_upload.py +53 -0
- code/tools/qa_contact_sheets.py +95 -0
- code/tools/qa_overlays.py +55 -0
- code/tools/qa_sample_hf.py +119 -0
code/README.md
ADDED
|
@@ -0,0 +1,104 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# `forge_mosaics/` — synthetic bench generation
|
| 2 |
+
|
| 3 |
+
Turns a painting into a LEGO mosaic whose brick grid is known exactly, then
|
| 4 |
+
fragments, corrupts and scatters it. Every label — instance masks, adjacencies,
|
| 5 |
+
poses — is *derived* from the grid, never hand-annotated.
|
| 6 |
+
|
| 7 |
+
```
|
| 8 |
+
painting → forge_LAR_2mosaic → mosaic2fragments → features → tools
|
| 9 |
+
mosaic + grid fragments + labels GNN input publish / QA
|
| 10 |
+
```
|
| 11 |
+
|
| 12 |
+
Python 3.11+, `pip install -r requirements.txt`. Run everything from the repository root.
|
| 13 |
+
|
| 14 |
+
```
|
| 15 |
+
forge_mosaics/
|
| 16 |
+
├── forge_LAR_2mosaic/ palette · mosaic · cli · batch · wikiart · remono
|
| 17 |
+
├── mosaic2fragments/ forge_dataset · curriculum · batch · visualize
|
| 18 |
+
├── features/ fragment_features · collate
|
| 19 |
+
└── tools/ hf_upload · hf_export_tars · qa_sample_hf · qa_overlays · qa_contact_sheets
|
| 20 |
+
```
|
| 21 |
+
|
| 22 |
+
## 1. `forge_LAR_2mosaic/` — painting → LEGO mosaic
|
| 23 |
+
|
| 24 |
+
Clean-room, pure Python: Lab quantisation onto a LEGO palette, greedy tiling, render
|
| 25 |
+
with grey joints. Outputs `canvas_mosaic_<id>.png` **and** `piece_grid_<id>.json`,
|
| 26 |
+
the exact position and colour of every brick.
|
| 27 |
+
|
| 28 |
+
| script | role |
|
| 29 |
+
|---|---|
|
| 30 |
+
| `palette.py` | 82 solid LEGO colours, **greys and silvers removed** — the joint detector downstream would take them for joints. Data table, no CLI. |
|
| 31 |
+
| `mosaic.py` | the forge: square crop → Lab quantisation → greedy largest-monochrome-rectangle packing → render. Deterministic. Imported by all the others. |
|
| 32 |
+
| `cli.py` | one image → one mosaic. |
|
| 33 |
+
| `batch.py` | one folder of images → as many mosaics. |
|
| 34 |
+
| `wikiart.py` | streams *N* images from `huggan/wikiart` through the forge, downloading nothing in full. `--max-dominant-frac` rejects near-monochrome paintings, which carry no cut signal. |
|
| 35 |
+
| `remono.py` | re-forges an existing mosaic as **1×1** tiles from its `piece_grid.json`, without re-quantising. The id is kept, so both tilings stay paired. |
|
| 36 |
+
|
| 37 |
+
```bash
|
| 38 |
+
python3 forge_mosaics/forge_LAR_2mosaic/wikiart.py --n 5000 \
|
| 39 |
+
--out-dir output/inputs --grid 96 --mode tile --max-dominant-frac 0.55
|
| 40 |
+
```
|
| 41 |
+
|
| 42 |
+
## 2. `mosaic2fragments/` — mosaic → fragments and labels
|
| 43 |
+
|
| 44 |
+
Joint detection → union-find over cells → pieces → adjacency graph → region growing
|
| 45 |
+
**along the joints** (no piece is ever cut) → corruption → placement → export.
|
| 46 |
+
|
| 47 |
+
| script | role |
|
| 48 |
+
|---|---|
|
| 49 |
+
| `forge_dataset.py` | the full pipeline on **one** mosaic; carries every knob: `--n-frag-min/max`, `--frag-distribution {balanced,compact,ultracompact}`, `--erode-px-min/max --holes-min/max --missing-min/max`, `--placement {scatter,explode}`, `--no-rotation`, `--max-area-loss`. |
|
| 50 |
+
| `curriculum.py` | **the usual entry point.** Replays the *same* mosaics under the *same* seeds across the five levels `L0_explode → L4_strong`, so that only input difficulty varies. Parallel (`--jobs`), resumes where it stopped. |
|
| 51 |
+
| `batch.py` | simpler driver: several mosaics or several seeds, no levels. |
|
| 52 |
+
| `visualize.py` | visual checks: `source_yolo.txt` polygons repainted on `source.png`; with `--recob`, what contour encoding dropped as opposed to what corruption removed. |
|
| 53 |
+
|
| 54 |
+
```bash
|
| 55 |
+
python3 forge_mosaics/mosaic2fragments/curriculum.py \
|
| 56 |
+
--inputs-glob 'output/inputs/canvas_mosaic_*.png' \
|
| 57 |
+
--frag-distribution balanced --out output --jobs 6
|
| 58 |
+
```
|
| 59 |
+
|
| 60 |
+
Each instance is one `mosaic_<id>/` folder:
|
| 61 |
+
|
| 62 |
+
| file | role |
|
| 63 |
+
|---|---|
|
| 64 |
+
| `source.png` | fragments scattered on a blank ground — segmentation and VLM **input** |
|
| 65 |
+
| `graph_fragments.json` | nodes only, no edge, no pose — GNN **input** |
|
| 66 |
+
| `source_yolo.txt` | one normalised polygon per fragment — segmentation ground truth |
|
| 67 |
+
| `graph_complete.json` | the same nodes **plus** adjacency edges — GNN target |
|
| 68 |
+
| `gt_layout.json` | exact footprint and pose of each fragment — scores the reconstruction |
|
| 69 |
+
| `target.png` | the assembled mosaic |
|
| 70 |
+
| `fragments/frag_XX.png` | one alpha crop per fragment |
|
| 71 |
+
| `degradation.md` | corruption report (empty means clean) |
|
| 72 |
+
|
| 73 |
+
> Edges are a **label read off the intact grid**, never re-derived from the corrupted
|
| 74 |
+
> geometry: two eroded fragments no longer touch, and their adjacency would be lost —
|
| 75 |
+
> exactly as two real shards no longer fit together.
|
| 76 |
+
|
| 77 |
+
## 3. `features/` — descriptors and GNN input
|
| 78 |
+
|
| 79 |
+
**Shared** between synthetic and real: works on a (mask, image) pair and knows nothing
|
| 80 |
+
about LEGO, so the same code applies to a mask from the forge or one detected by YOLO
|
| 81 |
+
on a fresco.
|
| 82 |
+
|
| 83 |
+
| script | role |
|
| 84 |
+
|---|---|
|
| 85 |
+
| `fragment_features.py` | contour, simplification under an **area-loss budget** (`--max-area-loss`, 1 % — the polygon stays inscribed in the true outline and the vertex count is an *output*), PCA canonicalisation (rotation-invariant descriptor), then per-fragment and per-side geometric and colour features. |
|
| 86 |
+
| `collate.py` | folds a dataset into one `gnn_ready.npz` per instance under **two contracts**: *dense* (padded to `n_max` + validity mask) and *ragged* (concatenated vertices + per-fragment start index, no `n_max`, for a length-agnostic encoder). Also writes `gnn_meta.json`; `--n-max` forces a value shared across datasets. |
|
| 87 |
+
|
| 88 |
+
```bash
|
| 89 |
+
for L in output/balanced/L*_*; do
|
| 90 |
+
python3 forge_mosaics/features/collate.py --dataset "$L"
|
| 91 |
+
done
|
| 92 |
+
```
|
| 93 |
+
|
| 94 |
+
## 4. `tools/` — publishing and quality control
|
| 95 |
+
|
| 96 |
+
All four Hub scripts take `--repo-id` and assume `huggingface-cli login`.
|
| 97 |
+
|
| 98 |
+
| script | role |
|
| 99 |
+
|---|---|
|
| 100 |
+
| `hf_upload.py` | pushes a folder to the Hub as a dataset. For small volumes. |
|
| 101 |
+
| `hf_export_tars.py` | for large ones: archives **one `.tar.gz` per level**, uploads it, deletes the local archive before the next (a full dataset exceeds the file count a browsable repo supports; peak disk stays one archive). Resumable. |
|
| 102 |
+
| `qa_sample_hf.py` | samples ~1 % of each published dataset **without downloading the archives**: reads the HTTP stream and cuts the connection once *N* mosaics have gone by. The sample is an archive prefix — unbiased as to content, not as to generation order. |
|
| 103 |
+
| `qa_overlays.py` | on that sample, repaints segmentation polygons and encoding loss, at constant content across levels. |
|
| 104 |
+
| `qa_contact_sheets.py` | `target ǀ source` contact sheets and target-only grids: a bad placement, a degenerate mosaic or a repeated painting shows up at a glance. |
|
code/features/__init__.py
ADDED
|
@@ -0,0 +1,18 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Module post-YOLO / pré-GNN — PARTAGÉ entre forge synthétique et chaîne réelle.
|
| 2 |
+
|
| 3 |
+
- fragment_features : (masque, image) → polygon_n (reco B + PCA), side_features, gnn_input.
|
| 4 |
+
- collate : dataset → n_max → padding + masque de validité → entrée GNN à dim fixe.
|
| 5 |
+
"""
|
| 6 |
+
from .fragment_features import (
|
| 7 |
+
extract_polygon,
|
| 8 |
+
resample_reflex_aware,
|
| 9 |
+
pca_canonical_rotation,
|
| 10 |
+
compute_fragment_features,
|
| 11 |
+
)
|
| 12 |
+
|
| 13 |
+
__all__ = [
|
| 14 |
+
"extract_polygon",
|
| 15 |
+
"resample_reflex_aware",
|
| 16 |
+
"pca_canonical_rotation",
|
| 17 |
+
"compute_fragment_features",
|
| 18 |
+
]
|
code/features/collate.py
ADDED
|
@@ -0,0 +1,177 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Préparation des entrées GNN (post-YOLO / pré-GNN) — DEUX contrats dans le même npz.
|
| 2 |
+
|
| 3 |
+
`polygon_n_canonical` et `side_features` ont une longueur VARIABLE par nœud
|
| 4 |
+
(`n_sides`). La dimension fixe qu'exige le GNN porte sur l'**EMBEDDING de nœud**,
|
| 5 |
+
pas sur le polygone brut → deux représentations livrées côte à côte :
|
| 6 |
+
|
| 7 |
+
**(A) pad à n_max + masque** (`polygon_n_canonical`, `side_features`, `valid_mask`)
|
| 8 |
+
— pour un modèle qui consomme des tenseurs denses. `n_max` est calculé sur
|
| 9 |
+
l'ENSEMBLE du dataset. ⚠️ n_max varie par dataset → pas de transfert cross-dataset
|
| 10 |
+
ni vers le réel (RePAIR : contours 403-1457 sommets).
|
| 11 |
+
|
| 12 |
+
**(B) ragged** (`verts_flat`, `sides_flat`, `node_offsets`) — les MÊMES données
|
| 13 |
+
SANS padding, concaténées sur l'axe des sommets ; le nœud i occupe la tranche
|
| 14 |
+
`[node_offsets[i]:node_offsets[i+1]]`. C'est l'entrée naturelle d'un encodeur
|
| 15 |
+
agnostique à la longueur (PointNet/Deep Sets : MLP partagé par sommet +
|
| 16 |
+
scatter-pooling par segment — `torch_scatter.scatter_max(h, seg_ids)`), qui
|
| 17 |
+
supprime n_max et rend le même modèle transférable LEGO↔réel.
|
| 18 |
+
|
| 19 |
+
Numpy pur, sans dépendance torch (le wrap torch_geometric se fait côté GNN,
|
| 20 |
+
ex. option A : `x = concat([gnn_input, poly.flatten(), sides.flatten()])`).
|
| 21 |
+
|
| 22 |
+
Usage :
|
| 23 |
+
nodes_per_graph = [json.load(open(p))['nodes'] for p in graph_fragments_paths]
|
| 24 |
+
n_max = compute_n_max(nodes_per_graph)
|
| 25 |
+
padded = [[pad_node(n, n_max) for n in nodes] for nodes in nodes_per_graph]
|
| 26 |
+
"""
|
| 27 |
+
import argparse
|
| 28 |
+
import glob
|
| 29 |
+
import json
|
| 30 |
+
import os
|
| 31 |
+
|
| 32 |
+
import numpy as np
|
| 33 |
+
|
| 34 |
+
|
| 35 |
+
def compute_n_max(nodes_per_graph):
|
| 36 |
+
"""max(n_sides) sur tous les nœuds de tous les graphes (cible de padding)."""
|
| 37 |
+
return max((len(n["polygon_n_canonical"])
|
| 38 |
+
for nodes in nodes_per_graph for n in nodes), default=0)
|
| 39 |
+
|
| 40 |
+
|
| 41 |
+
def pad_node(node, n_max):
|
| 42 |
+
"""Renvoie les features du nœud à dimension fixe + masque de validité.
|
| 43 |
+
|
| 44 |
+
- gnn_input : (7,) inchangé (déjà fixe)
|
| 45 |
+
- polygon_n_canonical : (n_max, 2) paddé à 0 au-delà de n_sides
|
| 46 |
+
- side_features : (n_max, 5) paddé à 0 au-delà de n_sides
|
| 47 |
+
- valid_mask : (n_max,) 1.0 pour les vrais sommets, 0.0 pour le padding
|
| 48 |
+
"""
|
| 49 |
+
poly = np.asarray(node["polygon_n_canonical"], dtype=np.float32)
|
| 50 |
+
sides = np.asarray(node["side_features"], dtype=np.float32)
|
| 51 |
+
n = len(poly)
|
| 52 |
+
if n > n_max:
|
| 53 |
+
raise ValueError(f"n_sides={n} > n_max={n_max} : recalculer n_max sur tout le dataset")
|
| 54 |
+
poly_p = np.zeros((n_max, 2), dtype=np.float32)
|
| 55 |
+
poly_p[:n] = poly
|
| 56 |
+
sides_p = np.zeros((n_max, 5), dtype=np.float32)
|
| 57 |
+
sides_p[:n] = sides
|
| 58 |
+
mask = np.zeros(n_max, dtype=np.float32)
|
| 59 |
+
mask[:n] = 1.0
|
| 60 |
+
return {
|
| 61 |
+
"gnn_input": np.asarray(node["gnn_input"], dtype=np.float32),
|
| 62 |
+
"polygon_n_canonical": poly_p,
|
| 63 |
+
"side_features": sides_p,
|
| 64 |
+
"valid_mask": mask,
|
| 65 |
+
}
|
| 66 |
+
|
| 67 |
+
|
| 68 |
+
def write_dataset_readme(dataset_dir, n_max, n_graphs):
|
| 69 |
+
"""README d'arborescence à la racine du dataset (auto-généré par collate)."""
|
| 70 |
+
name = os.path.basename(os.path.normpath(dataset_dir))
|
| 71 |
+
txt = f"""# `{name}/` — mosaïques LEGO fragmentées
|
| 72 |
+
|
| 73 |
+
## Fichiers par mosaïque (`mosaic_<id>/`)
|
| 74 |
+
| Fichier | Rôle |
|
| 75 |
+
|---|---|
|
| 76 |
+
| `target.png` | mosaïque **complète** = cible de reconstruction / contexte VLM |
|
| 77 |
+
| `source.png` | fragments **éclatés** sur fond blanc = **entrée YOLO** (et entrée VLM) |
|
| 78 |
+
| `source_yolo.txt` | **vérité terrain YOLO-Seg** : 1 polygone/fragment, normalisé `[0,1]` (entraîne YOLO avec `source.png`) |
|
| 79 |
+
| `source_yolo_viz.png` | overlay debug des labels (inspection humaine) — *absent par défaut* (opt-in `--debug`) |
|
| 80 |
+
| `pieces.json` | debug pur (pièces LEGO détectées) ; **n'entraîne PAS le YOLO** (c'est `source_yolo.txt`), lu nulle part — *absent par défaut* (opt-in `--debug`) |
|
| 81 |
+
| `graph_fragments.json` | **ENTRÉE GNN** : nœuds (features/fragment), zéro arête, sans `target_info` |
|
| 82 |
+
| `graph_complete.json` | **CIBLE GNN** : mêmes nœuds + `target_info` (leak) + arêtes (mating graph) |
|
| 83 |
+
| `gt_layout.json` | **GT de RECONSTRUCTION** (réponse finale) : footprint exact + pose `(x,y,rot)` de chaque fragment dans `target.png`. Sert à **scorer l'assemblage** (IoU/Q_pos) APRÈS la tête de pose du GNN / le VLM qui place les fragments |
|
| 84 |
+
| `gnn_ready.npz` | entrée GNN, **2 contrats** : (A) dense pad `n_max={n_max}`+masque ; (B) **ragged** `verts_flat`/`sides_flat`+`node_offsets` (sans n_max, pour encodeur agnostique) |
|
| 85 |
+
| `degradation.md` | rapport de dégradation (0 si clean) |
|
| 86 |
+
| `fragments/frag_XX.png` | crop alpha par fragment = **entrée VLM** |
|
| 87 |
+
|
| 88 |
+
Au niveau dataset : **`gnn_meta.json`** (`n_max={n_max}`, noms de features).
|
| 89 |
+
|
| 90 |
+
## Schéma d'un nœud (`graph_fragments.json`)
|
| 91 |
+
| Bloc | Contenu | Taille |
|
| 92 |
+
|---|---|---|
|
| 93 |
+
| `node_id` | identifiant stable (référencé par les arêtes) | — |
|
| 94 |
+
| `gnn_input` | `[area, perimeter, R, G, B, bbox_w, bbox_h]` — domaine-agnostique | 7 |
|
| 95 |
+
| `polygon_n_canonical` | contour en repère **PCA canonique** (invariant rotation) | `n_sides`×2 |
|
| 96 |
+
| `side_features` | par côté `[length, angle, R, G, B]` | `n_sides`×5 |
|
| 97 |
+
|
| 98 |
+
`n_sides` **variable** par nœud → paddé à `n_max={n_max}` + masque dans `gnn_ready.npz`.
|
| 99 |
+
⚠️ **`target_info`** (dans `graph_complete`) = vérité terrain → ne **jamais** donner en entrée.
|
| 100 |
+
|
| 101 |
+
_Régénéré par `forge_mosaics/mosaic2fragments/{{batch,curriculum}}.py` + `forge_mosaics/features/collate.py`._
|
| 102 |
+
"""
|
| 103 |
+
open(os.path.join(dataset_dir, "README.md"), "w", encoding="utf-8").write(txt)
|
| 104 |
+
|
| 105 |
+
|
| 106 |
+
def build_dataset(dataset_dir, out_name="gnn_ready.npz", n_max=None):
|
| 107 |
+
"""Étape COLLATE (post-YOLO / pré-GNN) : scanne tout le dataset, calcule n_max,
|
| 108 |
+
pad+masque chaque nœud, et écrit par mosaïque un `gnn_ready.npz` (entrée GNN à
|
| 109 |
+
dimension fixe) + un `gnn_meta.json` (n_max + schéma). Numpy pur : les amis
|
| 110 |
+
GNN chargent et wrappent en torch_geometric.Data eux-mêmes (les arêtes/cibles
|
| 111 |
+
restent dans graph_complete.json)."""
|
| 112 |
+
mosaics = sorted(glob.glob(os.path.join(dataset_dir, "mosaic_*")))
|
| 113 |
+
nodes_per = [json.load(open(os.path.join(m, "graph_fragments.json")))["nodes"]
|
| 114 |
+
for m in mosaics]
|
| 115 |
+
observed = compute_n_max(nodes_per)
|
| 116 |
+
if n_max is None:
|
| 117 |
+
n_max = observed # auto : max du dataset (historique)
|
| 118 |
+
elif observed > n_max: # n_max JOINT imposé → doit couvrir
|
| 119 |
+
raise ValueError(f"--n-max {n_max} < max observé {observed} : "
|
| 120 |
+
"choisir un n_max joint qui couvre ce dataset")
|
| 121 |
+
for m, nodes in zip(mosaics, nodes_per):
|
| 122 |
+
if not nodes:
|
| 123 |
+
continue
|
| 124 |
+
padded = [pad_node(n, n_max) for n in nodes]
|
| 125 |
+
# contrat B (ragged) : mêmes données sans padding, concaténées par sommet
|
| 126 |
+
lens = np.array([len(n["polygon_n_canonical"]) for n in nodes], dtype=np.int64)
|
| 127 |
+
np.savez(
|
| 128 |
+
os.path.join(m, out_name),
|
| 129 |
+
gnn_input=np.stack([p["gnn_input"] for p in padded]), # (N,7)
|
| 130 |
+
polygon_n_canonical=np.stack([p["polygon_n_canonical"] for p in padded]), # (N,n_max,2)
|
| 131 |
+
side_features=np.stack([p["side_features"] for p in padded]), # (N,n_max,5)
|
| 132 |
+
valid_mask=np.stack([p["valid_mask"] for p in padded]), # (N,n_max)
|
| 133 |
+
node_id=np.array([n["node_id"] for n in nodes], dtype=np.int64), # (N,)
|
| 134 |
+
verts_flat=np.concatenate( # (Σn,2) ragged
|
| 135 |
+
[np.asarray(n["polygon_n_canonical"], dtype=np.float32) for n in nodes]),
|
| 136 |
+
sides_flat=np.concatenate( # (Σn,5) ragged
|
| 137 |
+
[np.asarray(n["side_features"], dtype=np.float32) for n in nodes]),
|
| 138 |
+
node_offsets=np.concatenate([[0], np.cumsum(lens)]), # (N+1,)
|
| 139 |
+
)
|
| 140 |
+
sample = json.load(open(os.path.join(mosaics[0], "graph_fragments.json")))
|
| 141 |
+
meta = {
|
| 142 |
+
"n_max": n_max,
|
| 143 |
+
"n_max_observed": observed, # max réel du dataset (≠ n_max si joint imposé)
|
| 144 |
+
"n_graphs": len(mosaics),
|
| 145 |
+
"gnn_input_dim": len(sample["gnn_input_feature_names"]),
|
| 146 |
+
"gnn_input_feature_names": sample["gnn_input_feature_names"],
|
| 147 |
+
"side_feature_names": sample["side_feature_names"],
|
| 148 |
+
"note": ("gnn_ready.npz par mosaïque, DEUX contrats : (A) dense = "
|
| 149 |
+
"polygon_n_canonical/side_features paddés à n_max + valid_mask ; "
|
| 150 |
+
"(B) ragged = verts_flat/sides_flat + node_offsets (nœud i = "
|
| 151 |
+
"tranche [offsets[i]:offsets[i+1]]), SANS n_max — pour encodeur "
|
| 152 |
+
"agnostique (PointNet/Deep Sets, scatter-pooling par segment). "
|
| 153 |
+
"Cibles/arêtes : graph_complete.json. Régénérable via "
|
| 154 |
+
"`python3 forge_mosaics/features/collate.py --dataset <dir>`."),
|
| 155 |
+
}
|
| 156 |
+
json.dump(meta, open(os.path.join(dataset_dir, "gnn_meta.json"), "w"),
|
| 157 |
+
indent=2, ensure_ascii=False)
|
| 158 |
+
write_dataset_readme(dataset_dir, n_max, len(mosaics))
|
| 159 |
+
return n_max
|
| 160 |
+
|
| 161 |
+
|
| 162 |
+
def main():
|
| 163 |
+
p = argparse.ArgumentParser(description="Collate post-YOLO → tenseurs GNN à dim fixe")
|
| 164 |
+
p.add_argument("--dataset", required=True, help="dossier dataset (contient mosaic_*/)")
|
| 165 |
+
p.add_argument("--out-name", default="gnn_ready.npz")
|
| 166 |
+
p.add_argument("--n-max", type=int, default=None,
|
| 167 |
+
help="n_max JOINT imposé (au lieu du max du dataset) : le même sur "
|
| 168 |
+
"plusieurs paliers/datasets/le réel → un seul modèle pad+masque "
|
| 169 |
+
"cross-datasets. Erreur si < max observé. Repère : RePAIR réel "
|
| 170 |
+
"≤291 après reco-B ε=1 %% → 304 couvre LEGO et réel")
|
| 171 |
+
args = p.parse_args()
|
| 172 |
+
n_max = build_dataset(args.dataset, args.out_name, n_max=args.n_max)
|
| 173 |
+
print(f"[collate] {args.dataset} : n_max={n_max} → {args.out_name} par mosaïque + gnn_meta.json")
|
| 174 |
+
|
| 175 |
+
|
| 176 |
+
if __name__ == "__main__":
|
| 177 |
+
main()
|
code/features/fragment_features.py
ADDED
|
@@ -0,0 +1,264 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Extraction de features par fragment — module PARTAGÉ (post-YOLO / pré-GNN).
|
| 2 |
+
|
| 3 |
+
Domaine-agnostique : marche sur n'importe quel masque de fragment (+ image),
|
| 4 |
+
qu'il vienne de la forge synthétique LEGO OU de YOLO sur une mosaïque réelle.
|
| 5 |
+
C'est la garantie d'alignement de schéma entre synthétique et réel : le même
|
| 6 |
+
code produit les mêmes features dans les deux chaînes.
|
| 7 |
+
|
| 8 |
+
Contenu : contour, resample reflex-aware (reco B, polygon_n ⊆ polygon_raw),
|
| 9 |
+
canonicalisation PCA (invariance en rotation), et features par fragment
|
| 10 |
+
(géométrie + couleur de bord). Aucune dépendance LEGO ici (pas de studs, pas de
|
| 11 |
+
grille) — aucune notion LEGO ici. Les bits LEGO/GT (adjacence grille, n_pieces en
|
| 12 |
+
métadonnée) restent dans la forge.
|
| 13 |
+
"""
|
| 14 |
+
import numpy as np
|
| 15 |
+
import cv2
|
| 16 |
+
|
| 17 |
+
|
| 18 |
+
# ============================================================
|
| 19 |
+
# Contour
|
| 20 |
+
# ============================================================
|
| 21 |
+
|
| 22 |
+
def extract_polygon(mask):
|
| 23 |
+
contours, _ = cv2.findContours(mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
|
| 24 |
+
if not contours:
|
| 25 |
+
return np.zeros((0, 2), dtype=np.float32)
|
| 26 |
+
cnt = max(contours, key=cv2.contourArea)
|
| 27 |
+
return cnt.reshape(-1, 2).astype(np.float32)
|
| 28 |
+
|
| 29 |
+
|
| 30 |
+
# ============================================================
|
| 31 |
+
# Resample reflex-aware (reco B) : polygon_n ⊆ polygon_raw
|
| 32 |
+
# ============================================================
|
| 33 |
+
|
| 34 |
+
def _polygon_signed_area(P):
|
| 35 |
+
x, y = P[:, 0], P[:, 1]
|
| 36 |
+
return 0.5 * float(np.dot(x, np.roll(y, -1)) - np.dot(y, np.roll(x, -1)))
|
| 37 |
+
|
| 38 |
+
|
| 39 |
+
def _classify_vertices(P):
|
| 40 |
+
"""Per vertex: +1 convex, -1 reflex (concave), 0 ~collinear."""
|
| 41 |
+
M = len(P)
|
| 42 |
+
o = np.sign(_polygon_signed_area(P)) or 1.0
|
| 43 |
+
kind = np.zeros(M, dtype=int)
|
| 44 |
+
for i in range(M):
|
| 45 |
+
a, b, c = P[i - 1], P[i], P[(i + 1) % M]
|
| 46 |
+
cross = (b[0] - a[0]) * (c[1] - b[1]) - (b[1] - a[1]) * (c[0] - b[0])
|
| 47 |
+
v = cross * o
|
| 48 |
+
kind[i] = 0 if abs(v) < 1e-6 else (1 if v > 0 else -1)
|
| 49 |
+
return kind
|
| 50 |
+
|
| 51 |
+
|
| 52 |
+
def _clip_to_polygon(poly_enc, poly_raw):
|
| 53 |
+
"""Safety clamp: intersect the encoded polygon with the real piece so that
|
| 54 |
+
polygon_n ⊆ polygon_raw is GUARANTEED even on a pathological contour.
|
| 55 |
+
No-op if shapely is unavailable (reflex-aware alone already gives ~0 gain)."""
|
| 56 |
+
try:
|
| 57 |
+
from shapely.geometry import Polygon
|
| 58 |
+
except Exception:
|
| 59 |
+
return poly_enc
|
| 60 |
+
C = Polygon(poly_enc).buffer(0).intersection(Polygon(poly_raw).buffer(0))
|
| 61 |
+
if C.geom_type == 'Polygon':
|
| 62 |
+
polys = [C]
|
| 63 |
+
elif C.geom_type in ('MultiPolygon', 'GeometryCollection'):
|
| 64 |
+
polys = [g for g in C.geoms if g.geom_type == 'Polygon' and not g.is_empty]
|
| 65 |
+
else:
|
| 66 |
+
polys = []
|
| 67 |
+
if not polys:
|
| 68 |
+
return poly_enc
|
| 69 |
+
best = max(polys, key=lambda g: g.area)
|
| 70 |
+
return np.array(best.exterior.coords[:-1], dtype=np.float32)
|
| 71 |
+
|
| 72 |
+
|
| 73 |
+
MAX_AREA_LOSS_DEFAULT = 0.01 # ε : perte d'aire tolérée pour polygon_n (1 %)
|
| 74 |
+
|
| 75 |
+
|
| 76 |
+
def resample_reflex_aware(polygon_raw, n_min=0, clip=True,
|
| 77 |
+
max_area_loss=MAX_AREA_LOSS_DEFAULT):
|
| 78 |
+
"""Reco B — simplifie le contour en **budgétant la PERTE D'AIRE**, `n` en sortie.
|
| 79 |
+
|
| 80 |
+
Garantit `polygon_n ⊆ polygon_raw` : on ne *gagne* jamais d'aire, on en perd
|
| 81 |
+
(« comme si les coins étaient ébréchés ») → les fragments réassemblés ne se
|
| 82 |
+
chevauchent jamais, ils laissent des jours (régime physiquement réalisable).
|
| 83 |
+
Renvoie un polygone de longueur VARIABLE.
|
| 84 |
+
|
| 85 |
+
Algorithme = **Visvalingam-Whyatt** (1993) *restreint aux sommets convexes* :
|
| 86 |
+
on retire itérativement le sommet dont le triangle (préc., lui, suiv.) a la
|
| 87 |
+
plus petite aire, tant que la perte CUMULÉE reste ≤ `max_area_loss` (fraction
|
| 88 |
+
de l'aire du polygone). Deux invariants :
|
| 89 |
+
- **seuls les convexes sont retirables** — retirer un sommet reflex
|
| 90 |
+
pontifierait une concavité et **re-gagnerait** de l'aire (viole ⊆ raw) ;
|
| 91 |
+
- les sommets ~colinéaires ont un triangle d'aire nulle → retirés en
|
| 92 |
+
premier, **gratuitement** (c'est le gros du bruit de contour LEGO).
|
| 93 |
+
|
| 94 |
+
⚠️ **Corrigé le 23/07/2026.** L'implémentation précédente budgétait le NOMBRE
|
| 95 |
+
de sommets (`n_target` ∈ [16,24]) en amorçant `keep` avec tous les reflex :
|
| 96 |
+
comme `#reflex` (médiane **58** sur LEGO) ≥ `n_target` dans **100 %** des
|
| 97 |
+
fragments, la boucle sortait immédiatement et **aucun convexe n'était jamais
|
| 98 |
+
gardé**. Résultat : perte d'aire médiane **21,4 %** (p95 85,8 %, un quart des
|
| 99 |
+
fragments perdant >50 % de leur aire) sur la config k=10-15 du dataset 25k,
|
| 100 |
+
et jusqu'à 99 % à k=2. L'invariant ⊆ raw tenait (pas d'aire fantôme) mais
|
| 101 |
+
l'AMPLEUR de la perte n'était ni bornée ni mesurée. On budgète désormais la
|
| 102 |
+
grandeur qui compte (la fidélité), et `n` en découle.
|
| 103 |
+
|
| 104 |
+
`n_min` : plancher optionnel sur le nombre de sommets (0 = pas de plancher).
|
| 105 |
+
Le plancher réel reste **#reflex**, atteint automatiquement.
|
| 106 |
+
"""
|
| 107 |
+
import heapq
|
| 108 |
+
|
| 109 |
+
P = np.asarray(polygon_raw, dtype=np.float32)
|
| 110 |
+
M = len(P)
|
| 111 |
+
if M < 4:
|
| 112 |
+
return P.copy()
|
| 113 |
+
|
| 114 |
+
total_area = abs(_polygon_signed_area(P))
|
| 115 |
+
if total_area <= 0:
|
| 116 |
+
return P.copy()
|
| 117 |
+
budget = max_area_loss * total_area
|
| 118 |
+
|
| 119 |
+
orient = np.sign(_polygon_signed_area(P)) or 1.0
|
| 120 |
+
prev = [(i - 1) % M for i in range(M)]
|
| 121 |
+
nxt = [(i + 1) % M for i in range(M)]
|
| 122 |
+
alive = [True] * M
|
| 123 |
+
n_alive = M
|
| 124 |
+
|
| 125 |
+
def tri(i):
|
| 126 |
+
"""(aire du triangle retiré, convexe ?) pour le sommet i dans l'état courant."""
|
| 127 |
+
a, b, c = P[prev[i]], P[i], P[nxt[i]]
|
| 128 |
+
cross = (b[0] - a[0]) * (c[1] - b[1]) - (b[1] - a[1]) * (c[0] - b[0])
|
| 129 |
+
return abs(cross) / 2.0, (cross * orient) >= 0 # ≥ 0 : convexe ou colinéaire
|
| 130 |
+
|
| 131 |
+
heap = []
|
| 132 |
+
for i in range(M):
|
| 133 |
+
a, cvx = tri(i)
|
| 134 |
+
if cvx:
|
| 135 |
+
heapq.heappush(heap, (a, i))
|
| 136 |
+
|
| 137 |
+
lost = 0.0
|
| 138 |
+
while heap and n_alive > 3:
|
| 139 |
+
a, i = heapq.heappop(heap)
|
| 140 |
+
if not alive[i]:
|
| 141 |
+
continue
|
| 142 |
+
a_now, cvx_now = tri(i) # ré-évaluation paresseuse
|
| 143 |
+
if not cvx_now: # devenu reflex → intouchable
|
| 144 |
+
continue
|
| 145 |
+
if a_now > a + 1e-9: # coût périmé → re-pousser
|
| 146 |
+
heapq.heappush(heap, (a_now, i))
|
| 147 |
+
continue
|
| 148 |
+
if lost + a_now > budget: # budget épuisé
|
| 149 |
+
break
|
| 150 |
+
if n_min and n_alive <= n_min:
|
| 151 |
+
break
|
| 152 |
+
alive[i] = False
|
| 153 |
+
n_alive -= 1
|
| 154 |
+
lost += a_now
|
| 155 |
+
p, n = prev[i], nxt[i]
|
| 156 |
+
nxt[p], prev[n] = n, p
|
| 157 |
+
for j in (p, n): # coûts des voisins modifiés
|
| 158 |
+
aj, cj = tri(j)
|
| 159 |
+
if cj and alive[j]:
|
| 160 |
+
heapq.heappush(heap, (aj, j))
|
| 161 |
+
|
| 162 |
+
poly = P[[i for i in range(M) if alive[i]]]
|
| 163 |
+
if clip:
|
| 164 |
+
poly = _clip_to_polygon(poly, P) # filet de sécurité ⊆ raw
|
| 165 |
+
return np.asarray(poly, dtype=np.float32)
|
| 166 |
+
|
| 167 |
+
|
| 168 |
+
# ============================================================
|
| 169 |
+
# Canonicalisation PCA (invariance en rotation)
|
| 170 |
+
# ============================================================
|
| 171 |
+
|
| 172 |
+
def pca_canonical_rotation(mask):
|
| 173 |
+
"""Rotation 2×2 amenant le fragment dans son repère PCA canonique (axe
|
| 174 |
+
principal → x), signe déterministe (skewness du 3e moment sur l'axe majeur).
|
| 175 |
+
|
| 176 |
+
Rend la forme INVARIANTE EN ROTATION : un même fragment, à n'importe quelle
|
| 177 |
+
orientation, donne le même descripteur → plus de fuite d'orientation, et
|
| 178 |
+
repère-target ≡ repère-source (ce que produirait YOLO). Rotation PURE
|
| 179 |
+
(det=+1, pas de miroir : la chiralité compte pour le matching). Ambigu pour
|
| 180 |
+
les formes ~symétriques / ~carrées (axe instable) — limitation acceptée.
|
| 181 |
+
"""
|
| 182 |
+
ys, xs = np.where(mask > 0)
|
| 183 |
+
if len(xs) < 3:
|
| 184 |
+
return np.eye(2, dtype=np.float64)
|
| 185 |
+
pts = np.stack([xs, ys], axis=1).astype(np.float64)
|
| 186 |
+
pts -= pts.mean(axis=0)
|
| 187 |
+
evals, evecs = np.linalg.eigh(pts.T @ pts)
|
| 188 |
+
u = evecs[:, int(np.argmax(evals))] # axe principal (unitaire)
|
| 189 |
+
if float(((pts @ u) ** 3).sum()) < 0: # désambiguïse le flip 180°
|
| 190 |
+
u = -u
|
| 191 |
+
return np.array([[u[0], u[1]], [-u[1], u[0]]], dtype=np.float64) # u → axe x, det=+1
|
| 192 |
+
|
| 193 |
+
|
| 194 |
+
# ============================================================
|
| 195 |
+
# Features par fragment (géométrie canonique + couleur de bord)
|
| 196 |
+
# ============================================================
|
| 197 |
+
|
| 198 |
+
def _is_joint_pixel(px, target=136, tol=30):
|
| 199 |
+
r, g, b = int(px[0]), int(px[1]), int(px[2])
|
| 200 |
+
return (abs(r - target) < tol and abs(g - target) < tol and abs(b - target) < tol
|
| 201 |
+
and max(r, g, b) - min(r, g, b) < 25)
|
| 202 |
+
|
| 203 |
+
|
| 204 |
+
def _sample_inward(p0, p1, centroid, img, offset=6):
|
| 205 |
+
"""Sample piece interior color along a polygon side.
|
| 206 |
+
|
| 207 |
+
The naive midpoint + small offset can land on a LEGO joint (gray) — we
|
| 208 |
+
therefore try a few samples along the side and reject pixels that look like
|
| 209 |
+
joints, falling back to the median of all samples if everything is gray.
|
| 210 |
+
(Sur fresque réelle : à enrichir vers un profil de couleur le long du bord.)
|
| 211 |
+
"""
|
| 212 |
+
samples = []
|
| 213 |
+
for t in (0.3, 0.5, 0.7):
|
| 214 |
+
pt = (1 - t) * p0 + t * p1
|
| 215 |
+
inward = centroid - pt
|
| 216 |
+
norm = float(np.linalg.norm(inward))
|
| 217 |
+
if norm < 1e-6:
|
| 218 |
+
continue
|
| 219 |
+
inward = inward / norm * offset
|
| 220 |
+
s = pt + inward
|
| 221 |
+
sx = int(np.clip(s[0], 0, img.shape[1] - 1))
|
| 222 |
+
sy = int(np.clip(s[1], 0, img.shape[0] - 1))
|
| 223 |
+
px = img[sy, sx][:3]
|
| 224 |
+
samples.append(px)
|
| 225 |
+
if not samples:
|
| 226 |
+
cx = int(np.clip(centroid[0], 0, img.shape[1] - 1))
|
| 227 |
+
cy = int(np.clip(centroid[1], 0, img.shape[0] - 1))
|
| 228 |
+
return img[cy, cx][:3].tolist()
|
| 229 |
+
non_joint = [s for s in samples if not _is_joint_pixel(s)]
|
| 230 |
+
pool = non_joint if non_joint else samples
|
| 231 |
+
return np.median(np.array(pool), axis=0).astype(int).tolist()
|
| 232 |
+
|
| 233 |
+
|
| 234 |
+
def compute_fragment_features(polygon_n, area, mean_color, img, polygon_geom=None):
|
| 235 |
+
"""GÉOMÉTRIE (perimeter, bbox, length, angle) calculée sur `polygon_geom`
|
| 236 |
+
(repère canonique PCA → invariant en rotation, si fourni) ; COULEUR
|
| 237 |
+
échantillonnée sur `polygon_n` (repère image, pour indexer les pixels).
|
| 238 |
+
|
| 239 |
+
`area` est passé en scalaire (la forge : len(cells)·stud² ; un chemin post-YOLO
|
| 240 |
+
réel : masque.sum()) → fonction 100 % DOMAINE-AGNOSTIQUE (aucune notion LEGO).
|
| 241 |
+
cx,cy = centroïde image (GT de position, pas un input)."""
|
| 242 |
+
geom = polygon_n if polygon_geom is None else np.asarray(polygon_geom, dtype=np.float64)
|
| 243 |
+
centroid_img = polygon_n.mean(axis=0)
|
| 244 |
+
cx, cy = float(centroid_img[0]), float(centroid_img[1])
|
| 245 |
+
closed = np.vstack([geom, geom[0:1]])
|
| 246 |
+
perimeter = float(np.sqrt(((np.diff(closed, axis=0)) ** 2).sum(axis=1)).sum())
|
| 247 |
+
bbox_w = float(geom[:, 0].max() - geom[:, 0].min()) # repère canonique
|
| 248 |
+
bbox_h = float(geom[:, 1].max() - geom[:, 1].min())
|
| 249 |
+
|
| 250 |
+
side_features = []
|
| 251 |
+
n = len(polygon_n)
|
| 252 |
+
for i in range(n):
|
| 253 |
+
g0, g1 = geom[i], geom[(i + 1) % n]
|
| 254 |
+
length = float(np.linalg.norm(g1 - g0))
|
| 255 |
+
angle = float(np.arctan2(g1[1] - g0[1], g1[0] - g0[0])) # angle canonique
|
| 256 |
+
c = _sample_inward(polygon_n[i], polygon_n[(i + 1) % n], centroid_img, img)
|
| 257 |
+
side_features.append([length, angle, float(c[0]), float(c[1]), float(c[2])])
|
| 258 |
+
|
| 259 |
+
global_features = [
|
| 260 |
+
float(area), perimeter, cx, cy,
|
| 261 |
+
float(mean_color[0]), float(mean_color[1]), float(mean_color[2]),
|
| 262 |
+
bbox_w, bbox_h,
|
| 263 |
+
]
|
| 264 |
+
return global_features, side_features
|
code/forge_LAR_2mosaic/batch.py
ADDED
|
@@ -0,0 +1,67 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
"""Forge locale LAR — dossier d'images -> dataset_inputs/canvas_mosaic_XXX.png
|
| 3 |
+
+ piece_grid_XXX.json (numérotés, prêts pour `mosaic2fragments/batch.py`).
|
| 4 |
+
|
| 5 |
+
Exemple :
|
| 6 |
+
python3 forge_LAR_2mosaic/batch.py --input-dir mes_images --out-dir dataset_inputs \
|
| 7 |
+
--grid 96 --stud-size 40
|
| 8 |
+
|
| 9 |
+
Puis la pipeline aval consomme tout :
|
| 10 |
+
python3 mosaic2fragments/batch.py --inputs dataset_inputs/canvas_mosaic_*.png --out dataset
|
| 11 |
+
"""
|
| 12 |
+
import argparse
|
| 13 |
+
from pathlib import Path
|
| 14 |
+
|
| 15 |
+
from mosaic import forge_to_files
|
| 16 |
+
|
| 17 |
+
EXTS = {".png", ".jpg", ".jpeg", ".bmp", ".webp", ".tif", ".tiff", ".gif"}
|
| 18 |
+
|
| 19 |
+
|
| 20 |
+
def parse_args():
|
| 21 |
+
p = argparse.ArgumentParser(description="Dossier d'images -> mosaïques LEGO + GT.")
|
| 22 |
+
p.add_argument("--input-dir", required=True, help="dossier d'images sources")
|
| 23 |
+
p.add_argument("--out-dir", default="dataset_inputs", help="dossier de sortie")
|
| 24 |
+
p.add_argument("--grid", type=int, default=96, help="studs par côté")
|
| 25 |
+
p.add_argument("--stud-size", type=int, default=40, help="px par stud (≥40)")
|
| 26 |
+
p.add_argument("--joint-px", type=int, default=None, help="largeur du joint (défaut ~3)")
|
| 27 |
+
p.add_argument("--mode", choices=["tile", "plate", "brick", "mono"], default="tile",
|
| 28 |
+
help="jeu de pièces LEGO (tile par défaut, max 2×6 ; plate = jusqu'à 4×4/4×10)")
|
| 29 |
+
p.add_argument("--big-plates", action="store_true",
|
| 30 |
+
help="réactive 4×8/4×10 (décochées par défaut sur le site)")
|
| 31 |
+
p.add_argument("--start-index", type=int, default=0, help="indice de départ du nom")
|
| 32 |
+
p.add_argument("--max-dominant-frac", type=float, default=None,
|
| 33 |
+
help="garde-fou : rejeter une image si une couleur LEGO couvre > cette "
|
| 34 |
+
"fraction des cellules (aplats monochromes). None = désactivé")
|
| 35 |
+
return p.parse_args()
|
| 36 |
+
|
| 37 |
+
|
| 38 |
+
def main():
|
| 39 |
+
a = parse_args()
|
| 40 |
+
imgs = sorted(p for p in Path(a.input_dir).iterdir()
|
| 41 |
+
if p.suffix.lower() in EXTS)
|
| 42 |
+
if not imgs:
|
| 43 |
+
raise SystemExit(f"Aucune image dans {a.input_dir} (extensions {sorted(EXTS)})")
|
| 44 |
+
out = Path(a.out_dir)
|
| 45 |
+
out.mkdir(parents=True, exist_ok=True)
|
| 46 |
+
print(f"{len(imgs)} image(s) -> {out}/")
|
| 47 |
+
made = rejected = 0
|
| 48 |
+
for k, src in enumerate(imgs, start=a.start_index):
|
| 49 |
+
png = out / f"canvas_mosaic_{k:03d}.png"
|
| 50 |
+
js = out / f"piece_grid_{k:03d}.json"
|
| 51 |
+
g = forge_to_files(str(src), str(png), str(js),
|
| 52 |
+
grid_w=a.grid, grid_h=a.grid,
|
| 53 |
+
stud_size=a.stud_size, joint_px=a.joint_px,
|
| 54 |
+
mode=a.mode, big_plates=a.big_plates,
|
| 55 |
+
max_dominant_frac=a.max_dominant_frac)
|
| 56 |
+
if g.get("rejected"): # garde-fou : aplat monochrome
|
| 57 |
+
rejected += 1
|
| 58 |
+
print(f" [{k:03d}] {src.name:30s} -> REJETÉ (couleur dominante {g['dominant_frac']:.0%})")
|
| 59 |
+
continue
|
| 60 |
+
made += 1
|
| 61 |
+
print(f" [{k:03d}] {src.name:30s} -> {png.name} ({g['n_pieces']} pièces)")
|
| 62 |
+
print(f"Fait : {made} mosaïques (rejetées par garde-fou : {rejected}).")
|
| 63 |
+
print(f"Ensuite : python3 forge_mosaics/mosaic2fragments/batch.py --inputs {out}/canvas_mosaic_*.png --out dataset")
|
| 64 |
+
|
| 65 |
+
|
| 66 |
+
if __name__ == "__main__":
|
| 67 |
+
main()
|
code/forge_LAR_2mosaic/cli.py
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
"""Forge locale LAR — une image -> canvas_mosaic.png + piece_grid.json.
|
| 3 |
+
|
| 4 |
+
Exemple :
|
| 5 |
+
python3 forge_LAR_2mosaic/cli.py --input photo.jpg --out-png out/canvas_mosaic.png \
|
| 6 |
+
--out-json out/piece_grid.json --grid 96 --stud-size 40
|
| 7 |
+
"""
|
| 8 |
+
import argparse
|
| 9 |
+
from pathlib import Path
|
| 10 |
+
|
| 11 |
+
from mosaic import forge_to_files
|
| 12 |
+
|
| 13 |
+
|
| 14 |
+
def parse_args():
|
| 15 |
+
p = argparse.ArgumentParser(description="Image -> mosaïque LEGO (tile/plate/brick) + GT.")
|
| 16 |
+
p.add_argument("--input", required=True, help="image source (jpg/png/…)")
|
| 17 |
+
p.add_argument("--out-png", help="chemin du canvas_mosaic.png (défaut : à côté de --input)")
|
| 18 |
+
p.add_argument("--out-json", help="chemin du piece_grid.json")
|
| 19 |
+
p.add_argument("--grid", type=int, default=96, help="studs par côté (48/64/96/100/128)")
|
| 20 |
+
p.add_argument("--stud-size", type=int, default=40, help="px par stud (≥40)")
|
| 21 |
+
p.add_argument("--joint-px", type=int, default=None, help="largeur du joint (défaut ~3)")
|
| 22 |
+
p.add_argument("--mode", choices=["tile", "plate", "brick", "mono"], default="tile",
|
| 23 |
+
help="jeu de pièces LEGO (tile par défaut ; mono = 1×1 seul, ultracompact)")
|
| 24 |
+
p.add_argument("--big-plates", action="store_true",
|
| 25 |
+
help="réactive 4×8/4×10 (décochées par défaut sur le site)")
|
| 26 |
+
return p.parse_args()
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
def main():
|
| 30 |
+
a = parse_args()
|
| 31 |
+
out_png = a.out_png or str(Path(a.input).with_name("canvas_mosaic.png"))
|
| 32 |
+
out_json = a.out_json or str(Path(out_png).with_name("piece_grid.json"))
|
| 33 |
+
grid = forge_to_files(
|
| 34 |
+
a.input, out_png, out_json,
|
| 35 |
+
grid_w=a.grid, grid_h=a.grid, stud_size=a.stud_size, joint_px=a.joint_px,
|
| 36 |
+
mode=a.mode, big_plates=a.big_plates,
|
| 37 |
+
)
|
| 38 |
+
print(f"{out_png} ({grid['grid_width']}×{grid['grid_height']} studs, mode={grid['mode']}, "
|
| 39 |
+
f"max {grid['max_piece'][0]}×{grid['max_piece'][1]}, "
|
| 40 |
+
f"{grid['n_pieces']} pièces, joint {grid['joint_width_px']}px)")
|
| 41 |
+
print(f"{out_json}")
|
| 42 |
+
|
| 43 |
+
|
| 44 |
+
if __name__ == "__main__":
|
| 45 |
+
main()
|
code/forge_LAR_2mosaic/mosaic.py
ADDED
|
@@ -0,0 +1,232 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Forge de mosaïques LEGO « variable-tile/plate/brick » — clean-room, Python pur.
|
| 2 |
+
|
| 3 |
+
Remplace lego-art-remix.com pour notre usage : image quelconque → mosaïque LEGO
|
| 4 |
+
à pièces plates de tailles variables (1×1 … 4×10) avec joints gris, PLUS la
|
| 5 |
+
vérité-terrain exacte des pièces (`piece_grid.json`).
|
| 6 |
+
|
| 7 |
+
On NE copie aucun code de l'outil source (GPL-3.0) : on réimplémente la logique
|
| 8 |
+
publique (quantif vers palette + packing glouton du plus grand rectangle). Les
|
| 9 |
+
JEUX DE DIMENSIONS et n° de pièce BrickLink par mode sont des FAITS extraits de
|
| 10 |
+
`algo.js` (TILE/PLATE/BRICK_DIMENSIONS_TO_PART_ID). Sortie au FORMAT attendu par
|
| 11 |
+
`mosaic2fragments/forge_dataset.py` (carré, ≥40 px/stud, joints ~RGB(136), palette sans gris).
|
| 12 |
+
|
| 13 |
+
Pipeline :
|
| 14 |
+
1. crop carré + downsample area-average à grid×grid (1 couleur/cellule)
|
| 15 |
+
2. quantif nearest-color en espace Lab vers la palette LEGO (sans gris)
|
| 16 |
+
3. packing glouton : plus grand rectangle MONOCHROME d'abord (aire ↓, comme le
|
| 17 |
+
site) ; reste comblé en 1×1
|
| 18 |
+
4. rendu : pièces plates pleines + joints gris ; export PNG + piece_grid.json
|
| 19 |
+
|
| 20 |
+
Déterministe (aucun aléa) : même image + mêmes params → même sortie.
|
| 21 |
+
"""
|
| 22 |
+
import json
|
| 23 |
+
from pathlib import Path
|
| 24 |
+
|
| 25 |
+
import numpy as np
|
| 26 |
+
from PIL import Image
|
| 27 |
+
|
| 28 |
+
from palette import LEGO_PALETTE, JOINT_COLOR, hex_of
|
| 29 |
+
|
| 30 |
+
|
| 31 |
+
# ============================================================
|
| 32 |
+
# Jeux de dimensions LEGO par mode (studs) + n° de pièce BrickLink.
|
| 33 |
+
# Faits extraits de lego-art-remix/app/js/algo.js (clean-room).
|
| 34 |
+
# ============================================================
|
| 35 |
+
|
| 36 |
+
TILE_DIMS = { # "Variable Tile" (plat, lisse)
|
| 37 |
+
(1, 1): "3070b", (1, 2): "3069b", (1, 3): "63864", (1, 4): "2431",
|
| 38 |
+
(1, 6): "6636", (1, 8): "4162", (2, 2): "3068b", (2, 3): "26603",
|
| 39 |
+
(2, 4): "87079", (2, 6): "69729",
|
| 40 |
+
}
|
| 41 |
+
PLATE_DIMS = { # "Variable Plate" (le + de tailles)
|
| 42 |
+
(1, 1): "3024", (1, 2): "3023", (1, 3): "3623", (1, 4): "3710",
|
| 43 |
+
(1, 6): "3666", (1, 8): "3460", (2, 2): "3022", (2, 3): "3021",
|
| 44 |
+
(2, 4): "3020", (2, 6): "3795", (2, 8): "3034", (4, 4): "3031",
|
| 45 |
+
(4, 8): "3035", (4, 10): "3030",
|
| 46 |
+
}
|
| 47 |
+
BRICK_DIMS = { # "Variable Brick"
|
| 48 |
+
(1, 1): "3005", (1, 2): "3004", (1, 3): "3622", (1, 4): "3010",
|
| 49 |
+
(1, 6): "3009", (1, 8): "3008", (2, 2): "3003", (2, 3): "3002",
|
| 50 |
+
(2, 4): "3001", (2, 6): "2456", (2, 8): "3007",
|
| 51 |
+
}
|
| 52 |
+
MONO_DIMS = {(1, 1): "3070b"} # 1×1 SEUL (mode ultracompact)
|
| 53 |
+
MODE_DIMS = {"tile": TILE_DIMS, "plate": PLATE_DIMS, "brick": BRICK_DIMS, "mono": MONO_DIMS}
|
| 54 |
+
# Le site DÉCOCHE par défaut les plus grosses plates (UI : DEFAULT_DISABLED_DEPTH_PLATES).
|
| 55 |
+
DEFAULT_DISABLED = {(4, 8), (4, 10)}
|
| 56 |
+
|
| 57 |
+
|
| 58 |
+
def build_allowed_parts(mode="tile", big_plates=False):
|
| 59 |
+
"""Liste (w, h, part_id) pour le mode donné, 2 orientations, triée comme le
|
| 60 |
+
site : aire décroissante puis plus petite 1re dimension d'abord (déterministe).
|
| 61 |
+
`big_plates=True` réactive 4×8/4×10 (décochées par défaut sur le site)."""
|
| 62 |
+
dims = dict(MODE_DIMS[mode])
|
| 63 |
+
if not big_plates:
|
| 64 |
+
for d in DEFAULT_DISABLED:
|
| 65 |
+
dims.pop(d, None)
|
| 66 |
+
seen, parts = set(), []
|
| 67 |
+
for (a, b), pid in dims.items():
|
| 68 |
+
for w, h in {(a, b), (b, a)}:
|
| 69 |
+
if (w, h) not in seen:
|
| 70 |
+
seen.add((w, h))
|
| 71 |
+
parts.append((w, h, pid))
|
| 72 |
+
parts.sort(key=lambda p: (p[0] * p[1], -p[0]), reverse=True)
|
| 73 |
+
return parts
|
| 74 |
+
|
| 75 |
+
|
| 76 |
+
# ============================================================
|
| 77 |
+
# Espace couleur : sRGB [0,255] -> CIE Lab (D65)
|
| 78 |
+
# ============================================================
|
| 79 |
+
|
| 80 |
+
def srgb_to_lab(rgb):
|
| 81 |
+
"""rgb : (..., 3) en [0,255] -> Lab (..., 3). Pur numpy (pas de dépendance)."""
|
| 82 |
+
c = np.asarray(rgb, dtype=np.float64) / 255.0
|
| 83 |
+
lin = np.where(c <= 0.04045, c / 12.92, ((c + 0.055) / 1.055) ** 2.4)
|
| 84 |
+
r, g, b = lin[..., 0], lin[..., 1], lin[..., 2]
|
| 85 |
+
x = 0.4124 * r + 0.3576 * g + 0.1805 * b
|
| 86 |
+
y = 0.2126 * r + 0.7152 * g + 0.0722 * b
|
| 87 |
+
z = 0.0193 * r + 0.1192 * g + 0.9505 * b
|
| 88 |
+
x /= 0.95047; z /= 1.08883 # point blanc D65 (Yn = 1)
|
| 89 |
+
def f(t):
|
| 90 |
+
return np.where(t > 0.008856, np.cbrt(t), 7.787 * t + 16.0 / 116.0)
|
| 91 |
+
fx, fy, fz = f(x), f(y), f(z)
|
| 92 |
+
return np.stack([116.0 * fy - 16.0, 500.0 * (fx - fy), 200.0 * (fy - fz)], axis=-1)
|
| 93 |
+
|
| 94 |
+
|
| 95 |
+
# ============================================================
|
| 96 |
+
# Étapes 1-2 : image -> grille d'indices de palette
|
| 97 |
+
# ============================================================
|
| 98 |
+
|
| 99 |
+
def image_to_color_grid(image, grid_w, grid_h):
|
| 100 |
+
"""Crop carré centré + downsample area-average -> (grid_h, grid_w, 3) uint8.
|
| 101 |
+
`image` : chemin OU objet PIL.Image (utile pour le streaming wikiart)."""
|
| 102 |
+
im = (image if isinstance(image, Image.Image) else Image.open(image)).convert("RGB")
|
| 103 |
+
w, h = im.size
|
| 104 |
+
s = min(w, h) # crop carré centré
|
| 105 |
+
im = im.crop(((w - s) // 2, (h - s) // 2, (w - s) // 2 + s, (h - s) // 2 + s))
|
| 106 |
+
im = im.resize((grid_w, grid_h), Image.BOX) # moyenne par cellule
|
| 107 |
+
return np.asarray(im, dtype=np.uint8)
|
| 108 |
+
|
| 109 |
+
|
| 110 |
+
def quantize_to_palette(color_grid, palette=LEGO_PALETTE):
|
| 111 |
+
"""(H,W,3) -> (H,W) indices de la palette (nearest en Lab)."""
|
| 112 |
+
pal = np.array(palette, dtype=np.float64)
|
| 113 |
+
pal_lab = srgb_to_lab(pal) # (P,3)
|
| 114 |
+
cells_lab = srgb_to_lab(color_grid.reshape(-1, 3)) # (H*W,3)
|
| 115 |
+
d = ((cells_lab[:, None, :] - pal_lab[None, :, :]) ** 2).sum(-1) # (H*W,P)
|
| 116 |
+
idx = d.argmin(1).reshape(color_grid.shape[:2])
|
| 117 |
+
return idx.astype(np.int32)
|
| 118 |
+
|
| 119 |
+
|
| 120 |
+
# ============================================================
|
| 121 |
+
# Étape 3 : packing glouton du plus grand rectangle monochrome
|
| 122 |
+
# ============================================================
|
| 123 |
+
|
| 124 |
+
def pack_variable_tiles(idx_grid, allowed_parts):
|
| 125 |
+
"""Glouton : pose le plus grand rectangle MONOCHROME possible, sans
|
| 126 |
+
chevauchement, en balayant en row-major ; reste comblé par des 1×1.
|
| 127 |
+
|
| 128 |
+
`allowed_parts` : liste (w, h, part_id) déjà triée (aire ↓). Retourne la
|
| 129 |
+
liste des pièces (col, row, width, height, color_idx, part_id). Chaque pièce
|
| 130 |
+
est mono-couleur (une pièce LEGO = une couleur) ; deux pièces voisines de même
|
| 131 |
+
couleur restent séparées par un joint -> deux pièces distinctes en aval.
|
| 132 |
+
"""
|
| 133 |
+
H, W = idx_grid.shape
|
| 134 |
+
occupied = np.zeros((H, W), dtype=bool)
|
| 135 |
+
pieces = []
|
| 136 |
+
for pw, ph, pid in allowed_parts:
|
| 137 |
+
if pw > W or ph > H:
|
| 138 |
+
continue
|
| 139 |
+
for r in range(H - ph + 1):
|
| 140 |
+
for c in range(W - pw + 1):
|
| 141 |
+
if occupied[r:r + ph, c:c + pw].any():
|
| 142 |
+
continue
|
| 143 |
+
block = idx_grid[r:r + ph, c:c + pw]
|
| 144 |
+
if (block == block[0, 0]).all():
|
| 145 |
+
occupied[r:r + ph, c:c + pw] = True
|
| 146 |
+
pieces.append((c, r, pw, ph, int(block[0, 0]), pid))
|
| 147 |
+
# garde-fou : toute cellule restante -> 1×1 (ne devrait pas arriver)
|
| 148 |
+
unit_pid = next((pid for w, h, pid in allowed_parts if (w, h) == (1, 1)), None)
|
| 149 |
+
for r, c in np.argwhere(~occupied):
|
| 150 |
+
pieces.append((int(c), int(r), 1, 1, int(idx_grid[r, c]), unit_pid))
|
| 151 |
+
return pieces
|
| 152 |
+
|
| 153 |
+
|
| 154 |
+
# ============================================================
|
| 155 |
+
# Étape 4 : rendu PNG (pièces plates + joints gris)
|
| 156 |
+
# ============================================================
|
| 157 |
+
|
| 158 |
+
def render_mosaic(pieces, grid_w, grid_h, stud_size, joint_px, palette=LEGO_PALETTE):
|
| 159 |
+
"""Fond gris (joints) + chaque pièce peinte en retrait de joint_px/2 sur
|
| 160 |
+
chaque bord -> joint de joint_px px entre pièces voisines. Aucun anti-alias :
|
| 161 |
+
pixels de pièce = couleur palette exacte, pixels de joint = RGB(136) exact."""
|
| 162 |
+
H_px, W_px = grid_h * stud_size, grid_w * stud_size
|
| 163 |
+
img = np.empty((H_px, W_px, 3), dtype=np.uint8)
|
| 164 |
+
img[:] = JOINT_COLOR
|
| 165 |
+
li, ri = joint_px - joint_px // 2, joint_px // 2 # ceil / floor -> joint exact
|
| 166 |
+
for c, r, w, h, ci, _pid in pieces:
|
| 167 |
+
x0, y0 = c * stud_size + li, r * stud_size + li
|
| 168 |
+
x1, y1 = (c + w) * stud_size - ri, (r + h) * stud_size - ri
|
| 169 |
+
img[y0:y1, x0:x1] = palette[ci]
|
| 170 |
+
return Image.fromarray(img, "RGB")
|
| 171 |
+
|
| 172 |
+
|
| 173 |
+
# ============================================================
|
| 174 |
+
# Orchestration
|
| 175 |
+
# ============================================================
|
| 176 |
+
|
| 177 |
+
def forge_mosaic(image, grid_w=96, grid_h=96, stud_size=40, joint_px=None,
|
| 178 |
+
mode="tile", big_plates=False, palette=LEGO_PALETTE, source_name=None,
|
| 179 |
+
max_dominant_frac=None):
|
| 180 |
+
"""image (chemin ou PIL.Image) -> (PIL.Image mosaïque, dict piece_grid). mode: tile|plate|brick.
|
| 181 |
+
|
| 182 |
+
Garde-fou `max_dominant_frac` (∈ ]0,1]) : si une seule couleur de palette couvre
|
| 183 |
+
> cette fraction des cellules, l'image est REJETÉE → renvoie `(None, stats)`.
|
| 184 |
+
Cible : les peintures (souvent impressionnistes) à variations infimes d'une même
|
| 185 |
+
couleur, qui s'effondrent en un gros aplat monochrome une fois quantifiées (peu de
|
| 186 |
+
pièces, détection de joints faussée). None par défaut = garde-fou désactivé.
|
| 187 |
+
"""
|
| 188 |
+
if joint_px is None:
|
| 189 |
+
joint_px = max(2, round(stud_size * 0.075)) # ~3 px à 40 px/stud
|
| 190 |
+
allowed = build_allowed_parts(mode, big_plates)
|
| 191 |
+
color_grid = image_to_color_grid(image, grid_w, grid_h)
|
| 192 |
+
idx_grid = quantize_to_palette(color_grid, palette)
|
| 193 |
+
counts = np.bincount(idx_grid.ravel(), minlength=len(palette))
|
| 194 |
+
dominant_frac = float(counts.max() / idx_grid.size)
|
| 195 |
+
n_colors = int((counts > 0).sum())
|
| 196 |
+
if max_dominant_frac is not None and dominant_frac > max_dominant_frac:
|
| 197 |
+
return None, {"rejected": True, "reason": "dominant_color",
|
| 198 |
+
"dominant_frac": dominant_frac, "n_colors": n_colors,
|
| 199 |
+
"max_dominant_frac": max_dominant_frac}
|
| 200 |
+
pieces = pack_variable_tiles(idx_grid, allowed)
|
| 201 |
+
img = render_mosaic(pieces, grid_w, grid_h, stud_size, joint_px, palette)
|
| 202 |
+
piece_grid = {
|
| 203 |
+
"grid_width": grid_w,
|
| 204 |
+
"grid_height": grid_h,
|
| 205 |
+
"stud_size_px": stud_size,
|
| 206 |
+
"joint_width_px": joint_px,
|
| 207 |
+
"joint_color_rgb": list(JOINT_COLOR),
|
| 208 |
+
"mode": mode,
|
| 209 |
+
"max_piece": list(max((p[:2] for p in allowed), key=lambda d: d[0] * d[1])),
|
| 210 |
+
"source_image": source_name or (Path(image).name if isinstance(image, (str, Path)) else "image"),
|
| 211 |
+
"dominant_frac": dominant_frac,
|
| 212 |
+
"n_colors": n_colors,
|
| 213 |
+
"n_pieces": len(pieces),
|
| 214 |
+
"n_cells": grid_w * grid_h,
|
| 215 |
+
"pieces": [
|
| 216 |
+
{"col": c, "row": r, "width": w, "height": h,
|
| 217 |
+
"part_id": pid, "color_rgb": list(palette[ci]),
|
| 218 |
+
"color_hex": hex_of(palette[ci])}
|
| 219 |
+
for c, r, w, h, ci, pid in pieces
|
| 220 |
+
],
|
| 221 |
+
}
|
| 222 |
+
return img, piece_grid
|
| 223 |
+
|
| 224 |
+
|
| 225 |
+
def forge_to_files(image_path, out_png, out_json, **kw):
|
| 226 |
+
img, grid = forge_mosaic(image_path, **kw)
|
| 227 |
+
if img is None: # rejeté par le garde-fou
|
| 228 |
+
return grid # stats, rien écrit
|
| 229 |
+
Path(out_png).parent.mkdir(parents=True, exist_ok=True)
|
| 230 |
+
img.save(out_png)
|
| 231 |
+
Path(out_json).write_text(json.dumps(grid, indent=2))
|
| 232 |
+
return grid
|
code/forge_LAR_2mosaic/palette.py
ADDED
|
@@ -0,0 +1,103 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Palette LEGO — couleurs solides BrickLink, SANS les gris/argentés.
|
| 2 |
+
|
| 3 |
+
Les valeurs hex sont des FAITS (couleurs LEGO officielles, non protégeables) ;
|
| 4 |
+
elles ont été extraites de `bricklink-colors.js` (lego-art-remix, GPL-3.0) puis
|
| 5 |
+
filtrées hors-ligne. AUCUN code de l'outil source n'est importé ici (clean-room).
|
| 6 |
+
|
| 7 |
+
Filtre appliqué : on retire toute couleur « confusable joint », i.e. quasi-grise
|
| 8 |
+
(max-min des canaux < 25) ET d'intensité moyenne (70 ≤ moyenne ≤ 200). C'est la
|
| 9 |
+
fenêtre où le détecteur de joints de `mosaic2fragments/` (gris ~RGB 136 ± 30) ferait un
|
| 10 |
+
faux positif. Blanc (255) et Noir (33) sont conservés (loin de 136).
|
| 11 |
+
"""
|
| 12 |
+
|
| 13 |
+
# (R, G, B) — 82 couleurs LEGO solides, gris/argent retirés.
|
| 14 |
+
LEGO_PALETTE = [
|
| 15 |
+
(255, 255, 255), # White
|
| 16 |
+
(232, 232, 232), # Very Light Gray
|
| 17 |
+
(228, 232, 232), # Very Light Bluish Gray
|
| 18 |
+
( 33, 33, 33), # Black
|
| 19 |
+
(106, 14, 21), # Dark Red
|
| 20 |
+
(179, 0, 6), # Red
|
| 21 |
+
(181, 44, 32), # Rust
|
| 22 |
+
(248, 131, 121), # Coral
|
| 23 |
+
(244, 92, 64), # Salmon
|
| 24 |
+
(255, 222, 220), # Light Salmon
|
| 25 |
+
(140, 107, 107), # Sand Red
|
| 26 |
+
(137, 53, 29), # Reddish Brown
|
| 27 |
+
( 83, 33, 21), # Brown
|
| 28 |
+
( 51, 0, 0), # Dark Brown
|
| 29 |
+
(144, 116, 80), # Dark Tan
|
| 30 |
+
(222, 198, 156), # Tan
|
| 31 |
+
(254, 204, 176), # Light Nougat
|
| 32 |
+
(255, 175, 125), # Nougat
|
| 33 |
+
(227, 160, 91), # Medium Nougat
|
| 34 |
+
(231, 139, 62), # Dark Nougat
|
| 35 |
+
(161, 108, 66), # Medium Brown
|
| 36 |
+
(179, 105, 78), # Fabuland Brown
|
| 37 |
+
(239, 145, 33), # Fabuland Orange
|
| 38 |
+
(230, 136, 29), # Earth Orange
|
| 39 |
+
(179, 84, 8), # Dark Orange
|
| 40 |
+
(250, 89, 71), # Neon Orange
|
| 41 |
+
(255, 126, 20), # Orange
|
| 42 |
+
(255, 165, 49), # Medium Orange
|
| 43 |
+
(247, 186, 48), # Bright Light Orange
|
| 44 |
+
(247, 173, 99), # Light Orange
|
| 45 |
+
(230, 192, 93), # Very Light Orange
|
| 46 |
+
(221, 152, 46), # Dark Yellow
|
| 47 |
+
(247, 209, 23), # Yellow
|
| 48 |
+
(243, 224, 85), # Bright Light Yellow
|
| 49 |
+
(255, 227, 131), # Light Yellow
|
| 50 |
+
(235, 238, 143), # Light Lime
|
| 51 |
+
(223, 238, 165), # Yellowish Green
|
| 52 |
+
(188, 239, 102), # Neon Green
|
| 53 |
+
(189, 198, 24), # Medium Lime
|
| 54 |
+
(166, 202, 85), # Lime
|
| 55 |
+
(124, 144, 81), # Olive Green
|
| 56 |
+
( 46, 85, 67), # Dark Green
|
| 57 |
+
( 0, 100, 46), # Green
|
| 58 |
+
( 16, 203, 49), # Bright Green
|
| 59 |
+
( 98, 245, 142), # Medium Green
|
| 60 |
+
(165, 219, 181), # Light Green
|
| 61 |
+
(118, 162, 144), # Sand Green
|
| 62 |
+
( 0, 138, 128), # Dark Turquoise
|
| 63 |
+
( 49, 181, 202), # Light Turquoise
|
| 64 |
+
(181, 211, 214), # Aqua
|
| 65 |
+
(204, 255, 255), # Light Aqua
|
| 66 |
+
( 20, 48, 68), # Dark Blue
|
| 67 |
+
( 0, 87, 166), # Blue
|
| 68 |
+
( 51, 153, 255), # Dark Azure
|
| 69 |
+
( 66, 192, 251), # Medium Azure
|
| 70 |
+
( 97, 175, 255), # Medium Blue
|
| 71 |
+
(107, 173, 214), # Maersk Blue
|
| 72 |
+
(159, 195, 233), # Bright Light Blue
|
| 73 |
+
(180, 210, 227), # Light Blue
|
| 74 |
+
(125, 191, 221), # Sky Blue
|
| 75 |
+
( 90, 113, 132), # Sand Blue
|
| 76 |
+
( 80, 108, 239), # Blue-Violet
|
| 77 |
+
( 32, 50, 176), # Dark Blue-Violet
|
| 78 |
+
( 52, 72, 164), # Violet
|
| 79 |
+
(147, 145, 228), # Medium Violet
|
| 80 |
+
(201, 202, 226), # Light Violet
|
| 81 |
+
( 95, 38, 131), # Dark Purple
|
| 82 |
+
(165, 73, 156), # Purple
|
| 83 |
+
(218, 112, 214), # Light Purple
|
| 84 |
+
(136, 94, 158), # Medium Lavender
|
| 85 |
+
(224, 170, 217), # Clikits Lavender
|
| 86 |
+
(177, 140, 191), # Lavender
|
| 87 |
+
(181, 125, 165), # Sand Purple
|
| 88 |
+
(181, 41, 82), # Magenta
|
| 89 |
+
(200, 112, 128), # Dark Pink
|
| 90 |
+
(247, 133, 177), # Medium Dark Pink
|
| 91 |
+
(255, 187, 255), # Bright Pink
|
| 92 |
+
(255, 192, 203), # Pink
|
| 93 |
+
(255, 225, 255), # Light Pink
|
| 94 |
+
(255, 252, 0), # Neon Yellow
|
| 95 |
+
(231, 149, 0), # Pearl Gold
|
| 96 |
+
(217, 197, 148), # Medium Tan
|
| 97 |
+
]
|
| 98 |
+
|
| 99 |
+
JOINT_COLOR = (136, 136, 136) # gris des joints — RGB(136) ≈ rendu lego-art-remix
|
| 100 |
+
|
| 101 |
+
|
| 102 |
+
def hex_of(rgb):
|
| 103 |
+
return "#{:02x}{:02x}{:02x}".format(*rgb)
|
code/forge_LAR_2mosaic/remono.py
ADDED
|
@@ -0,0 +1,84 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
"""Re-forge une mosaïque existante en mode MONO (tuiles 1×1) — pour `ultracompact`.
|
| 3 |
+
|
| 4 |
+
Entrée = le `piece_grid_<uuid>.json` d'une mosaïque déjà forgée (n'importe quel
|
| 5 |
+
mode) : il porte la couleur EXACTE de chaque cellule → on reconstruit la même
|
| 6 |
+
image de mosaïque, mais où chaque cellule est une pièce 1×1 (joints partout).
|
| 7 |
+
Aucune re-quantification (repasser le canvas rendu dans la forge moyennerait les
|
| 8 |
+
joints gris dans les cellules) ; l'uuid est conservé → l'appariement des
|
| 9 |
+
curriculums (même mosaïque de base) reste valable entre variable-tile et mono.
|
| 10 |
+
|
| 11 |
+
Usage :
|
| 12 |
+
python3 forge_mosaics/forge_LAR_2mosaic/remono.py \
|
| 13 |
+
--piece-grids 'output/wikiart_inputs/piece_grid_*.json' \
|
| 14 |
+
--out-dir output/wikiart_inputs_mono [--limit 100]
|
| 15 |
+
"""
|
| 16 |
+
import argparse
|
| 17 |
+
import glob
|
| 18 |
+
import json
|
| 19 |
+
from pathlib import Path
|
| 20 |
+
|
| 21 |
+
import numpy as np
|
| 22 |
+
|
| 23 |
+
import sys
|
| 24 |
+
sys.path.insert(0, str(Path(__file__).parent))
|
| 25 |
+
from mosaic import render_mosaic, MONO_DIMS, JOINT_COLOR # noqa: E402
|
| 26 |
+
|
| 27 |
+
|
| 28 |
+
def mono_from_piece_grid(pg):
|
| 29 |
+
"""piece_grid dict (mode quelconque) -> (PIL.Image mono, piece_grid mono)."""
|
| 30 |
+
W, H = pg["grid_width"], pg["grid_height"]
|
| 31 |
+
stud, joint = pg["stud_size_px"], pg["joint_width_px"]
|
| 32 |
+
# 1) grille de couleurs exacte depuis les pièces (couverture 1× garantie)
|
| 33 |
+
grid = np.zeros((H, W, 3), dtype=np.uint8)
|
| 34 |
+
for p in pg["pieces"]:
|
| 35 |
+
c, r, w, h = p["col"], p["row"], p["width"], p["height"]
|
| 36 |
+
grid[r:r + h, c:c + w] = p["color_rgb"]
|
| 37 |
+
# 2) palette locale = couleurs distinctes de CETTE mosaïque (rendu exact)
|
| 38 |
+
colors = np.unique(grid.reshape(-1, 3), axis=0)
|
| 39 |
+
lut = {tuple(c): i for i, c in enumerate(colors)}
|
| 40 |
+
unit_pid = MONO_DIMS[(1, 1)]
|
| 41 |
+
pieces = [(c, r, 1, 1, lut[tuple(grid[r, c])], unit_pid)
|
| 42 |
+
for r in range(H) for c in range(W)]
|
| 43 |
+
img = render_mosaic(pieces, W, H, stud, joint, palette=colors)
|
| 44 |
+
piece_grid = {
|
| 45 |
+
**{k: pg[k] for k in ("grid_width", "grid_height", "stud_size_px",
|
| 46 |
+
"joint_width_px", "source_image")},
|
| 47 |
+
"joint_color_rgb": list(JOINT_COLOR),
|
| 48 |
+
"mode": "mono", "max_piece": [1, 1],
|
| 49 |
+
"remono_from_mode": pg.get("mode"),
|
| 50 |
+
"n_pieces": len(pieces), "n_cells": W * H,
|
| 51 |
+
"pieces": [{"col": c, "row": r, "width": 1, "height": 1, "part_id": pid,
|
| 52 |
+
"color_rgb": [int(v) for v in colors[ci]]}
|
| 53 |
+
for c, r, _w, _h, ci, pid in pieces],
|
| 54 |
+
}
|
| 55 |
+
return img, piece_grid
|
| 56 |
+
|
| 57 |
+
|
| 58 |
+
def main():
|
| 59 |
+
ap = argparse.ArgumentParser(description=__doc__)
|
| 60 |
+
ap.add_argument("--piece-grids", required=True,
|
| 61 |
+
help="glob des piece_grid_*.json source")
|
| 62 |
+
ap.add_argument("--out-dir", required=True)
|
| 63 |
+
ap.add_argument("--limit", type=int, default=None,
|
| 64 |
+
help="ne traiter que les N premiers (ordre trié)")
|
| 65 |
+
a = ap.parse_args()
|
| 66 |
+
out = Path(a.out_dir)
|
| 67 |
+
out.mkdir(parents=True, exist_ok=True)
|
| 68 |
+
paths = sorted(glob.glob(a.piece_grids))[: a.limit]
|
| 69 |
+
for i, jp in enumerate(paths, 1):
|
| 70 |
+
pg = json.load(open(jp))
|
| 71 |
+
uid = Path(jp).stem[len("piece_grid_"):]
|
| 72 |
+
png = out / f"canvas_mosaic_{uid}.png"
|
| 73 |
+
if png.exists(): # resume
|
| 74 |
+
continue
|
| 75 |
+
img, mono = mono_from_piece_grid(pg)
|
| 76 |
+
img.save(png)
|
| 77 |
+
json.dump(mono, open(out / f"piece_grid_{uid}.json", "w"))
|
| 78 |
+
if i % 20 == 0 or i == len(paths):
|
| 79 |
+
print(f"[{i}/{len(paths)}] {uid}")
|
| 80 |
+
print(f"done → {out}")
|
| 81 |
+
|
| 82 |
+
|
| 83 |
+
if __name__ == "__main__":
|
| 84 |
+
main()
|
code/forge_LAR_2mosaic/wikiart.py
ADDED
|
@@ -0,0 +1,87 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
"""Stream wikiart (HuggingFace) → mosaïques LEGO, SANS télécharger le dataset.
|
| 3 |
+
|
| 4 |
+
Tire N images de `huggan/wikiart` à la volée et les passe dans la forge locale →
|
| 5 |
+
`dataset_inputs/canvas_mosaic_XXX.png` + `piece_grid_XXX.json` (prêts pour
|
| 6 |
+
`mosaic2fragments`). Aucun téléchargement complet (~1-2 Mo/image en flux).
|
| 7 |
+
|
| 8 |
+
Exemple :
|
| 9 |
+
python3 forge_LAR_2mosaic/wikiart.py --n 200 --out-dir dataset_inputs --grid 96 --mode plate
|
| 10 |
+
python3 mosaic2fragments/batch.py --inputs dataset_inputs/canvas_mosaic_*.png --out dataset_wikiart
|
| 11 |
+
|
| 12 |
+
Dépendance : `pip install datasets` (HuggingFace). Réseau requis.
|
| 13 |
+
"""
|
| 14 |
+
import argparse
|
| 15 |
+
import json
|
| 16 |
+
import uuid
|
| 17 |
+
from itertools import islice
|
| 18 |
+
from pathlib import Path
|
| 19 |
+
|
| 20 |
+
from mosaic import forge_mosaic
|
| 21 |
+
|
| 22 |
+
|
| 23 |
+
def parse_args():
|
| 24 |
+
p = argparse.ArgumentParser(description="wikiart (streaming) -> mosaïques LEGO + GT.")
|
| 25 |
+
p.add_argument("--n", type=int, default=200, help="nb de mosaïques à générer")
|
| 26 |
+
p.add_argument("--skip", type=int, default=0, help="sauter les K premières images du flux")
|
| 27 |
+
p.add_argument("--out-dir", default="dataset_inputs", help="dossier de sortie")
|
| 28 |
+
p.add_argument("--grid", type=int, default=96, help="studs par côté")
|
| 29 |
+
p.add_argument("--stud-size", type=int, default=40, help="px par stud (≥40)")
|
| 30 |
+
p.add_argument("--joint-px", type=int, default=None, help="largeur du joint (défaut ~3)")
|
| 31 |
+
p.add_argument("--mode", choices=["tile", "plate", "brick", "mono"], default="tile",
|
| 32 |
+
help="mono = tuiles 1×1 seules (pour le curriculum ultracompact)")
|
| 33 |
+
p.add_argument("--big-plates", action="store_true", help="réactive 4×8/4×10")
|
| 34 |
+
p.add_argument("--max-dominant-frac", type=float, default=0.55,
|
| 35 |
+
help="garde-fou : rejeter une image si une seule couleur LEGO couvre "
|
| 36 |
+
"> cette fraction des cellules (impressionnistes → aplats monochromes). "
|
| 37 |
+
"1.0 ou plus = désactivé")
|
| 38 |
+
p.add_argument("--dataset", default="huggan/wikiart",
|
| 39 |
+
help="dataset HF en streaming (image attendue dans ex['image'])")
|
| 40 |
+
return p.parse_args()
|
| 41 |
+
|
| 42 |
+
|
| 43 |
+
def main():
|
| 44 |
+
a = parse_args()
|
| 45 |
+
from datasets import load_dataset # import tardif : message clair si absent
|
| 46 |
+
out = Path(a.out_dir)
|
| 47 |
+
out.mkdir(parents=True, exist_ok=True)
|
| 48 |
+
print(f"streaming {a.dataset} (train) → {a.n} mosaïques "
|
| 49 |
+
f"(mode={a.mode}, grid={a.grid}, skip={a.skip})…")
|
| 50 |
+
ds = load_dataset(a.dataset, split="train", streaming=True)
|
| 51 |
+
|
| 52 |
+
made = 0
|
| 53 |
+
rejected = 0 # garde-fou couleur dominante
|
| 54 |
+
for ex in islice(ds, a.skip, None):
|
| 55 |
+
if made >= a.n:
|
| 56 |
+
break
|
| 57 |
+
try:
|
| 58 |
+
img = ex["image"].convert("RGB")
|
| 59 |
+
except Exception as e: # image illisible / champ manquant
|
| 60 |
+
print(f" skip (image illisible): {e}")
|
| 61 |
+
continue
|
| 62 |
+
if min(img.size) < a.grid: # trop petite pour la grille demandée
|
| 63 |
+
continue
|
| 64 |
+
uid = str(uuid.uuid4()) # nom unique, uuid4 canonique (wikiart n'a pas d'ID)
|
| 65 |
+
try:
|
| 66 |
+
mos, grid = forge_mosaic(
|
| 67 |
+
img, grid_w=a.grid, grid_h=a.grid, stud_size=a.stud_size,
|
| 68 |
+
joint_px=a.joint_px, mode=a.mode, big_plates=a.big_plates,
|
| 69 |
+
source_name=uid, max_dominant_frac=a.max_dominant_frac)
|
| 70 |
+
except Exception as e: # forge échoue → on saute l'image
|
| 71 |
+
print(f" skip (forge: {e})")
|
| 72 |
+
continue
|
| 73 |
+
if mos is None: # garde-fou : aplat monochrome → rejeté
|
| 74 |
+
rejected += 1
|
| 75 |
+
continue
|
| 76 |
+
mos.save(out / f"canvas_mosaic_{uid}.png")
|
| 77 |
+
(out / f"piece_grid_{uid}.json").write_text(json.dumps(grid))
|
| 78 |
+
made += 1
|
| 79 |
+
if made % 20 == 0:
|
| 80 |
+
print(f" {made}/{a.n}… (rejetées par garde-fou : {rejected})")
|
| 81 |
+
print(f"Fait : {made} mosaïques dans {out}/ (rejetées par garde-fou couleur : {rejected})")
|
| 82 |
+
print(f"Ensuite : python3 forge_mosaics/mosaic2fragments/batch.py "
|
| 83 |
+
f"--inputs {out}/canvas_mosaic_*.png --out dataset_wikiart")
|
| 84 |
+
|
| 85 |
+
|
| 86 |
+
if __name__ == "__main__":
|
| 87 |
+
main()
|
code/mosaic2fragments/batch.py
ADDED
|
@@ -0,0 +1,90 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
"""Forge multiple mosaic samples by varying random seeds (and optionally inputs).
|
| 3 |
+
|
| 4 |
+
For 50 mosaics with a single input image, this produces 50 distinct
|
| 5 |
+
fragmentations of the same target. To grow real diversity, feed different
|
| 6 |
+
canvas_mosaic.png files via --inputs.
|
| 7 |
+
"""
|
| 8 |
+
|
| 9 |
+
import argparse
|
| 10 |
+
from pathlib import Path
|
| 11 |
+
|
| 12 |
+
from forge_dataset import forge_one
|
| 13 |
+
from visualize import visualize
|
| 14 |
+
|
| 15 |
+
|
| 16 |
+
def main():
|
| 17 |
+
p = argparse.ArgumentParser()
|
| 18 |
+
p.add_argument('--inputs', nargs='+', default=['exp1/canvas_mosaic.png'],
|
| 19 |
+
help='One or more canvas_mosaic.png paths')
|
| 20 |
+
p.add_argument('--out', default='dataset',
|
| 21 |
+
help='Root output directory (one subfolder per sample)')
|
| 22 |
+
p.add_argument('--n-samples', type=int, default=5,
|
| 23 |
+
help='Total number of mosaic samples to generate')
|
| 24 |
+
p.add_argument('--seed-start', type=int, default=0)
|
| 25 |
+
p.add_argument('--n-sides-min', type=int, default=16)
|
| 26 |
+
p.add_argument('--n-sides-max', type=int, default=24)
|
| 27 |
+
p.add_argument('--max-area-loss', type=float, default=0.01,
|
| 28 |
+
help="reco B — perte d'aire tolérée pour polygon_n (défaut 1%%) ; "
|
| 29 |
+
"c'est LE budget, n en découle")
|
| 30 |
+
p.add_argument('--n-frag-min', type=int, default=10)
|
| 31 |
+
p.add_argument('--n-frag-max', type=int, default=15)
|
| 32 |
+
p.add_argument('--canvas-w', type=int, default=3500)
|
| 33 |
+
p.add_argument('--canvas-h', type=int, default=3500)
|
| 34 |
+
p.add_argument('--debug', action='store_true',
|
| 35 |
+
help='écrit les fichiers DEBUG (pieces.json + source_yolo_viz.png) ; OFF par défaut')
|
| 36 |
+
p.add_argument('--name-from-input', action='store_true',
|
| 37 |
+
help="Nomme le dossier mosaic_<id> d'après le PNG d'entrée "
|
| 38 |
+
"(canvas_mosaic_<id>.png) au lieu de mosaic_000 séquentiel "
|
| 39 |
+
"→ noms uniques (uuid) préservés en sortie")
|
| 40 |
+
p.add_argument('--no-rotation', action='store_true',
|
| 41 |
+
help='Curriculum L1 : fragments éclatés SANS rotation (translation seule)')
|
| 42 |
+
p.add_argument('--placement', choices=['scatter', 'explode'], default='scatter',
|
| 43 |
+
help='explode = curriculum L0 : vue éclatée (positions relatives gardées)')
|
| 44 |
+
p.add_argument('--frag-distribution', choices=['balanced', 'compact'], default='balanced',
|
| 45 |
+
help='balanced (défaut) ou compact (~rectangulaire, moins de sommets)')
|
| 46 |
+
p.add_argument('--erode-px-min', type=int, default=0)
|
| 47 |
+
p.add_argument('--erode-px-max', type=int, default=0)
|
| 48 |
+
p.add_argument('--holes-min', type=int, default=0)
|
| 49 |
+
p.add_argument('--holes-max', type=int, default=0)
|
| 50 |
+
p.add_argument('--missing-min', type=int, default=0)
|
| 51 |
+
p.add_argument('--missing-max', type=int, default=0)
|
| 52 |
+
args = p.parse_args()
|
| 53 |
+
degrade = {
|
| 54 |
+
'erode_px': (args.erode_px_min, args.erode_px_max),
|
| 55 |
+
'holes': (args.holes_min, args.holes_max),
|
| 56 |
+
'missing': (args.missing_min, args.missing_max),
|
| 57 |
+
}
|
| 58 |
+
|
| 59 |
+
out_root = Path(args.out)
|
| 60 |
+
for i in range(args.n_samples):
|
| 61 |
+
inp = args.inputs[i % len(args.inputs)]
|
| 62 |
+
seed = args.seed_start + i
|
| 63 |
+
if args.name_from_input:
|
| 64 |
+
stem = Path(inp).stem
|
| 65 |
+
uid = stem[len('canvas_mosaic_'):] if stem.startswith('canvas_mosaic_') else stem
|
| 66 |
+
sample_dir = out_root / f'mosaic_{uid}'
|
| 67 |
+
else:
|
| 68 |
+
sample_dir = out_root / f'mosaic_{i:03d}'
|
| 69 |
+
print(f"\n=== sample {i:03d} (seed={seed}, input={inp}) ===")
|
| 70 |
+
forge_one(
|
| 71 |
+
target_path=inp,
|
| 72 |
+
out_dir=sample_dir,
|
| 73 |
+
n_sides_range=(args.n_sides_min, args.n_sides_max),
|
| 74 |
+
max_area_loss=args.max_area_loss,
|
| 75 |
+
n_frag_range=(args.n_frag_min, args.n_frag_max),
|
| 76 |
+
canvas_size=(args.canvas_w, args.canvas_h),
|
| 77 |
+
stud_size=None,
|
| 78 |
+
seed=seed,
|
| 79 |
+
degrade=degrade,
|
| 80 |
+
rotate=not args.no_rotation,
|
| 81 |
+
placement=args.placement,
|
| 82 |
+
frag_distribution=args.frag_distribution,
|
| 83 |
+
debug=args.debug,
|
| 84 |
+
)
|
| 85 |
+
if args.debug:
|
| 86 |
+
visualize(sample_dir)
|
| 87 |
+
|
| 88 |
+
|
| 89 |
+
if __name__ == '__main__':
|
| 90 |
+
main()
|
code/mosaic2fragments/curriculum.py
ADDED
|
@@ -0,0 +1,217 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
"""Curriculum de dégradation — génère les 5 paliers L0→L4 dans des sous-dossiers
|
| 3 |
+
`<out>/L0_explode` … `<out>/L4_strong`, depuis UN MÊME jeu de mosaïques de base.
|
| 4 |
+
|
| 5 |
+
Design **apparié** : même mosaïque + même seed à tous les paliers → la GT
|
| 6 |
+
(fragmentation + arêtes, fixée avant tout tirage de dégradation) est IDENTIQUE
|
| 7 |
+
d'un palier à l'autre ; seule la difficulté de l'INPUT change. → la courbe
|
| 8 |
+
score(difficulté) du GNN se lit à variance faible.
|
| 9 |
+
|
| 10 |
+
Paliers (ladder monotone) :
|
| 11 |
+
L0_explode vue éclatée (positions relatives conservées), sans rotation ← le + facile
|
| 12 |
+
L1_translation scatter aléatoire, sans rotation, aucune dégradation
|
| 13 |
+
L2_rotation scatter + rotation, aucune dégradation
|
| 14 |
+
L3_light scatter + rotation, dégradation LÉGÈRE (érosion 1-2 px, 0-1 trou)
|
| 15 |
+
L4_strong scatter + rotation, dégradation FORTE (érosion 2-6 px, trous, manquants)
|
| 16 |
+
|
| 17 |
+
⚠️ DAFNE B (distribution de découpe) n'est instrumenté que partiellement, via
|
| 18 |
+
`--frag-distribution` ; DAFNE D (fragments parasites) n'est pas implémenté.
|
| 19 |
+
|
| 20 |
+
Inputs = mosaïques rendues : soit des `canvas_mosaic_*.png`, soit les `target.png`
|
| 21 |
+
d'un dataset clean existant (target.png = copie de la mosaïque d'entrée).
|
| 22 |
+
|
| 23 |
+
Exemples :
|
| 24 |
+
# apparié 1:1 sur les mosaïques wikiart déjà rendues
|
| 25 |
+
python3 forge_mosaics/mosaic2fragments/curriculum.py \
|
| 26 |
+
--inputs-glob 'output/wikiart_inputs/canvas_mosaic_*.png' --out output --jobs 6
|
| 27 |
+
|
| 28 |
+
# augmenter : 5 fragmentations (seeds) par mosaïque → 500/palier à partir de 100 images
|
| 29 |
+
python3 forge_mosaics/mosaic2fragments/curriculum.py \
|
| 30 |
+
--inputs output/wikiart_inputs/canvas_mosaic_*.png --out output --n-per-input 5
|
| 31 |
+
|
| 32 |
+
Puis collate par palier :
|
| 33 |
+
for L in output/L?_*; do python3 forge_mosaics/features/collate.py --dataset "$L"; done
|
| 34 |
+
"""
|
| 35 |
+
import argparse
|
| 36 |
+
import glob
|
| 37 |
+
import zlib
|
| 38 |
+
from pathlib import Path
|
| 39 |
+
|
| 40 |
+
from forge_dataset import forge_one
|
| 41 |
+
from visualize import visualize
|
| 42 |
+
|
| 43 |
+
# (placement, rotate, érosion px, trous, fragments manquants) par palier.
|
| 44 |
+
# placement explode = vue éclatée (positions relatives gardées, juste espacées)
|
| 45 |
+
# placement scatter = placement aléatoire (peut échanger les positions)
|
| 46 |
+
LEVELS = {
|
| 47 |
+
"L0_explode": dict(placement="explode", rotate=False, erode=(0, 0), holes=(0, 0), missing=(0, 0)),
|
| 48 |
+
"L1_translation": dict(placement="scatter", rotate=False, erode=(0, 0), holes=(0, 0), missing=(0, 0)),
|
| 49 |
+
"L2_rotation": dict(placement="scatter", rotate=True, erode=(0, 0), holes=(0, 0), missing=(0, 0)),
|
| 50 |
+
"L3_light": dict(placement="scatter", rotate=True, erode=(1, 2), holes=(0, 1), missing=(0, 0)),
|
| 51 |
+
"L4_strong": dict(placement="scatter", rotate=True, erode=(2, 6), holes=(1, 3), missing=(0, 2)), # = degraded actuel
|
| 52 |
+
}
|
| 53 |
+
# L5 « knobs poussés » RETIRÉ (saturait : 0.092 vs 0.088 pour L4). Le vrai durcissement
|
| 54 |
+
# passe par l'axe ORTHOGONAL --frag-distribution {voronoi, clusters} (DAFNE B), pas par
|
| 55 |
+
# plus d'érosion. Structure de sortie = output/<frag-distribution>/<palier>/mosaic_*.
|
| 56 |
+
|
| 57 |
+
FRAG_DISTRIBUTIONS = ["balanced", "compact", "ultracompact", "voronoi", "clusters"]
|
| 58 |
+
|
| 59 |
+
|
| 60 |
+
def input_uid(path):
|
| 61 |
+
"""Identifiant stable d'une mosaïque de base (pour apparier les paliers)."""
|
| 62 |
+
p = Path(path)
|
| 63 |
+
if p.stem == "target": # .../mosaic_<uuid>/target.png
|
| 64 |
+
name = p.parent.name
|
| 65 |
+
return name[len("mosaic_"):] if name.startswith("mosaic_") else name
|
| 66 |
+
if p.stem.startswith("canvas_mosaic_"): # canvas_mosaic_<id>.png
|
| 67 |
+
return p.stem[len("canvas_mosaic_"):]
|
| 68 |
+
return p.stem
|
| 69 |
+
|
| 70 |
+
|
| 71 |
+
def parse_args():
|
| 72 |
+
p = argparse.ArgumentParser(description=__doc__,
|
| 73 |
+
formatter_class=argparse.RawDescriptionHelpFormatter)
|
| 74 |
+
p.add_argument("--inputs", nargs="+", default=[],
|
| 75 |
+
help="mosaïques de base (canvas_mosaic_*.png ou target.png d'un dataset clean)")
|
| 76 |
+
p.add_argument("--inputs-glob", default=None,
|
| 77 |
+
help="motif glob résolu en interne (robuste pour des milliers d'entrées, "
|
| 78 |
+
"évite ARG_MAX) ; ex. 'output/wikiart_inputs/canvas_mosaic_*.png'")
|
| 79 |
+
p.add_argument("--out", default="output", help="racine → output/<frag-distribution>/<palier>/")
|
| 80 |
+
p.add_argument("--frag-distribution", choices=FRAG_DISTRIBUTIONS, default="balanced",
|
| 81 |
+
help="motif de découpe (axe orthogonal aux paliers). balanced (défaut), "
|
| 82 |
+
"compact et ultracompact (= compact sur une mosaïque 1×1, cf. "
|
| 83 |
+
"forge_LAR_2mosaic --mode mono) sont implémentés ; voronoi/clusters "
|
| 84 |
+
"(DAFNE B, distribution de tailles continue) = à venir")
|
| 85 |
+
p.add_argument("--n-per-input", type=int, default=1,
|
| 86 |
+
help="nb de fragmentations (seeds) par mosaïque de base (augmentation)")
|
| 87 |
+
p.add_argument("--seed-start", type=int, default=0)
|
| 88 |
+
p.add_argument("--seed-mode", choices=("uid", "constant"), default="uid",
|
| 89 |
+
help="uid (défaut) : seed = crc32(uuid de la mosaïque) + seed-start "
|
| 90 |
+
"→ une découpe DIFFÉRENTE par mosaïque. constant : ancien "
|
| 91 |
+
"comportement (seed = seed-start pour toutes) — à n'utiliser "
|
| 92 |
+
"que pour reproduire un run d'avant le 27/07/2026")
|
| 93 |
+
p.add_argument("--n-sides-min", type=int, default=16)
|
| 94 |
+
p.add_argument("--n-sides-max", type=int, default=24)
|
| 95 |
+
p.add_argument("--n-frag-min", type=int, default=10)
|
| 96 |
+
p.add_argument("--n-frag-max", type=int, default=15)
|
| 97 |
+
p.add_argument("--max-area-loss", type=float, default=0.01,
|
| 98 |
+
help="reco B — fraction d'aire que polygon_n a le droit de PERDRE "
|
| 99 |
+
"vs polygon_raw (défaut 0.01 = 1 %%). C'est LE budget ; le "
|
| 100 |
+
"nombre de sommets en découle (correctif 23/07/2026).")
|
| 101 |
+
p.add_argument("--missing-max", type=int, default=None,
|
| 102 |
+
help="plafond du nb de fragments MANQUANTS (DAFNE C) des paliers. "
|
| 103 |
+
"DAFNE C est un POURCENTAGE, notre `missing` un COMPTE ABSOLU → "
|
| 104 |
+
"défaut = round(0.15·n_frag_max) (~15 %% de k, quel que soit k). "
|
| 105 |
+
"À k=10-15 ça donne 2 = le réglage historique exact (run 25k "
|
| 106 |
+
"reproductible) ; à k=2/3 → 0 ; k=5 → 1. Passer une valeur pour "
|
| 107 |
+
"outrepasser.")
|
| 108 |
+
p.add_argument("--canvas-w", type=int, default=4096)
|
| 109 |
+
p.add_argument("--canvas-h", type=int, default=4096)
|
| 110 |
+
p.add_argument("--levels", nargs="+", default=list(LEVELS),
|
| 111 |
+
help=f"sous-ensemble de paliers (défaut : tous → {list(LEVELS)})")
|
| 112 |
+
p.add_argument("--jobs", type=int, default=1,
|
| 113 |
+
help="processus parallèles (1 = série). Forge = CPU-bound, embarrassingly "
|
| 114 |
+
"parallel. Sur machine partagée : plafonner (ex. 12 sur 24), pas tous les cœurs")
|
| 115 |
+
p.add_argument("--debug", action="store_true",
|
| 116 |
+
help="écrit les fichiers DEBUG (pieces.json + source_yolo_viz.png), "
|
| 117 |
+
"lus nulle part par l'entraînement ; OFF par défaut")
|
| 118 |
+
return p.parse_args()
|
| 119 |
+
|
| 120 |
+
|
| 121 |
+
def _run_task(t):
|
| 122 |
+
"""Worker (1 instance). Top-level → picklable par multiprocessing (spawn macOS).
|
| 123 |
+
|
| 124 |
+
ROBUSTESSE (grosses générations) :
|
| 125 |
+
- **resume** : si l'instance est déjà complète (`degradation.md` = dernier fichier
|
| 126 |
+
écrit), on la saute → un relancement reprend où ça s'est arrêté.
|
| 127 |
+
- **tolérance** : toute exception d'une instance est CAPTURÉE et renvoyée comme
|
| 128 |
+
`FAIL …` au lieu de planter tout le batch (1 mosaïque pourrie ≠ 25 000 perdues).
|
| 129 |
+
"""
|
| 130 |
+
if (Path(t["dir"]) / "degradation.md").exists():
|
| 131 |
+
return f"skip {t['label']}"
|
| 132 |
+
try:
|
| 133 |
+
forge_one(
|
| 134 |
+
target_path=t["inp"], out_dir=t["dir"],
|
| 135 |
+
n_sides_range=t["n_sides"], n_frag_range=t["n_frag"], canvas_size=t["canvas"],
|
| 136 |
+
stud_size=None, seed=t["seed"], degrade=t["degrade"],
|
| 137 |
+
rotate=t["rotate"], placement=t["placement"], frag_distribution=t["frag_distribution"],
|
| 138 |
+
debug=t["debug"], max_area_loss=t["max_area_loss"],
|
| 139 |
+
)
|
| 140 |
+
if t["viz"]:
|
| 141 |
+
visualize(t["dir"])
|
| 142 |
+
return t["label"]
|
| 143 |
+
except Exception as e:
|
| 144 |
+
return f"FAIL {t['label']}: {type(e).__name__}: {e}"
|
| 145 |
+
|
| 146 |
+
|
| 147 |
+
def main():
|
| 148 |
+
a = parse_args()
|
| 149 |
+
if a.frag_distribution in ("voronoi", "clusters"):
|
| 150 |
+
raise SystemExit(f"--frag-distribution {a.frag_distribution} : pas encore "
|
| 151 |
+
"implémenté (DAFNE B à venir ; dispo : balanced, compact)")
|
| 152 |
+
inputs = list(a.inputs) + (sorted(glob.glob(a.inputs_glob)) if a.inputs_glob else [])
|
| 153 |
+
if not inputs:
|
| 154 |
+
raise SystemExit("aucune entrée : fournir --inputs ... ou --inputs-glob 'motif'")
|
| 155 |
+
out_root = Path(a.out) / a.frag_distribution
|
| 156 |
+
tasks = []
|
| 157 |
+
for inp in inputs:
|
| 158 |
+
uid = input_uid(inp)
|
| 159 |
+
for s in range(a.n_per_input):
|
| 160 |
+
# ⚠️ Le seed DOIT dépendre de la mosaïque. Avec un seed constant, en
|
| 161 |
+
# `mono` (tuiles 1×1) la grille d'adjacence est identique pour toute
|
| 162 |
+
# image → la fragmentation, déterministe, rendait la MÊME découpe et
|
| 163 |
+
# les MÊMES poses pour les 5000 mosaïques (bug trouvé le 27/07/2026 :
|
| 164 |
+
# les 4 datasets mono_compact publiés n'avaient qu'une seule
|
| 165 |
+
# géométrie). En `poly` le bug était masqué : la découpe suit les
|
| 166 |
+
# joints de tuiles variables, qui dépendent de l'image.
|
| 167 |
+
# crc32(uid) ne dépend QUE de la mosaïque → l'appariement reste
|
| 168 |
+
# exact entre paliers ET entre valeurs de k.
|
| 169 |
+
seed = (zlib.crc32(uid.encode()) + a.seed_start + s) % (2 ** 31) \
|
| 170 |
+
if a.seed_mode == "uid" else a.seed_start + s
|
| 171 |
+
# même seed sur tous les paliers → mosaïque/fragmentation appariée
|
| 172 |
+
suffix = f"_s{seed}" if a.n_per_input > 1 else ""
|
| 173 |
+
for level in a.levels:
|
| 174 |
+
cfg = LEVELS[level]
|
| 175 |
+
# DAFNE C est un %, `missing` un compte absolu → plafond ~15 % de k.
|
| 176 |
+
# round(0.15·15)=2 ⟹ identique au réglage historique à k=10-15.
|
| 177 |
+
cap = a.missing_max if a.missing_max is not None \
|
| 178 |
+
else round(0.15 * a.n_frag_max)
|
| 179 |
+
miss = (min(cfg["missing"][0], cap), min(cfg["missing"][1], cap))
|
| 180 |
+
tasks.append(dict(
|
| 181 |
+
inp=inp, dir=str(out_root / level / f"mosaic_{uid}{suffix}"),
|
| 182 |
+
n_sides=(a.n_sides_min, a.n_sides_max),
|
| 183 |
+
n_frag=(a.n_frag_min, a.n_frag_max),
|
| 184 |
+
canvas=(a.canvas_w, a.canvas_h), seed=seed,
|
| 185 |
+
degrade={"erode_px": cfg["erode"], "holes": cfg["holes"], "missing": miss},
|
| 186 |
+
rotate=cfg["rotate"], placement=cfg["placement"],
|
| 187 |
+
frag_distribution=a.frag_distribution, viz=a.debug, debug=a.debug,
|
| 188 |
+
max_area_loss=a.max_area_loss,
|
| 189 |
+
label=f"{level} mosaic_{uid}{suffix}"))
|
| 190 |
+
total = len(tasks)
|
| 191 |
+
print(f"curriculum [{a.frag_distribution}] : {total} instances → {out_root}/ (jobs={a.jobs})")
|
| 192 |
+
fails = skips = 0
|
| 193 |
+
|
| 194 |
+
def _tally(lab, i):
|
| 195 |
+
nonlocal fails, skips
|
| 196 |
+
if lab.startswith("FAIL"):
|
| 197 |
+
fails += 1
|
| 198 |
+
elif lab.startswith("skip"):
|
| 199 |
+
skips += 1
|
| 200 |
+
if lab.startswith("FAIL") or i % 100 == 0 or i == total:
|
| 201 |
+
print(f"[{i}/{total}] {lab}")
|
| 202 |
+
|
| 203 |
+
if a.jobs <= 1:
|
| 204 |
+
for i, t in enumerate(tasks, 1):
|
| 205 |
+
_tally(_run_task(t), i)
|
| 206 |
+
else:
|
| 207 |
+
import multiprocessing as mp
|
| 208 |
+
with mp.Pool(a.jobs) as pool:
|
| 209 |
+
for i, lab in enumerate(pool.imap_unordered(_run_task, tasks), 1):
|
| 210 |
+
_tally(lab, i)
|
| 211 |
+
print(f"\nFait : {total - fails - skips} générées, {skips} sautées (resume), {fails} échecs.")
|
| 212 |
+
print(f"Collate par palier :\n"
|
| 213 |
+
f" for L in {out_root}/L*_*; do python3 forge_mosaics/features/collate.py --dataset \"$L\"; done")
|
| 214 |
+
|
| 215 |
+
|
| 216 |
+
if __name__ == "__main__":
|
| 217 |
+
main()
|
code/mosaic2fragments/forge_dataset.py
ADDED
|
@@ -0,0 +1,1161 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
"""Forge a LEGO mosaic dataset for YOLO + GNN training.
|
| 3 |
+
|
| 4 |
+
Inputs: a canvas_mosaic.png produced by ../forge_LAR_2mosaic/ (square, >=40 px per
|
| 5 |
+
stud, grey joints, LEGO palette without greys).
|
| 6 |
+
Outputs (per mosaic, in <out_dir>/mosaic_XXX/):
|
| 7 |
+
target.png — copy of the input mosaic
|
| 8 |
+
source.png — fragments exploded on a white canvas (YOLO-Seg input)
|
| 9 |
+
source_yolo.txt — YOLO-Seg polygons, one line per fragment (THE YOLO ground truth)
|
| 10 |
+
pieces.json — detected LEGO pieces (debug only, read nowhere) — opt-in via --debug
|
| 11 |
+
graph_complete.json — N nodes (fragment features) + adjacency edges (target)
|
| 12 |
+
graph_fragments.json — same N nodes, zero edges (GNN input)
|
| 13 |
+
gt_layout.json — exact footprint and pose (x, y, rot) of each fragment
|
| 14 |
+
degradation.md — degradation report (empty means clean)
|
| 15 |
+
fragments/frag_XX.png — per-fragment alpha crop in target frame
|
| 16 |
+
|
| 17 |
+
Pipeline:
|
| 18 |
+
1. Joint detection on the rendered mosaic → LEGO piece layout
|
| 19 |
+
2. Union-find on cells → pieces (rectangles on the stud grid)
|
| 20 |
+
3. Piece adjacency graph → region-growing fragmentation (k ∈ [10,15] by default)
|
| 21 |
+
4. Per-fragment polygon → simplified under an AREA-LOSS budget (--max-area-loss,
|
| 22 |
+
default 1%), the vertex count being an output, not a target
|
| 23 |
+
5. Per-fragment features (global) + per-side features
|
| 24 |
+
6. Fragment adjacency = boundary segments shared between fragments
|
| 25 |
+
7. Random placement of rotated fragments on a white canvas (axes-aligned + ±5°)
|
| 26 |
+
8. Polygon export in YOLO-Seg format (normalized)
|
| 27 |
+
"""
|
| 28 |
+
|
| 29 |
+
import argparse
|
| 30 |
+
import collections
|
| 31 |
+
import heapq
|
| 32 |
+
import json
|
| 33 |
+
import os
|
| 34 |
+
import random
|
| 35 |
+
import sys
|
| 36 |
+
from pathlib import Path
|
| 37 |
+
|
| 38 |
+
import cv2
|
| 39 |
+
import networkx as nx
|
| 40 |
+
import numpy as np
|
| 41 |
+
from PIL import Image
|
| 42 |
+
|
| 43 |
+
# Extraction de features = module PARTAGÉ post-YOLO/pré-GNN (features/), importé
|
| 44 |
+
# depuis la racine du repo → forge synthétique et future chaîne YOLO utilisent le
|
| 45 |
+
# MÊME code (alignement de schéma garanti).
|
| 46 |
+
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
|
| 47 |
+
from features.fragment_features import ( # noqa: E402
|
| 48 |
+
_polygon_signed_area, extract_polygon, resample_reflex_aware,
|
| 49 |
+
pca_canonical_rotation, compute_fragment_features, MAX_AREA_LOSS_DEFAULT,
|
| 50 |
+
)
|
| 51 |
+
|
| 52 |
+
|
| 53 |
+
# ============================================================
|
| 54 |
+
# 1. Joint detection
|
| 55 |
+
# ============================================================
|
| 56 |
+
|
| 57 |
+
def is_joint_pixel(px, target=136, tol=35):
|
| 58 |
+
r, g, b = int(px[0]), int(px[1]), int(px[2])
|
| 59 |
+
if abs(r - target) > tol or abs(g - target) > tol or abs(b - target) > tol:
|
| 60 |
+
return False
|
| 61 |
+
if max(r, g, b) - min(r, g, b) > 25:
|
| 62 |
+
return False
|
| 63 |
+
return True
|
| 64 |
+
|
| 65 |
+
|
| 66 |
+
def _cell_is_gray(img, i, j, stud_size):
|
| 67 |
+
"""True if the interior of cell (i, j) is itself joint-gray (= a gray piece)."""
|
| 68 |
+
cy = min(j * stud_size + stud_size // 2, img.shape[0] - 1)
|
| 69 |
+
cx = min(i * stud_size + stud_size // 2, img.shape[1] - 1)
|
| 70 |
+
return is_joint_pixel(img[cy, cx])
|
| 71 |
+
|
| 72 |
+
|
| 73 |
+
def detect_joints(img, stud_size, tolerate_gray=True):
|
| 74 |
+
"""Return joint_v[i, j] (vertical joints) and joint_h[i, j] (horizontal joints).
|
| 75 |
+
|
| 76 |
+
joint_v[i, j] = True iff there is a joint between cell (i, j) and (i+1, j).
|
| 77 |
+
joint_h[i, j] = True iff there is a joint between cell (i, j) and (i, j+1).
|
| 78 |
+
Convention: i = column index (x-axis), j = row index (y-axis).
|
| 79 |
+
|
| 80 |
+
With tolerate_gray=True, a gray boundary is NOT a joint when BOTH adjacent
|
| 81 |
+
cell interiors are themselves gray — we are inside a gray LEGO piece, not on
|
| 82 |
+
a real joint. This lets gray pieces stay in the palette. The only
|
| 83 |
+
unresolvable case is two *different* gray pieces sharing a gray joint (rare,
|
| 84 |
+
accepted). For mosaics generated without gray pieces this is a strict no-op.
|
| 85 |
+
"""
|
| 86 |
+
H, W = img.shape[:2]
|
| 87 |
+
n_cols = W // stud_size
|
| 88 |
+
n_rows = H // stud_size
|
| 89 |
+
|
| 90 |
+
joint_v = np.zeros((n_cols - 1, n_rows), dtype=bool)
|
| 91 |
+
joint_h = np.zeros((n_cols, n_rows - 1), dtype=bool)
|
| 92 |
+
|
| 93 |
+
for i in range(n_cols - 1):
|
| 94 |
+
x = (i + 1) * stud_size
|
| 95 |
+
for j in range(n_rows):
|
| 96 |
+
y_mid = j * stud_size + stud_size // 2
|
| 97 |
+
samples = [
|
| 98 |
+
is_joint_pixel(img[j * stud_size + stud_size // 4, x]),
|
| 99 |
+
is_joint_pixel(img[y_mid, x]),
|
| 100 |
+
is_joint_pixel(img[j * stud_size + 3 * stud_size // 4, x]),
|
| 101 |
+
]
|
| 102 |
+
is_joint = sum(samples) >= 2
|
| 103 |
+
if (is_joint and tolerate_gray
|
| 104 |
+
and _cell_is_gray(img, i, j, stud_size)
|
| 105 |
+
and _cell_is_gray(img, i + 1, j, stud_size)):
|
| 106 |
+
is_joint = False # interior of a gray piece, not a joint
|
| 107 |
+
joint_v[i, j] = is_joint
|
| 108 |
+
|
| 109 |
+
for j in range(n_rows - 1):
|
| 110 |
+
y = (j + 1) * stud_size
|
| 111 |
+
for i in range(n_cols):
|
| 112 |
+
x_mid = i * stud_size + stud_size // 2
|
| 113 |
+
samples = [
|
| 114 |
+
is_joint_pixel(img[y, i * stud_size + stud_size // 4]),
|
| 115 |
+
is_joint_pixel(img[y, x_mid]),
|
| 116 |
+
is_joint_pixel(img[y, i * stud_size + 3 * stud_size // 4]),
|
| 117 |
+
]
|
| 118 |
+
is_joint = sum(samples) >= 2
|
| 119 |
+
if (is_joint and tolerate_gray
|
| 120 |
+
and _cell_is_gray(img, i, j, stud_size)
|
| 121 |
+
and _cell_is_gray(img, i, j + 1, stud_size)):
|
| 122 |
+
is_joint = False
|
| 123 |
+
joint_h[i, j] = is_joint
|
| 124 |
+
|
| 125 |
+
return joint_v, joint_h
|
| 126 |
+
|
| 127 |
+
|
| 128 |
+
# ============================================================
|
| 129 |
+
# 2. Piece detection via union-find on cells
|
| 130 |
+
# ============================================================
|
| 131 |
+
|
| 132 |
+
def detect_pieces(img, stud_size):
|
| 133 |
+
joint_v, joint_h = detect_joints(img, stud_size)
|
| 134 |
+
n_cols = joint_v.shape[0] + 1
|
| 135 |
+
n_rows = joint_h.shape[1] + 1
|
| 136 |
+
|
| 137 |
+
parent = {(i, j): (i, j) for i in range(n_cols) for j in range(n_rows)}
|
| 138 |
+
|
| 139 |
+
def find(c):
|
| 140 |
+
while parent[c] != c:
|
| 141 |
+
parent[c] = parent[parent[c]]
|
| 142 |
+
c = parent[c]
|
| 143 |
+
return c
|
| 144 |
+
|
| 145 |
+
def union(a, b):
|
| 146 |
+
ra, rb = find(a), find(b)
|
| 147 |
+
if ra != rb:
|
| 148 |
+
parent[ra] = rb
|
| 149 |
+
|
| 150 |
+
for i in range(n_cols - 1):
|
| 151 |
+
for j in range(n_rows):
|
| 152 |
+
if not joint_v[i, j]:
|
| 153 |
+
union((i, j), (i + 1, j))
|
| 154 |
+
for i in range(n_cols):
|
| 155 |
+
for j in range(n_rows - 1):
|
| 156 |
+
if not joint_h[i, j]:
|
| 157 |
+
union((i, j), (i, j + 1))
|
| 158 |
+
|
| 159 |
+
groups = collections.defaultdict(list)
|
| 160 |
+
for c in parent:
|
| 161 |
+
groups[find(c)].append(c)
|
| 162 |
+
|
| 163 |
+
pieces = []
|
| 164 |
+
for cells in groups.values():
|
| 165 |
+
cells = sorted(cells)
|
| 166 |
+
i_min = min(c[0] for c in cells)
|
| 167 |
+
i_max = max(c[0] for c in cells)
|
| 168 |
+
j_min = min(c[1] for c in cells)
|
| 169 |
+
j_max = max(c[1] for c in cells)
|
| 170 |
+
# Sample interior color from the center of an interior cell
|
| 171 |
+
ci, cj = cells[len(cells) // 2]
|
| 172 |
+
cx = ci * stud_size + stud_size // 2
|
| 173 |
+
cy = cj * stud_size + stud_size // 2
|
| 174 |
+
color = img[cy, cx][:3].tolist()
|
| 175 |
+
pieces.append({
|
| 176 |
+
'cells': cells,
|
| 177 |
+
'bbox_grid': [i_min, j_min, i_max - i_min + 1, j_max - j_min + 1],
|
| 178 |
+
'color': [int(c) for c in color],
|
| 179 |
+
})
|
| 180 |
+
return pieces
|
| 181 |
+
|
| 182 |
+
|
| 183 |
+
# ============================================================
|
| 184 |
+
# 2b. Sanity checks on detected pieces
|
| 185 |
+
# ============================================================
|
| 186 |
+
|
| 187 |
+
def sanity_check_pieces(pieces, n_cols, n_rows):
|
| 188 |
+
"""Return a list of warning strings if detected pieces look suspicious.
|
| 189 |
+
|
| 190 |
+
Checks:
|
| 191 |
+
- all grid cells are covered exactly once
|
| 192 |
+
- every piece is rectangular (its cells fill its bbox)
|
| 193 |
+
- share of 1x1 pieces is not abnormally high (a tell-tale sign of
|
| 194 |
+
gray-tile false-positive joints, since gray pieces get internally
|
| 195 |
+
split into 1x1s if joint detection is over-aggressive)
|
| 196 |
+
"""
|
| 197 |
+
warnings = []
|
| 198 |
+
total = sum(len(p['cells']) for p in pieces)
|
| 199 |
+
expected = n_cols * n_rows
|
| 200 |
+
if total != expected:
|
| 201 |
+
warnings.append(f"cell coverage: {total} from pieces, expected {expected}")
|
| 202 |
+
|
| 203 |
+
seen = set()
|
| 204 |
+
dup = 0
|
| 205 |
+
for p in pieces:
|
| 206 |
+
for c in p['cells']:
|
| 207 |
+
if c in seen:
|
| 208 |
+
dup += 1
|
| 209 |
+
seen.add(c)
|
| 210 |
+
if dup:
|
| 211 |
+
warnings.append(f"{dup} cells assigned to multiple pieces")
|
| 212 |
+
|
| 213 |
+
non_rect = 0
|
| 214 |
+
for p in pieces:
|
| 215 |
+
cells = p['cells']
|
| 216 |
+
i_min = min(c[0] for c in cells)
|
| 217 |
+
i_max = max(c[0] for c in cells)
|
| 218 |
+
j_min = min(c[1] for c in cells)
|
| 219 |
+
j_max = max(c[1] for c in cells)
|
| 220 |
+
if len(cells) != (i_max - i_min + 1) * (j_max - j_min + 1):
|
| 221 |
+
non_rect += 1
|
| 222 |
+
if non_rect:
|
| 223 |
+
warnings.append(f"{non_rect}/{len(pieces)} pieces are non-rectangular "
|
| 224 |
+
"(joint detection inconsistency)")
|
| 225 |
+
|
| 226 |
+
n_1x1 = sum(1 for p in pieces if len(p['cells']) == 1)
|
| 227 |
+
pct_1x1 = 100.0 * n_1x1 / max(len(pieces), 1)
|
| 228 |
+
if pct_1x1 > 60.0:
|
| 229 |
+
warnings.append(
|
| 230 |
+
f"{pct_1x1:.0f}% of pieces are 1x1 ({n_1x1}/{len(pieces)}). "
|
| 231 |
+
"Anything above ~60% suggests joint over-detection."
|
| 232 |
+
)
|
| 233 |
+
|
| 234 |
+
# Targeted gray-tile check: a large cluster of 1x1 with a similar near-gray
|
| 235 |
+
# color is the smoking gun for Light/Dark Bluish Gray false positives.
|
| 236 |
+
gray_1x1 = []
|
| 237 |
+
for p in pieces:
|
| 238 |
+
if len(p['cells']) != 1:
|
| 239 |
+
continue
|
| 240 |
+
r, g, b = p['color']
|
| 241 |
+
if (max(r, g, b) - min(r, g, b) < 25
|
| 242 |
+
and 80 <= (r + g + b) / 3 <= 200):
|
| 243 |
+
gray_1x1.append(p)
|
| 244 |
+
if len(gray_1x1) > 30:
|
| 245 |
+
warnings.append(
|
| 246 |
+
f"{len(gray_1x1)} gray-ish 1x1 pieces detected — likely "
|
| 247 |
+
"Light Bluish Gray / Dark Bluish Gray / Flat Silver / "
|
| 248 |
+
"Pearl Light Gray / Metallic Silver tiles being mistaken "
|
| 249 |
+
"for joints. Disable those colors in the palette."
|
| 250 |
+
)
|
| 251 |
+
return warnings
|
| 252 |
+
|
| 253 |
+
|
| 254 |
+
# ============================================================
|
| 255 |
+
# 3. Piece adjacency graph
|
| 256 |
+
# ============================================================
|
| 257 |
+
|
| 258 |
+
def build_piece_graph(pieces):
|
| 259 |
+
cell_to_pid = {}
|
| 260 |
+
for pid, p in enumerate(pieces):
|
| 261 |
+
for cell in p['cells']:
|
| 262 |
+
cell_to_pid[cell] = pid
|
| 263 |
+
|
| 264 |
+
g = nx.Graph()
|
| 265 |
+
for pid in range(len(pieces)):
|
| 266 |
+
g.add_node(pid)
|
| 267 |
+
for cell, pid in cell_to_pid.items():
|
| 268 |
+
i, j = cell
|
| 269 |
+
for di, dj in [(1, 0), (0, 1)]:
|
| 270 |
+
neighbor = (i + di, j + dj)
|
| 271 |
+
if neighbor in cell_to_pid:
|
| 272 |
+
pid2 = cell_to_pid[neighbor]
|
| 273 |
+
if pid != pid2:
|
| 274 |
+
g.add_edge(pid, pid2)
|
| 275 |
+
return g
|
| 276 |
+
|
| 277 |
+
|
| 278 |
+
# ============================================================
|
| 279 |
+
# 4. Fragmentation by balanced region growing
|
| 280 |
+
# ============================================================
|
| 281 |
+
|
| 282 |
+
def _bfs_distances(piece_graph, source, dist=None):
|
| 283 |
+
"""Unweighted shortest-path distances from `source` over `piece_graph`.
|
| 284 |
+
If `dist` is provided, distances are updated only where shorter."""
|
| 285 |
+
if dist is None:
|
| 286 |
+
dist = {n: float('inf') for n in piece_graph.nodes}
|
| 287 |
+
dist[source] = 0
|
| 288 |
+
queue = collections.deque([source])
|
| 289 |
+
while queue:
|
| 290 |
+
u = queue.popleft()
|
| 291 |
+
for v in piece_graph.neighbors(u):
|
| 292 |
+
if dist[u] + 1 < dist[v]:
|
| 293 |
+
dist[v] = dist[u] + 1
|
| 294 |
+
queue.append(v)
|
| 295 |
+
return dist
|
| 296 |
+
|
| 297 |
+
|
| 298 |
+
def select_seeds_farthest(piece_graph, k):
|
| 299 |
+
"""Farthest-point sampling: each new seed maximises distance to the nearest existing seed."""
|
| 300 |
+
nodes = list(piece_graph.nodes)
|
| 301 |
+
first = random.choice(nodes)
|
| 302 |
+
seeds = [first]
|
| 303 |
+
dist = _bfs_distances(piece_graph, first)
|
| 304 |
+
while len(seeds) < k:
|
| 305 |
+
# Argmax over nodes; ties broken randomly to avoid bias.
|
| 306 |
+
max_d = max(dist.values())
|
| 307 |
+
candidates = [n for n, d in dist.items() if d == max_d]
|
| 308 |
+
nxt = random.choice(candidates)
|
| 309 |
+
seeds.append(nxt)
|
| 310 |
+
# Update distance-to-nearest-seed by running BFS from new seed.
|
| 311 |
+
_bfs_distances(piece_graph, nxt, dist)
|
| 312 |
+
return seeds
|
| 313 |
+
|
| 314 |
+
|
| 315 |
+
def fragment_pieces(piece_graph, k):
|
| 316 |
+
"""Balanced multi-source region growing.
|
| 317 |
+
|
| 318 |
+
Seeds are picked via farthest-point sampling (k-center). At each step the
|
| 319 |
+
currently smallest fragment expands by one queued node — ensuring balanced
|
| 320 |
+
fragment sizes regardless of local connectivity differences.
|
| 321 |
+
"""
|
| 322 |
+
nodes = list(piece_graph.nodes)
|
| 323 |
+
if k >= len(nodes):
|
| 324 |
+
return {n: i for i, n in enumerate(nodes)}
|
| 325 |
+
|
| 326 |
+
seeds = select_seeds_farthest(piece_graph, k)
|
| 327 |
+
assigned = {s: i for i, s in enumerate(seeds)}
|
| 328 |
+
frontier = [collections.deque([s]) for s in seeds]
|
| 329 |
+
sizes = [1] * k
|
| 330 |
+
|
| 331 |
+
# Min-heap keyed on (current size, fragment_id) → smallest grows first.
|
| 332 |
+
# Tie-breaker is fragment_id which is stable.
|
| 333 |
+
heap = [(1, i) for i in range(k)]
|
| 334 |
+
heapq.heapify(heap)
|
| 335 |
+
|
| 336 |
+
while heap:
|
| 337 |
+
_, fid = heapq.heappop(heap)
|
| 338 |
+
if not frontier[fid]:
|
| 339 |
+
continue
|
| 340 |
+
current = frontier[fid].popleft()
|
| 341 |
+
for neighbor in piece_graph.neighbors(current):
|
| 342 |
+
if neighbor not in assigned:
|
| 343 |
+
assigned[neighbor] = fid
|
| 344 |
+
frontier[fid].append(neighbor)
|
| 345 |
+
sizes[fid] += 1
|
| 346 |
+
if frontier[fid]:
|
| 347 |
+
heapq.heappush(heap, (sizes[fid], fid))
|
| 348 |
+
return assigned
|
| 349 |
+
|
| 350 |
+
|
| 351 |
+
def fragment_pieces_compact(piece_graph, pieces, k):
|
| 352 |
+
"""Region growing biaisé COMPACITÉ — `--frag-distribution compact`.
|
| 353 |
+
|
| 354 |
+
Comme fragment_pieces, mais à chaque pas le fragment ajoute la pièce-frontière
|
| 355 |
+
qui **maximise le remplissage de sa bbox** (cellules remplies / aire bbox) →
|
| 356 |
+
fragments ~rectangulaires, silhouettes régulières, MOINS de coins concaves
|
| 357 |
+
→ `n_max` plus petit. On ne coupe JAMAIS une pièce (les bords suivent les
|
| 358 |
+
joints) : donc pas de rectangles parfaits à 4 sommets, le signal de bord pour
|
| 359 |
+
l'appariement GNN survit. **N'altère NI reco-B (⊆ raw) NI le pad+masque** :
|
| 360 |
+
seule la partition change ; le reste du pipeline est identique à balanced.
|
| 361 |
+
"""
|
| 362 |
+
nodes = list(piece_graph.nodes)
|
| 363 |
+
if k >= len(nodes):
|
| 364 |
+
return {n: i for i, n in enumerate(nodes)}
|
| 365 |
+
|
| 366 |
+
# Géométrie par pièce : cellules (set de (i,j)) + périmètre propre (rectangle).
|
| 367 |
+
pcells = [set(map(tuple, pieces[p]['cells'])) for p in range(len(pieces))]
|
| 368 |
+
pperim = [2 * (pieces[p]['bbox_grid'][2] + pieces[p]['bbox_grid'][3])
|
| 369 |
+
for p in range(len(pieces))]
|
| 370 |
+
|
| 371 |
+
seeds = select_seeds_farthest(piece_graph, k)
|
| 372 |
+
assigned = {s: i for i, s in enumerate(seeds)}
|
| 373 |
+
frag_cells = [set(pcells[s]) for s in seeds] # cellules par fragment
|
| 374 |
+
sizes = [1] * k
|
| 375 |
+
frontier = [set(n for n in piece_graph.neighbors(s) if n not in assigned)
|
| 376 |
+
for s in seeds]
|
| 377 |
+
|
| 378 |
+
heap = [(1, i) for i in range(k)]
|
| 379 |
+
heapq.heapify(heap)
|
| 380 |
+
while heap:
|
| 381 |
+
_, fid = heapq.heappop(heap)
|
| 382 |
+
cand = [p for p in frontier[fid] if p not in assigned]
|
| 383 |
+
frontier[fid] = set(cand)
|
| 384 |
+
if not cand:
|
| 385 |
+
continue
|
| 386 |
+
fc = frag_cells[fid]
|
| 387 |
+
|
| 388 |
+
def delta_perim(p):
|
| 389 |
+
# Δpérimètre en ajoutant p = périmètre propre de p − 2×(arêtes partagées
|
| 390 |
+
# avec le fragment). Remplir une concavité → Δ négatif → bord plus lisse
|
| 391 |
+
# → moins de coins concaves → reco-B sort moins de sommets.
|
| 392 |
+
shared = 0
|
| 393 |
+
for (i, j) in pcells[p]:
|
| 394 |
+
shared += ((i + 1, j) in fc) + ((i - 1, j) in fc) \
|
| 395 |
+
+ ((i, j + 1) in fc) + ((i, j - 1) in fc)
|
| 396 |
+
return (pperim[p] - 2 * shared, -shared, len(pcells[p]))
|
| 397 |
+
|
| 398 |
+
best = min(cand, key=delta_perim) # pièce qui lisse le plus le bord
|
| 399 |
+
assigned[best] = fid
|
| 400 |
+
frag_cells[fid] |= pcells[best]
|
| 401 |
+
sizes[fid] += 1
|
| 402 |
+
frontier[fid].discard(best)
|
| 403 |
+
for nb in piece_graph.neighbors(best):
|
| 404 |
+
if nb not in assigned:
|
| 405 |
+
frontier[fid].add(nb)
|
| 406 |
+
heapq.heappush(heap, (sizes[fid], fid))
|
| 407 |
+
|
| 408 |
+
# composantes non atteintes (rare : graphe normalement connexe) → rattachées
|
| 409 |
+
changed = True
|
| 410 |
+
while changed:
|
| 411 |
+
changed = False
|
| 412 |
+
for n in nodes:
|
| 413 |
+
if n in assigned:
|
| 414 |
+
continue
|
| 415 |
+
nbf = [assigned[x] for x in piece_graph.neighbors(n) if x in assigned]
|
| 416 |
+
if nbf:
|
| 417 |
+
assigned[n] = nbf[0]
|
| 418 |
+
changed = True
|
| 419 |
+
for i, n in enumerate(nodes): # tout isolé restant → fragment propre
|
| 420 |
+
assigned.setdefault(n, k + i)
|
| 421 |
+
return assigned
|
| 422 |
+
|
| 423 |
+
|
| 424 |
+
# ============================================================
|
| 425 |
+
# 5. Fragment polygon and resampling
|
| 426 |
+
# ============================================================
|
| 427 |
+
|
| 428 |
+
def fragment_mask(fragment_cells, img_shape, stud_size):
|
| 429 |
+
H, W = img_shape[:2]
|
| 430 |
+
mask = np.zeros((H, W), dtype=np.uint8)
|
| 431 |
+
for (i, j) in fragment_cells:
|
| 432 |
+
x0, y0 = i * stud_size, j * stud_size
|
| 433 |
+
mask[y0:y0 + stud_size, x0:x0 + stud_size] = 1
|
| 434 |
+
return mask
|
| 435 |
+
|
| 436 |
+
|
| 437 |
+
# extract_polygon, resample_reflex_aware (+ helpers PCA), compute_fragment_features
|
| 438 |
+
# → déplacés dans features/fragment_features.py (module partagé), importés en tête.
|
| 439 |
+
# (resample_polygon équidistant supprimé : remplacé par resample_reflex_aware.)
|
| 440 |
+
|
| 441 |
+
|
| 442 |
+
# ============================================================
|
| 443 |
+
# 7. Fragment adjacencies (segments shared between fragments)
|
| 444 |
+
# ============================================================
|
| 445 |
+
|
| 446 |
+
def compute_fragment_adjacencies(pieces, piece_to_frag, stud_size):
|
| 447 |
+
cell_to_frag = {}
|
| 448 |
+
for pid, p in enumerate(pieces):
|
| 449 |
+
fid = piece_to_frag[pid]
|
| 450 |
+
for cell in p['cells']:
|
| 451 |
+
cell_to_frag[cell] = fid
|
| 452 |
+
|
| 453 |
+
adj = collections.defaultdict(list)
|
| 454 |
+
for cell, fid in cell_to_frag.items():
|
| 455 |
+
i, j = cell
|
| 456 |
+
right = (i + 1, j)
|
| 457 |
+
if right in cell_to_frag and cell_to_frag[right] != fid:
|
| 458 |
+
other = cell_to_frag[right]
|
| 459 |
+
a, b = min(fid, other), max(fid, other)
|
| 460 |
+
x = (i + 1) * stud_size
|
| 461 |
+
adj[(a, b)].append((x, j * stud_size, x, (j + 1) * stud_size))
|
| 462 |
+
bottom = (i, j + 1)
|
| 463 |
+
if bottom in cell_to_frag and cell_to_frag[bottom] != fid:
|
| 464 |
+
other = cell_to_frag[bottom]
|
| 465 |
+
a, b = min(fid, other), max(fid, other)
|
| 466 |
+
y = (j + 1) * stud_size
|
| 467 |
+
adj[(a, b)].append((i * stud_size, y, (i + 1) * stud_size, y))
|
| 468 |
+
return adj, cell_to_frag
|
| 469 |
+
|
| 470 |
+
|
| 471 |
+
def find_outer_fragments_per_side(polygon_n, centroid, frag_id, cell_to_frag,
|
| 472 |
+
stud_size, offset=6, n_samples=7):
|
| 473 |
+
"""For each side of polygon_n, vote over multiple samples along the side
|
| 474 |
+
to find the most common adjacent fragment (or None if image edge)."""
|
| 475 |
+
out = []
|
| 476 |
+
n = len(polygon_n)
|
| 477 |
+
for i in range(n):
|
| 478 |
+
p0, p1 = polygon_n[i], polygon_n[(i + 1) % n]
|
| 479 |
+
votes = collections.Counter()
|
| 480 |
+
for t in np.linspace(0.15, 0.85, n_samples):
|
| 481 |
+
sample = (1 - t) * p0 + t * p1
|
| 482 |
+
outward = sample - centroid
|
| 483 |
+
norm = float(np.linalg.norm(outward))
|
| 484 |
+
if norm < 1e-6:
|
| 485 |
+
continue
|
| 486 |
+
outer = sample + outward / norm * offset
|
| 487 |
+
ci, cj = int(outer[0] // stud_size), int(outer[1] // stud_size)
|
| 488 |
+
if (ci, cj) in cell_to_frag:
|
| 489 |
+
o = cell_to_frag[(ci, cj)]
|
| 490 |
+
if o != frag_id:
|
| 491 |
+
votes[o] += 1
|
| 492 |
+
if votes:
|
| 493 |
+
best, _ = votes.most_common(1)[0]
|
| 494 |
+
out.append(int(best))
|
| 495 |
+
else:
|
| 496 |
+
out.append(None)
|
| 497 |
+
return out
|
| 498 |
+
|
| 499 |
+
|
| 500 |
+
# ============================================================
|
| 501 |
+
# 8. Exploded image: cut + rotate + place
|
| 502 |
+
# ============================================================
|
| 503 |
+
|
| 504 |
+
def _disk(radius):
|
| 505 |
+
d = 2 * int(radius) + 1
|
| 506 |
+
return cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (d, d))
|
| 507 |
+
|
| 508 |
+
|
| 509 |
+
def degrade_mask(mask, erode_px=0, n_holes=0, stud_size=40, max_hole_area_frac=0.10):
|
| 510 |
+
"""Degraded copy of a fragment alpha mask — only ever REMOVES material, so
|
| 511 |
+
degraded ⊆ perfect (keeps the "area never gained" invariant).
|
| 512 |
+
|
| 513 |
+
erode_px : morphological erosion radius in px (boundary wear).
|
| 514 |
+
n_holes : interior holes punched (missing tesselae *inside*).
|
| 515 |
+
max_hole_area_frac : hard cap on TOTAL hole area (default 10% of the
|
| 516 |
+
post-erosion fragment area). The n_holes disks share
|
| 517 |
+
this budget, so holes never remove more than the cap
|
| 518 |
+
(overlaps / boundary clipping only reduce it further).
|
| 519 |
+
Uses the module RNG (seeded in forge_one) for reproducibility.
|
| 520 |
+
"""
|
| 521 |
+
m = (mask > 0).astype(np.uint8)
|
| 522 |
+
if erode_px > 0:
|
| 523 |
+
m = cv2.erode(m, _disk(erode_px))
|
| 524 |
+
if n_holes > 0:
|
| 525 |
+
ys, xs = np.where(m > 0)
|
| 526 |
+
area = len(xs)
|
| 527 |
+
if area:
|
| 528 |
+
budget = max_hole_area_frac * area # ≤10% of fragment area
|
| 529 |
+
r = max(1, int(np.sqrt(budget / (n_holes * np.pi))))
|
| 530 |
+
for _ in range(n_holes):
|
| 531 |
+
k = random.randint(0, area - 1)
|
| 532 |
+
cv2.circle(m, (int(xs[k]), int(ys[k])), r, 0, -1)
|
| 533 |
+
return m
|
| 534 |
+
|
| 535 |
+
|
| 536 |
+
def make_fragment_rgba(fragment_cells, target_img, stud_size, mask=None):
|
| 537 |
+
if mask is None:
|
| 538 |
+
mask = fragment_mask(fragment_cells, target_img.shape, stud_size)
|
| 539 |
+
mask = (mask > 0).astype(np.uint8)
|
| 540 |
+
ys, xs = np.where(mask > 0)
|
| 541 |
+
y_min, y_max = int(ys.min()), int(ys.max())
|
| 542 |
+
x_min, x_max = int(xs.min()), int(xs.max())
|
| 543 |
+
crop = target_img[y_min:y_max + 1, x_min:x_max + 1]
|
| 544 |
+
crop_mask = mask[y_min:y_max + 1, x_min:x_max + 1]
|
| 545 |
+
rgba = np.dstack([crop, crop_mask * 255]).astype(np.uint8)
|
| 546 |
+
return rgba, (x_min, y_min, x_max, y_max)
|
| 547 |
+
|
| 548 |
+
|
| 549 |
+
def rotate_rgba(rgba, angle):
|
| 550 |
+
pil = Image.fromarray(rgba, 'RGBA')
|
| 551 |
+
rot = pil.rotate(angle, resample=Image.BILINEAR, expand=True)
|
| 552 |
+
return np.array(rot)
|
| 553 |
+
|
| 554 |
+
|
| 555 |
+
def _collides(x, y, alpha, occupied):
|
| 556 |
+
"""True if fragment alpha placed at (x, y) overlaps any occupied mask."""
|
| 557 |
+
h, w = alpha.shape
|
| 558 |
+
for (ox, oy, omask) in occupied:
|
| 559 |
+
oh, ow = omask.shape
|
| 560 |
+
ix1, iy1 = max(x, ox), max(y, oy)
|
| 561 |
+
ix2, iy2 = min(x + w, ox + ow), min(y + h, oy + oh)
|
| 562 |
+
if ix1 >= ix2 or iy1 >= iy2:
|
| 563 |
+
continue
|
| 564 |
+
a = alpha[iy1 - y:iy2 - y, ix1 - x:ix2 - x]
|
| 565 |
+
b = omask[iy1 - oy:iy2 - oy, ix1 - ox:ix2 - ox]
|
| 566 |
+
if np.logical_and(a, b).any():
|
| 567 |
+
return True
|
| 568 |
+
return False
|
| 569 |
+
|
| 570 |
+
|
| 571 |
+
def _scan_free_position(alpha, occupied, cw, ch, padding, step):
|
| 572 |
+
"""Deterministic raster scan for the first non-overlapping top-left position.
|
| 573 |
+
Returns None only if no free spot exists anywhere on the current canvas."""
|
| 574 |
+
h, w = alpha.shape
|
| 575 |
+
y = padding
|
| 576 |
+
while y + h + padding <= ch:
|
| 577 |
+
x = padding
|
| 578 |
+
while x + w + padding <= cw:
|
| 579 |
+
if not _collides(x, y, alpha, occupied):
|
| 580 |
+
return x, y
|
| 581 |
+
x += step
|
| 582 |
+
y += step
|
| 583 |
+
return None
|
| 584 |
+
|
| 585 |
+
|
| 586 |
+
def explode_fragments(fragments_data, target_size, base_factor=1.8, padding=4,
|
| 587 |
+
max_factor=4.0, step=0.3):
|
| 588 |
+
"""Placement « vue éclatée » — curriculum L0.
|
| 589 |
+
|
| 590 |
+
Chaque fragment GARDE sa position RELATIVE dans la mosaïque, simplement
|
| 591 |
+
écarté de ses voisins (≈ schéma de montage éclaté). Contrairement au scatter
|
| 592 |
+
aléatoire (place_fragments), aucun échange de position : un fragment du coin
|
| 593 |
+
haut-gauche reste en haut-gauche. On dilate les centroïdes-cible par un
|
| 594 |
+
facteur autour du centre, en l'augmentant jusqu'à supprimer les
|
| 595 |
+
chevauchements. Renvoie (canvas, placements en ordre d'entrée, taille).
|
| 596 |
+
"""
|
| 597 |
+
W, H = target_size
|
| 598 |
+
fac = base_factor
|
| 599 |
+
while True:
|
| 600 |
+
cw, ch = int(round(W * fac)), int(round(H * fac))
|
| 601 |
+
cxt, cyt, cxc, cyc = W / 2.0, H / 2.0, cw / 2.0, ch / 2.0
|
| 602 |
+
placements, occupied, overlap = [], [], False
|
| 603 |
+
for f in fragments_data:
|
| 604 |
+
rgba = f['rgba_rotated']
|
| 605 |
+
h, w = rgba.shape[:2]
|
| 606 |
+
alpha = rgba[..., 3] > 0
|
| 607 |
+
x0, y0, x1, y1 = f['bbox_in_target']
|
| 608 |
+
x = int(round(cxc + fac * ((x0 + x1) / 2.0 - cxt) - w / 2.0))
|
| 609 |
+
y = int(round(cyc + fac * ((y0 + y1) / 2.0 - cyt) - h / 2.0))
|
| 610 |
+
x = max(padding, min(x, cw - w - padding))
|
| 611 |
+
y = max(padding, min(y, ch - h - padding))
|
| 612 |
+
if _collides(x, y, alpha, occupied):
|
| 613 |
+
overlap = True
|
| 614 |
+
placements.append({'x': x, 'y': y, 'w': w, 'h': h})
|
| 615 |
+
occupied.append((x, y, alpha))
|
| 616 |
+
if not overlap or fac >= max_factor:
|
| 617 |
+
break
|
| 618 |
+
fac += step
|
| 619 |
+
canvas = Image.new('RGB', (cw, ch), (255, 255, 255))
|
| 620 |
+
for f, place in zip(fragments_data, placements):
|
| 621 |
+
pil = Image.fromarray(f['rgba_rotated'], 'RGBA')
|
| 622 |
+
canvas.paste(pil, (place['x'], place['y']), pil)
|
| 623 |
+
return canvas, placements, (cw, ch)
|
| 624 |
+
|
| 625 |
+
|
| 626 |
+
def place_fragments(fragments_data, canvas_size, max_tries=200, padding=4):
|
| 627 |
+
"""Scatter fragments on a white canvas WITHOUT overlap.
|
| 628 |
+
|
| 629 |
+
Per fragment (largest first): (1) random search, to keep the scattered look;
|
| 630 |
+
(2) deterministic raster scan — guaranteed to find a free spot if one exists;
|
| 631 |
+
(3) last resort, grow the canvas downward. Steps 2-3 replace the old
|
| 632 |
+
"place at random, may overlap" fallback that caused the mosaic_004 overlaps.
|
| 633 |
+
Returns the (possibly grown) canvas size so the caller normalises YOLO
|
| 634 |
+
coordinates against the right dimensions.
|
| 635 |
+
"""
|
| 636 |
+
cw, ch = canvas_size
|
| 637 |
+
placements_by_id = {}
|
| 638 |
+
occupied = [] # list of (x, y, alpha_mask)
|
| 639 |
+
|
| 640 |
+
# Place larger fragments first to ease packing
|
| 641 |
+
order = sorted(range(len(fragments_data)),
|
| 642 |
+
key=lambda k: -fragments_data[k]['rgba_rotated'].shape[0] *
|
| 643 |
+
fragments_data[k]['rgba_rotated'].shape[1])
|
| 644 |
+
|
| 645 |
+
for k in order:
|
| 646 |
+
rgba = fragments_data[k]['rgba_rotated']
|
| 647 |
+
h, w = rgba.shape[:2]
|
| 648 |
+
alpha = rgba[..., 3] > 0
|
| 649 |
+
|
| 650 |
+
# Fragment larger than the canvas → grow so it (and the random range) fit.
|
| 651 |
+
while w + 2 * padding > cw:
|
| 652 |
+
cw += w
|
| 653 |
+
while h + 2 * padding > ch:
|
| 654 |
+
ch += h
|
| 655 |
+
|
| 656 |
+
pos = None
|
| 657 |
+
for _ in range(max_tries):
|
| 658 |
+
x = random.randint(padding, cw - w - padding)
|
| 659 |
+
y = random.randint(padding, ch - h - padding)
|
| 660 |
+
if not _collides(x, y, alpha, occupied):
|
| 661 |
+
pos = (x, y)
|
| 662 |
+
break
|
| 663 |
+
|
| 664 |
+
if pos is None:
|
| 665 |
+
# Random search exhausted: deterministic scan is guaranteed to find a
|
| 666 |
+
# gap if the canvas still has room (mosaic_004 fix).
|
| 667 |
+
step = max(16, min(h, w) // 2)
|
| 668 |
+
pos = _scan_free_position(alpha, occupied, cw, ch, padding, step)
|
| 669 |
+
|
| 670 |
+
if pos is None:
|
| 671 |
+
# Canvas genuinely full: add a strip at the bottom and drop it there.
|
| 672 |
+
pos = (padding, ch + padding)
|
| 673 |
+
ch += h + 2 * padding
|
| 674 |
+
|
| 675 |
+
x, y = pos
|
| 676 |
+
placements_by_id[k] = {'x': x, 'y': y, 'w': w, 'h': h}
|
| 677 |
+
occupied.append((x, y, alpha))
|
| 678 |
+
|
| 679 |
+
canvas = Image.new('RGB', (cw, ch), (255, 255, 255))
|
| 680 |
+
placements = []
|
| 681 |
+
# Paste in original index order (deterministic z-order).
|
| 682 |
+
for idx, fdata in enumerate(fragments_data):
|
| 683 |
+
place = placements_by_id[idx]
|
| 684 |
+
pil = Image.fromarray(fdata['rgba_rotated'], 'RGBA')
|
| 685 |
+
canvas.paste(pil, (place['x'], place['y']), pil)
|
| 686 |
+
placements.append(place)
|
| 687 |
+
return canvas, placements, (cw, ch)
|
| 688 |
+
|
| 689 |
+
|
| 690 |
+
# ============================================================
|
| 691 |
+
# 9. YOLO polygon export from rotated alpha
|
| 692 |
+
# ============================================================
|
| 693 |
+
|
| 694 |
+
def extract_yolo_polygon(rgba_rotated, place_x, place_y, canvas_size, simplify_epsilon=1.5):
|
| 695 |
+
alpha = (rgba_rotated[..., 3] > 0).astype(np.uint8)
|
| 696 |
+
contours, _ = cv2.findContours(alpha, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_NONE)
|
| 697 |
+
if not contours:
|
| 698 |
+
return None
|
| 699 |
+
cnt = max(contours, key=cv2.contourArea)
|
| 700 |
+
if simplify_epsilon > 0:
|
| 701 |
+
cnt = cv2.approxPolyDP(cnt, simplify_epsilon, True)
|
| 702 |
+
cnt = cnt.reshape(-1, 2).astype(np.float32)
|
| 703 |
+
cnt[:, 0] = (cnt[:, 0] + place_x) / canvas_size[0]
|
| 704 |
+
cnt[:, 1] = (cnt[:, 1] + place_y) / canvas_size[1]
|
| 705 |
+
return cnt
|
| 706 |
+
|
| 707 |
+
|
| 708 |
+
# ============================================================
|
| 709 |
+
# 10. Top-level orchestration
|
| 710 |
+
# ============================================================
|
| 711 |
+
|
| 712 |
+
def forge_one(target_path, out_dir, n_sides_range, n_frag_range, canvas_size,
|
| 713 |
+
stud_size, seed, degrade=None, rotate=True, placement='scatter',
|
| 714 |
+
frag_distribution='balanced', debug=False,
|
| 715 |
+
max_area_loss=MAX_AREA_LOSS_DEFAULT):
|
| 716 |
+
if seed is not None:
|
| 717 |
+
random.seed(seed)
|
| 718 |
+
np.random.seed(seed)
|
| 719 |
+
|
| 720 |
+
degrade = degrade or {}
|
| 721 |
+
erode_rng = tuple(degrade.get('erode_px', (0, 0)))
|
| 722 |
+
holes_rng = tuple(degrade.get('holes', (0, 0)))
|
| 723 |
+
missing_rng = tuple(degrade.get('missing', (0, 0))) # absolute count [min, max]
|
| 724 |
+
|
| 725 |
+
out_dir = Path(out_dir)
|
| 726 |
+
out_dir.mkdir(parents=True, exist_ok=True)
|
| 727 |
+
(out_dir / 'fragments').mkdir(exist_ok=True) # toujours écrit (entrée du VLM)
|
| 728 |
+
|
| 729 |
+
pil = Image.open(target_path).convert('RGB')
|
| 730 |
+
img = np.array(pil)
|
| 731 |
+
H, W = img.shape[:2]
|
| 732 |
+
if stud_size is None:
|
| 733 |
+
# Grille la PLUS FINE valide : plus grand g divisant l'image avec un
|
| 734 |
+
# stud résultant ≥ 40 px (contrainte projet : ≥40 px/stud). Désambiguïse
|
| 735 |
+
# les images divisibles par plusieurs tailles (ex. 3840 = 48·80 = 96·40
|
| 736 |
+
# → grille 96, stud 40) au lieu de retomber sur la grille la plus
|
| 737 |
+
# grossière. Les mosaïques 1920 restent en grille 48 (stud 40).
|
| 738 |
+
candidates = [g for g in (48, 50, 64, 96, 100, 128)
|
| 739 |
+
if W % g == 0 and H % g == 0 and W // g >= 40]
|
| 740 |
+
stud_size = W // max(candidates) if candidates else 40
|
| 741 |
+
print(f"[forge] image {W}x{H}, stud_size={stud_size} → grid {W // stud_size}x{H // stud_size}")
|
| 742 |
+
|
| 743 |
+
print("[forge] detecting LEGO pieces…")
|
| 744 |
+
pieces = detect_pieces(img, stud_size)
|
| 745 |
+
print(f"[forge] {len(pieces)} pieces detected")
|
| 746 |
+
|
| 747 |
+
n_cols, n_rows = W // stud_size, H // stud_size
|
| 748 |
+
warnings = sanity_check_pieces(pieces, n_cols, n_rows)
|
| 749 |
+
for w in warnings:
|
| 750 |
+
print(f"[warn] {w}")
|
| 751 |
+
|
| 752 |
+
piece_graph = build_piece_graph(pieces)
|
| 753 |
+
|
| 754 |
+
k = random.randint(n_frag_range[0], n_frag_range[1])
|
| 755 |
+
print(f"[forge] fragmenting into {k} fragments ({frag_distribution})…")
|
| 756 |
+
# ultracompact = compact appliqué à des mosaïques 1×1 (mode mono) → la
|
| 757 |
+
# fragmentation est la MÊME que compact ; "l'ultra" vient de l'input 1×1.
|
| 758 |
+
if frag_distribution in ('compact', 'ultracompact'):
|
| 759 |
+
piece_to_frag = fragment_pieces_compact(piece_graph, pieces, k)
|
| 760 |
+
else:
|
| 761 |
+
piece_to_frag = fragment_pieces(piece_graph, k)
|
| 762 |
+
frag_pieces = collections.defaultdict(list)
|
| 763 |
+
for pid, fid in piece_to_frag.items():
|
| 764 |
+
frag_pieces[fid].append(pid)
|
| 765 |
+
actual_k = len(frag_pieces)
|
| 766 |
+
print(f"[forge] {actual_k} non-empty fragments")
|
| 767 |
+
|
| 768 |
+
adjacencies, cell_to_frag = compute_fragment_adjacencies(pieces, piece_to_frag, stud_size)
|
| 769 |
+
print(f"[forge] {len(adjacencies)} fragment-fragment adjacencies")
|
| 770 |
+
|
| 771 |
+
print("[forge] extracting polygons & features per fragment…")
|
| 772 |
+
fragment_data = []
|
| 773 |
+
sorted_fids = sorted(frag_pieces.keys())
|
| 774 |
+
fid_to_idx = {fid: idx for idx, fid in enumerate(sorted_fids)}
|
| 775 |
+
|
| 776 |
+
for fid in sorted_fids:
|
| 777 |
+
all_cells = []
|
| 778 |
+
weighted_color = np.zeros(3, dtype=np.float64)
|
| 779 |
+
total_cells = 0
|
| 780 |
+
for pid in frag_pieces[fid]:
|
| 781 |
+
all_cells.extend(pieces[pid]['cells'])
|
| 782 |
+
n_cells = len(pieces[pid]['cells'])
|
| 783 |
+
weighted_color += np.array(pieces[pid]['color'], dtype=np.float64) * n_cells
|
| 784 |
+
total_cells += n_cells
|
| 785 |
+
mean_color = (weighted_color / max(total_cells, 1)).tolist()
|
| 786 |
+
|
| 787 |
+
mask = fragment_mask(all_cells, img.shape, stud_size)
|
| 788 |
+
polygon_raw = extract_polygon(mask) # PERFECT footprint → GT
|
| 789 |
+
# --- degradation: input only (the GT polygon_raw stays perfect) ---
|
| 790 |
+
erode_px = random.randint(*erode_rng) if erode_rng[1] > 0 else 0
|
| 791 |
+
n_holes = random.randint(*holes_rng) if holes_rng[1] > 0 else 0
|
| 792 |
+
if erode_px or n_holes:
|
| 793 |
+
mask_in = degrade_mask(mask, erode_px, n_holes, stud_size)
|
| 794 |
+
polygon_raw_deg = extract_polygon(mask_in)
|
| 795 |
+
if len(polygon_raw_deg) >= 4:
|
| 796 |
+
# erosion/holes add rasterisation noise (1px stair-steps) that
|
| 797 |
+
# would explode #reflex → simplify so reco B keeps n reasonable.
|
| 798 |
+
eps = max(1.5, stud_size * 0.06)
|
| 799 |
+
cnt = polygon_raw_deg.reshape(-1, 1, 2).astype(np.float32)
|
| 800 |
+
polygon_raw_deg = cv2.approxPolyDP(cnt, eps, True).reshape(-1, 2).astype(np.float32)
|
| 801 |
+
if len(polygon_raw_deg) < 4: # degradation wiped it → revert
|
| 802 |
+
mask_in = (mask > 0).astype(np.uint8)
|
| 803 |
+
polygon_raw_deg = polygon_raw
|
| 804 |
+
else:
|
| 805 |
+
mask_in = (mask > 0).astype(np.uint8)
|
| 806 |
+
polygon_raw_deg = polygon_raw
|
| 807 |
+
|
| 808 |
+
# Degradation severity = area lost by the degraded footprint vs the
|
| 809 |
+
# perfect one (DERIVED metric, not an input knob — cf. reco/06/06).
|
| 810 |
+
perf_area = abs(_polygon_signed_area(polygon_raw)) if len(polygon_raw) >= 3 else 0.0
|
| 811 |
+
deg_area = abs(_polygon_signed_area(np.asarray(polygon_raw_deg, dtype=np.float64))) \
|
| 812 |
+
if len(polygon_raw_deg) >= 3 else 0.0
|
| 813 |
+
lost_frac = float(1.0 - deg_area / perf_area) if perf_area > 0 else 0.0
|
| 814 |
+
|
| 815 |
+
# reco B : budget sur la PERTE D'AIRE (ε), n en sortie. `n_sides_range`
|
| 816 |
+
# est CONSERVÉ comme plancher optionnel mais n'est plus le budget — cf.
|
| 817 |
+
# resample_reflex_aware (correctif 23/07/2026).
|
| 818 |
+
polygon_n = resample_reflex_aware(polygon_raw_deg, n_min=0,
|
| 819 |
+
max_area_loss=max_area_loss) # INPUT, ⊆ perfect
|
| 820 |
+
centroid = polygon_n.mean(axis=0)
|
| 821 |
+
# rotation-invariance : repère canonique PCA (sur le masque observé)
|
| 822 |
+
R = pca_canonical_rotation(mask_in)
|
| 823 |
+
polygon_canon = ((polygon_n - centroid) @ R.T).astype(np.float32)
|
| 824 |
+
pca_angle = float(np.degrees(np.arctan2(R[0, 1], R[0, 0]))) # axe majeur (repère image)
|
| 825 |
+
|
| 826 |
+
global_feat, side_feat = compute_fragment_features(
|
| 827 |
+
polygon_n, float(len(all_cells) * stud_size * stud_size),
|
| 828 |
+
mean_color, img, polygon_geom=polygon_canon
|
| 829 |
+
)
|
| 830 |
+
outer_per_side = find_outer_fragments_per_side(
|
| 831 |
+
polygon_n, centroid, fid, cell_to_frag, stud_size
|
| 832 |
+
)
|
| 833 |
+
# Translate fragment_id → contiguous node index
|
| 834 |
+
outer_per_side = [fid_to_idx[o] if o is not None and o in fid_to_idx else None
|
| 835 |
+
for o in outer_per_side]
|
| 836 |
+
|
| 837 |
+
rgba, bbox = make_fragment_rgba(all_cells, img, stud_size, mask=mask_in)
|
| 838 |
+
# rotate=False → curriculum L1 : fragments éclatés par TRANSLATION seule
|
| 839 |
+
# (palier le plus facile : YOLO/pose n'a pas à résoudre l'orientation).
|
| 840 |
+
angle = (random.choice([0, 90, 180, 270]) + random.uniform(-5.0, 5.0)) if rotate else 0.0
|
| 841 |
+
rgba_rot = rotate_rgba(rgba, angle) if rotate else rgba
|
| 842 |
+
|
| 843 |
+
fragment_data.append({
|
| 844 |
+
'node_id': fid_to_idx[fid],
|
| 845 |
+
'cells': all_cells,
|
| 846 |
+
'piece_ids': frag_pieces[fid],
|
| 847 |
+
'polygon_raw': polygon_raw.tolist(),
|
| 848 |
+
'polygon_raw_degraded': np.asarray(polygon_raw_deg).tolist(),
|
| 849 |
+
'degradation_lost_frac': lost_frac,
|
| 850 |
+
'erode_px': int(erode_px),
|
| 851 |
+
'n_holes': int(n_holes),
|
| 852 |
+
'polygon_n': polygon_n.tolist(),
|
| 853 |
+
'polygon_n_canonical': polygon_canon.tolist(), # repère PCA (invariant rotation)
|
| 854 |
+
'pca_angle_deg': pca_angle,
|
| 855 |
+
'global_features': global_feat,
|
| 856 |
+
'side_features': side_feat,
|
| 857 |
+
'mean_color': [float(c) for c in mean_color],
|
| 858 |
+
'rgba': rgba,
|
| 859 |
+
'rgba_rotated': rgba_rot,
|
| 860 |
+
'rotation_angle': float(angle),
|
| 861 |
+
'bbox_in_target': bbox,
|
| 862 |
+
'outer_per_side': outer_per_side,
|
| 863 |
+
})
|
| 864 |
+
|
| 865 |
+
# --- missing fragments: removed from the exploded INPUT, kept in the GT ---
|
| 866 |
+
missing_ids = set()
|
| 867 |
+
if missing_rng[1] > 0 and len(fragment_data) > 1:
|
| 868 |
+
n_missing = min(random.randint(missing_rng[0], missing_rng[1]),
|
| 869 |
+
len(fragment_data) - 1) # keep at least one fragment
|
| 870 |
+
if n_missing > 0:
|
| 871 |
+
missing_ids = set(random.sample(
|
| 872 |
+
[f['node_id'] for f in fragment_data], n_missing))
|
| 873 |
+
print(f"[forge] {len(missing_ids)} fragment(s) marked missing (input only)")
|
| 874 |
+
present = [f for f in fragment_data if f['node_id'] not in missing_ids]
|
| 875 |
+
|
| 876 |
+
if placement == 'explode':
|
| 877 |
+
# L0 : vue éclatée (positions relatives conservées, juste espacées)
|
| 878 |
+
print("[forge] placing fragments (exploded view, relative layout kept)…")
|
| 879 |
+
canvas, placements, canvas_size = explode_fragments(present, (W, H))
|
| 880 |
+
else:
|
| 881 |
+
print(f"[forge] placing fragments on {canvas_size[0]}x{canvas_size[1]} white canvas…")
|
| 882 |
+
requested_size = canvas_size
|
| 883 |
+
canvas, placements, canvas_size = place_fragments(present, canvas_size)
|
| 884 |
+
if canvas_size != requested_size:
|
| 885 |
+
print(f"[forge] canvas grown to {canvas_size[0]}x{canvas_size[1]} to fit all fragments")
|
| 886 |
+
for f, place in zip(present, placements):
|
| 887 |
+
f['placement'] = place
|
| 888 |
+
|
| 889 |
+
print("[forge] extracting YOLO polygons…")
|
| 890 |
+
yolo_lines = []
|
| 891 |
+
for fdata, place in zip(present, placements):
|
| 892 |
+
poly = extract_yolo_polygon(
|
| 893 |
+
fdata['rgba_rotated'], place['x'], place['y'], canvas_size
|
| 894 |
+
)
|
| 895 |
+
if poly is None:
|
| 896 |
+
continue
|
| 897 |
+
coords = ' '.join(f'{v:.6f}' for v in poly.flatten())
|
| 898 |
+
yolo_lines.append(f'0 {coords}')
|
| 899 |
+
|
| 900 |
+
print("[forge] writing outputs…")
|
| 901 |
+
pil.save(out_dir / 'target.png')
|
| 902 |
+
canvas.save(out_dir / 'source.png')
|
| 903 |
+
(out_dir / 'source_yolo.txt').write_text('\n'.join(yolo_lines) + '\n')
|
| 904 |
+
|
| 905 |
+
# Crops par fragment : toujours écrits (entrée du VLM).
|
| 906 |
+
for fdata in present:
|
| 907 |
+
Image.fromarray(fdata['rgba'], 'RGBA').save(
|
| 908 |
+
out_dir / 'fragments' / f"frag_{fdata['node_id']:02d}.png"
|
| 909 |
+
)
|
| 910 |
+
# pieces.json = DEBUG pur (pièces LEGO détectées) : n'entraîne PAS le YOLO (c'est
|
| 911 |
+
# source_yolo.txt), pas lu par visualize.py, lu NULLE PART → OFF par défaut,
|
| 912 |
+
# opt-in via --debug (~−0.8 Mo/instance, soit ~20 Go sur 25 000).
|
| 913 |
+
if debug:
|
| 914 |
+
pieces_json = [{
|
| 915 |
+
'cells': p['cells'],
|
| 916 |
+
'bbox_grid': p['bbox_grid'],
|
| 917 |
+
'color': p['color'],
|
| 918 |
+
} for p in pieces]
|
| 919 |
+
(out_dir / 'pieces.json').write_text(json.dumps(pieces_json, indent=2))
|
| 920 |
+
|
| 921 |
+
# Build two node representations:
|
| 922 |
+
# - 'gnn_input': features the GNN may legitimately consume (no target-frame
|
| 923 |
+
# position leak). polygon_n_canonical = polygon centred at the origin.
|
| 924 |
+
# - 'target_info': ground-truth position/orientation, used as supervision
|
| 925 |
+
# and for viz. NOT to be fed to the GNN as input.
|
| 926 |
+
def split_node(f):
|
| 927 |
+
gf = f['global_features'] # [area, perim, cx, cy, R, G, B, bbox_w, bbox_h]
|
| 928 |
+
# gnn_input = DOMAINE-AGNOSTIQUE (transférable LEGO→fresque) — aucune feature LEGO-only.
|
| 929 |
+
gnn_input = [gf[0], gf[1], gf[4], gf[5], gf[6], gf[7], gf[8]] # area, perim, R, G, B, bbox_w, bbox_h
|
| 930 |
+
return {
|
| 931 |
+
'gnn_input': gnn_input,
|
| 932 |
+
'n_sides': len(f['polygon_n_canonical']),
|
| 933 |
+
'polygon_n_canonical': f['polygon_n_canonical'], # repère PCA (invariant rotation)
|
| 934 |
+
'side_features': f['side_features'],
|
| 935 |
+
'target_info': {
|
| 936 |
+
'centroid': [gf[2], gf[3]],
|
| 937 |
+
'pca_angle_deg': f['pca_angle_deg'], # orientation retirée de l'input → GT
|
| 938 |
+
'polygon_n_absolute': f['polygon_n'],
|
| 939 |
+
'polygon_raw': f['polygon_raw'],
|
| 940 |
+
'mean_color': f['mean_color'],
|
| 941 |
+
'n_pieces': len(f['piece_ids']),
|
| 942 |
+
},
|
| 943 |
+
}
|
| 944 |
+
|
| 945 |
+
mean_lost = (float(np.mean([f['degradation_lost_frac'] for f in fragment_data]))
|
| 946 |
+
if fragment_data else 0.0)
|
| 947 |
+
nodes_full = [{'node_id': f['node_id'], 'missing': f['node_id'] in missing_ids,
|
| 948 |
+
**split_node(f)} for f in fragment_data]
|
| 949 |
+
# graph_fragments = INPUT: missing fragments are not observed → excluded.
|
| 950 |
+
nodes_input_only = [
|
| 951 |
+
{k: v for k, v in n.items() if k not in ('target_info', 'missing')}
|
| 952 |
+
for n in nodes_full if not n['missing']
|
| 953 |
+
]
|
| 954 |
+
|
| 955 |
+
edges_json = []
|
| 956 |
+
for (fa, fb), segments in adjacencies.items():
|
| 957 |
+
a = fid_to_idx[fa]
|
| 958 |
+
b = fid_to_idx[fb]
|
| 959 |
+
total_length = float(sum(
|
| 960 |
+
np.sqrt((s[2] - s[0]) ** 2 + (s[3] - s[1]) ** 2) for s in segments
|
| 961 |
+
))
|
| 962 |
+
angles = [float(np.arctan2(s[3] - s[1], s[2] - s[0])) for s in segments]
|
| 963 |
+
mean_angle = float(np.mean(angles))
|
| 964 |
+
a_sides = [i for i, o in enumerate(fragment_data[a]['outer_per_side']) if o == b]
|
| 965 |
+
b_sides = [i for i, o in enumerate(fragment_data[b]['outer_per_side']) if o == a]
|
| 966 |
+
edges_json.append({
|
| 967 |
+
'src': a, 'dst': b,
|
| 968 |
+
'features': [total_length, mean_angle, float(len(segments))],
|
| 969 |
+
'src_side_idx': a_sides,
|
| 970 |
+
'dst_side_idx': b_sides,
|
| 971 |
+
'polyline_raw': [list(s) for s in segments],
|
| 972 |
+
})
|
| 973 |
+
|
| 974 |
+
common_meta = {
|
| 975 |
+
'max_area_loss': max_area_loss, # ε — LE budget de reco B (fidélité)
|
| 976 |
+
'n_sides_range': list(n_sides_range), # OBSOLÈTE depuis le 23/07/2026 (info seule)
|
| 977 |
+
'n_sides_note': ("polygon_n_canonical / side_features ont une longueur "
|
| 978 |
+
"VARIABLE par nœud (cf. champ 'n_sides' de chaque nœud). "
|
| 979 |
+
"Reco B reflex-aware : polygon_n ⊆ polygon_raw (l'aire "
|
| 980 |
+
"n'est jamais gagnée). Plancher dur = nb de sommets "
|
| 981 |
+
"concaves. Pour un tenseur fixe côté GNN : padder au max "
|
| 982 |
+
"des 'n_sides' du dataset + masque de validité."),
|
| 983 |
+
'gnn_input_feature_names': ['area', 'perimeter', 'R', 'G', 'B', 'bbox_w', 'bbox_h'],
|
| 984 |
+
'feature_convention_note': (
|
| 985 |
+
"gnn_input = features 100 % DOMAINE-AGNOSTIQUES (transférables LEGO→fresque ; "
|
| 986 |
+
"aucune feature LEGO-only). polygon_n_canonical / side_features ont une longueur "
|
| 987 |
+
"VARIABLE (n_sides) → côté GNN : padder à n_max + masque (cf. features/collate.py), "
|
| 988 |
+
"ou set-encoder. area/perimeter/bbox sont en PIXELS (non normalisés) ; "
|
| 989 |
+
"normaliser avant transfert cross-échelle."),
|
| 990 |
+
'side_feature_names': ['length', 'angle', 'R', 'G', 'B'],
|
| 991 |
+
'edge_feature_names': ['shared_length', 'mean_angle', 'n_segments'],
|
| 992 |
+
'target_info_note': "centroid/polygon_n_absolute/polygon_raw are in the "
|
| 993 |
+
"target frame and MUST NOT be fed to the GNN as "
|
| 994 |
+
"input — they are supervision targets / metadata.",
|
| 995 |
+
'degradation': {
|
| 996 |
+
'erode_px': list(erode_rng),
|
| 997 |
+
'holes': list(holes_rng),
|
| 998 |
+
'missing': list(missing_rng),
|
| 999 |
+
'n_missing': len(missing_ids),
|
| 1000 |
+
'mean_lost_frac': round(mean_lost, 4), # sévérité moyenne (métrique dérivée)
|
| 1001 |
+
'note': "Dégradation appliquée à l'INPUT seulement ; la GT (arêtes + "
|
| 1002 |
+
"target_info.polygon_raw perfect) reste sur la partition intacte. "
|
| 1003 |
+
"mean_lost_frac = aire moyenne perdue (parfait→dégradé), métrique "
|
| 1004 |
+
"post-hoc de sévérité, pas un paramètre d'entrée.",
|
| 1005 |
+
},
|
| 1006 |
+
}
|
| 1007 |
+
|
| 1008 |
+
graph_complete = dict(common_meta, nodes=nodes_full, edges=edges_json)
|
| 1009 |
+
(out_dir / 'graph_complete.json').write_text(json.dumps(graph_complete, indent=2))
|
| 1010 |
+
|
| 1011 |
+
graph_fragments = dict(common_meta, nodes=nodes_input_only, edges=[])
|
| 1012 |
+
(out_dir / 'graph_fragments.json').write_text(json.dumps(graph_fragments, indent=2))
|
| 1013 |
+
|
| 1014 |
+
# gt_layout.json — ground truth for the RECONSTRUCTION metric (IoU / Q_pos).
|
| 1015 |
+
# Per fragment: its exact footprint in target.png (polygon_raw) + the reco-B
|
| 1016 |
+
# polygon, both in target coords, plus where it sits in the exploded source.
|
| 1017 |
+
# The GNN+pose / VLM prediction is scored against these placed polygons.
|
| 1018 |
+
gt_fragments = []
|
| 1019 |
+
for f in fragment_data:
|
| 1020 |
+
gf = f['global_features']
|
| 1021 |
+
place = f.get('placement')
|
| 1022 |
+
gt_fragments.append({
|
| 1023 |
+
'node_id': f['node_id'],
|
| 1024 |
+
'missing': f['node_id'] in missing_ids,
|
| 1025 |
+
'polygon_raw': f['polygon_raw'], # PERFECT footprint, target coords
|
| 1026 |
+
'polygon_raw_degraded': f['polygon_raw_degraded'], # what the model sees (eroded/holed)
|
| 1027 |
+
'degradation_lost_frac': round(f['degradation_lost_frac'], 4),
|
| 1028 |
+
'polygon_n': f['polygon_n'], # reco-B polygon, target coords
|
| 1029 |
+
'centroid': [gf[2], gf[3]],
|
| 1030 |
+
'bbox_target': list(f['bbox_in_target']),
|
| 1031 |
+
'area': gf[0],
|
| 1032 |
+
'source_placement': ( # location in source.png (exploded)
|
| 1033 |
+
{'x': place['x'], 'y': place['y'], 'w': place['w'], 'h': place['h'],
|
| 1034 |
+
'rotation_deg': f['rotation_angle']}
|
| 1035 |
+
if place is not None else None # None ⟺ missing fragment
|
| 1036 |
+
),
|
| 1037 |
+
})
|
| 1038 |
+
gt_layout = {
|
| 1039 |
+
'mosaic': out_dir.name,
|
| 1040 |
+
'target_size': [W, H],
|
| 1041 |
+
'canvas_size': [canvas_size[0], canvas_size[1]],
|
| 1042 |
+
'mean_degradation_lost_frac': round(mean_lost, 4),
|
| 1043 |
+
'note': ("Vérité terrain pour la métrique de RECONSTRUCTION (IoU / Q_pos). "
|
| 1044 |
+
"polygon_raw = footprint exact du fragment dans target.png ; on y "
|
| 1045 |
+
"compare la prédiction GNN+pose / VLM (fragment replacé). "
|
| 1046 |
+
"source_placement = pose du fragment dans source.png (éclaté)."),
|
| 1047 |
+
'fragments': gt_fragments,
|
| 1048 |
+
}
|
| 1049 |
+
(out_dir / 'gt_layout.json').write_text(json.dumps(gt_layout, indent=2))
|
| 1050 |
+
|
| 1051 |
+
# degradation.md — rapport lisible : paramètres tirés + aire perdue par fragment
|
| 1052 |
+
report = [
|
| 1053 |
+
f"# Dégradation — {out_dir.name}", "",
|
| 1054 |
+
"## Configuration (intervalles ; tirage aléatoire par fragment)",
|
| 1055 |
+
f"- Érosion morpho : `{list(erode_rng)}` px",
|
| 1056 |
+
f"- Trous internes : `{list(holes_rng)}` (aire totale plafonnée à 10 % du fragment)",
|
| 1057 |
+
f"- Fragments manquants : `{list(missing_rng)}` → **{len(missing_ids)}** tiré(s)",
|
| 1058 |
+
f"- Reco B — perte d'aire tolérée : `{max_area_loss:.1%}` (ε ; n en sortie, "
|
| 1059 |
+
f"plancher dur = #reflex)", "",
|
| 1060 |
+
"## Par fragment",
|
| 1061 |
+
"| node_id | manquant | érosion px | trous | n sommets | aire perdue % |",
|
| 1062 |
+
"|---:|:---:|---:|---:|---:|---:|",
|
| 1063 |
+
]
|
| 1064 |
+
for f in fragment_data:
|
| 1065 |
+
report.append(
|
| 1066 |
+
f"| {f['node_id']} | {'oui' if f['node_id'] in missing_ids else '—'} "
|
| 1067 |
+
f"| {f['erode_px']} | {f['n_holes']} | {len(f['polygon_n'])} "
|
| 1068 |
+
f"| {100 * f['degradation_lost_frac']:.1f} |")
|
| 1069 |
+
report += [
|
| 1070 |
+
"", "## Résumé",
|
| 1071 |
+
f"- **Aire moyenne perdue par fragment : {100 * mean_lost:.2f} %**",
|
| 1072 |
+
f"- Fragments manquants : {sorted(missing_ids) if missing_ids else 'aucun'}",
|
| 1073 |
+
f"- Total fragments : {len(fragment_data)} (présents dans l'input : {len(present)})",
|
| 1074 |
+
"",
|
| 1075 |
+
]
|
| 1076 |
+
(out_dir / 'degradation.md').write_text("\n".join(report))
|
| 1077 |
+
|
| 1078 |
+
print(f"[forge] done → {out_dir}")
|
| 1079 |
+
print(f" pieces={len(pieces)} fragments={len(fragment_data)} "
|
| 1080 |
+
f"adjacencies={len(edges_json)}")
|
| 1081 |
+
|
| 1082 |
+
|
| 1083 |
+
def parse_args():
|
| 1084 |
+
p = argparse.ArgumentParser(description=__doc__)
|
| 1085 |
+
p.add_argument('--input', default='exp1/canvas_mosaic.png',
|
| 1086 |
+
help='Path to canvas_mosaic.png (target)')
|
| 1087 |
+
p.add_argument('--out', default='dataset/mosaic_001',
|
| 1088 |
+
help='Output directory')
|
| 1089 |
+
p.add_argument('--seed', type=int, default=42)
|
| 1090 |
+
p.add_argument('--n-sides-min', type=int, default=16,
|
| 1091 |
+
help='Min target vertex count per fragment (reco B, variable n)')
|
| 1092 |
+
p.add_argument('--n-sides-max', type=int, default=24,
|
| 1093 |
+
help='Max target vertex count per fragment (floor = #reflex)')
|
| 1094 |
+
p.add_argument('--max-area-loss', type=float, default=MAX_AREA_LOSS_DEFAULT,
|
| 1095 |
+
help="reco B — fraction d'aire que polygon_n a le droit de PERDRE "
|
| 1096 |
+
"vs polygon_raw (défaut 0.01 = 1%%). C'est LE budget ; le nombre "
|
| 1097 |
+
"de sommets en découle. Remplace --n-sides-min/max (23/07/2026)")
|
| 1098 |
+
p.add_argument('--n-frag-min', type=int, default=10)
|
| 1099 |
+
p.add_argument('--n-frag-max', type=int, default=15)
|
| 1100 |
+
p.add_argument('--canvas-w', type=int, default=3500)
|
| 1101 |
+
p.add_argument('--canvas-h', type=int, default=3500)
|
| 1102 |
+
p.add_argument('--stud-size', type=int, default=None,
|
| 1103 |
+
help='Pixels per stud (auto-detected from image size if omitted)')
|
| 1104 |
+
p.add_argument('--no-rotation', action='store_true',
|
| 1105 |
+
help='Curriculum L1 : fragments éclatés SANS rotation (translation seule)')
|
| 1106 |
+
p.add_argument('--placement', choices=['scatter', 'explode'], default='scatter',
|
| 1107 |
+
help="explode = curriculum L0 : vue éclatée (positions relatives "
|
| 1108 |
+
"gardées, juste espacées) ; scatter = placement aléatoire (défaut)")
|
| 1109 |
+
p.add_argument('--frag-distribution', choices=['balanced', 'compact', 'ultracompact'],
|
| 1110 |
+
default='balanced',
|
| 1111 |
+
help="motif de découpe : balanced (BFS, défaut), compact (~rectangulaire), "
|
| 1112 |
+
"ou ultracompact (= compact, à utiliser sur des mosaïques 1×1 `--mode mono`). "
|
| 1113 |
+
"voronoi/clusters = à venir")
|
| 1114 |
+
p.add_argument('--debug', action='store_true',
|
| 1115 |
+
help="écrit aussi pieces.json (debug du détecteur de pièces, lu nulle "
|
| 1116 |
+
"part) ; OFF par défaut")
|
| 1117 |
+
# --- degradation knobs (curriculum) ; all default OFF → clean reco-B ---
|
| 1118 |
+
p.add_argument('--erode-px-min', type=int, default=0,
|
| 1119 |
+
help='Min boundary-erosion radius in px (DAFNE E)')
|
| 1120 |
+
p.add_argument('--erode-px-max', type=int, default=0,
|
| 1121 |
+
help='Max boundary-erosion radius in px (0 = no erosion)')
|
| 1122 |
+
p.add_argument('--holes-min', type=int, default=0,
|
| 1123 |
+
help='Min interior holes per fragment')
|
| 1124 |
+
p.add_argument('--holes-max', type=int, default=0,
|
| 1125 |
+
help='Max interior holes per fragment (0 = no holes)')
|
| 1126 |
+
p.add_argument('--missing-min', type=int, default=0,
|
| 1127 |
+
help='Min number of fragments removed from the input (DAFNE C)')
|
| 1128 |
+
p.add_argument('--missing-max', type=int, default=0,
|
| 1129 |
+
help='Max number of fragments removed from the input (0 = none)')
|
| 1130 |
+
return p.parse_args()
|
| 1131 |
+
|
| 1132 |
+
|
| 1133 |
+
def degrade_from_args(args):
|
| 1134 |
+
return {
|
| 1135 |
+
'erode_px': (args.erode_px_min, args.erode_px_max),
|
| 1136 |
+
'holes': (args.holes_min, args.holes_max),
|
| 1137 |
+
'missing': (args.missing_min, args.missing_max),
|
| 1138 |
+
}
|
| 1139 |
+
|
| 1140 |
+
|
| 1141 |
+
def main():
|
| 1142 |
+
args = parse_args()
|
| 1143 |
+
forge_one(
|
| 1144 |
+
target_path=args.input,
|
| 1145 |
+
out_dir=args.out,
|
| 1146 |
+
n_sides_range=(args.n_sides_min, args.n_sides_max),
|
| 1147 |
+
max_area_loss=args.max_area_loss,
|
| 1148 |
+
n_frag_range=(args.n_frag_min, args.n_frag_max),
|
| 1149 |
+
canvas_size=(args.canvas_w, args.canvas_h),
|
| 1150 |
+
stud_size=args.stud_size,
|
| 1151 |
+
seed=args.seed,
|
| 1152 |
+
degrade=degrade_from_args(args),
|
| 1153 |
+
rotate=not args.no_rotation,
|
| 1154 |
+
placement=args.placement,
|
| 1155 |
+
frag_distribution=args.frag_distribution,
|
| 1156 |
+
debug=args.debug,
|
| 1157 |
+
)
|
| 1158 |
+
|
| 1159 |
+
|
| 1160 |
+
if __name__ == '__main__':
|
| 1161 |
+
main()
|
code/mosaic2fragments/visualize.py
ADDED
|
@@ -0,0 +1,126 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
"""Quick visualization: overlay YOLO polygons on source.png to check alignment.
|
| 3 |
+
|
| 4 |
+
Also reports any sanity-check failures (fragment count vs YOLO lines, etc.).
|
| 5 |
+
"""
|
| 6 |
+
|
| 7 |
+
import argparse
|
| 8 |
+
import json
|
| 9 |
+
from pathlib import Path
|
| 10 |
+
|
| 11 |
+
import numpy as np
|
| 12 |
+
from PIL import Image, ImageDraw
|
| 13 |
+
|
| 14 |
+
|
| 15 |
+
COLORS = [
|
| 16 |
+
(231, 76, 60), (52, 152, 219), (241, 196, 15), (46, 204, 113),
|
| 17 |
+
(155, 89, 182), (52, 73, 94), (230, 126, 34), (26, 188, 156),
|
| 18 |
+
(192, 57, 43), (41, 128, 185), (243, 156, 18), (39, 174, 96),
|
| 19 |
+
(142, 68, 173), (44, 62, 80), (211, 84, 0), (22, 160, 133),
|
| 20 |
+
]
|
| 21 |
+
|
| 22 |
+
|
| 23 |
+
def visualize(dataset_dir):
|
| 24 |
+
d = Path(dataset_dir)
|
| 25 |
+
src = Image.open(d / 'source.png').convert('RGBA')
|
| 26 |
+
W, H = src.size
|
| 27 |
+
|
| 28 |
+
overlay = Image.new('RGBA', src.size, (0, 0, 0, 0))
|
| 29 |
+
draw = ImageDraw.Draw(overlay)
|
| 30 |
+
|
| 31 |
+
with open(d / 'source_yolo.txt') as f:
|
| 32 |
+
lines = [l.strip() for l in f if l.strip()]
|
| 33 |
+
print(f"YOLO lines: {len(lines)}")
|
| 34 |
+
|
| 35 |
+
for idx, line in enumerate(lines):
|
| 36 |
+
parts = line.split()
|
| 37 |
+
cls = parts[0]
|
| 38 |
+
coords = [float(x) for x in parts[1:]]
|
| 39 |
+
pts = [(coords[i] * W, coords[i + 1] * H) for i in range(0, len(coords), 2)]
|
| 40 |
+
color = COLORS[idx % len(COLORS)]
|
| 41 |
+
draw.polygon(pts, outline=color + (255,), fill=color + (90,))
|
| 42 |
+
|
| 43 |
+
out = Image.alpha_composite(src, overlay)
|
| 44 |
+
out_path = d / 'source_yolo_viz.png'
|
| 45 |
+
out.convert('RGB').save(out_path)
|
| 46 |
+
print(f"Wrote {out_path}")
|
| 47 |
+
|
| 48 |
+
# Sanity: cross-check with graph_complete.json
|
| 49 |
+
g = json.loads((d / 'graph_complete.json').read_text())
|
| 50 |
+
print(f"Graph nodes: {len(g['nodes'])}")
|
| 51 |
+
print(f"Graph edges: {len(g['edges'])}")
|
| 52 |
+
sizes = [n['target_info']['n_pieces'] for n in g['nodes']] # n_pieces (métadonnée GT)
|
| 53 |
+
print(f"Fragment sizes (n_pieces): min={min(sizes):.0f} max={max(sizes):.0f} "
|
| 54 |
+
f"mean={np.mean(sizes):.1f}")
|
| 55 |
+
|
| 56 |
+
|
| 57 |
+
def visualize_recob(dataset_dir):
|
| 58 |
+
"""Viz des DEUX pertes d'aire, en repère target (depuis `gt_layout.json`).
|
| 59 |
+
|
| 60 |
+
Par fragment, trois couches :
|
| 61 |
+
couleur pâle = `polygon_raw` (empreinte PARFAITE)
|
| 62 |
+
hachures GRISES fines = perte de DÉGRADATION (raw − raw_degraded :
|
| 63 |
+
érosion, trous — voulue, c'est l'input L3/L4)
|
| 64 |
+
hachures COULEUR marquées = perte d'ENCODAGE reco-B (raw_degraded −
|
| 65 |
+
polygon_n : budget --max-area-loss, ~ε)
|
| 66 |
+
trait noir = `polygon_n` (ce que voit le GNN)
|
| 67 |
+
NB : `source.png` n'est PAS affecté par la perte d'encodage (vrais pixels
|
| 68 |
+
découpés par le masque dégradé) — elle n'existe que dans polygon_n.
|
| 69 |
+
"""
|
| 70 |
+
import cv2
|
| 71 |
+
|
| 72 |
+
d = Path(dataset_dir)
|
| 73 |
+
gl = json.loads((d / 'gt_layout.json').read_text())
|
| 74 |
+
W, H = gl['target_size']
|
| 75 |
+
img = np.full((H, W, 3), 255, dtype=np.uint8)
|
| 76 |
+
|
| 77 |
+
hatch = np.zeros((H, W), dtype=np.uint8) # trame diagonale //
|
| 78 |
+
for c in range(-H, W, 14):
|
| 79 |
+
cv2.line(hatch, (c, 0), (c + H, H), 1, 2)
|
| 80 |
+
hatch2 = np.zeros((H, W), dtype=np.uint8) # trame croisée \\ serrée
|
| 81 |
+
for c in range(-H, W, 8):
|
| 82 |
+
cv2.line(hatch2, (c + H, 0), (c, H), 1, 1)
|
| 83 |
+
|
| 84 |
+
deg_total = enc_total = raw_total = 0.0
|
| 85 |
+
for idx, f in enumerate(gl['fragments']):
|
| 86 |
+
raw = np.array(f['polygon_raw'], dtype=np.int32)
|
| 87 |
+
deg = np.array(f['polygon_raw_degraded'], dtype=np.int32)
|
| 88 |
+
enc = np.array(f['polygon_n'], dtype=np.int32)
|
| 89 |
+
color = COLORS[idx % len(COLORS)]
|
| 90 |
+
pale = tuple(int(v + (255 - v) * 0.72) for v in color)
|
| 91 |
+
m = {}
|
| 92 |
+
for k, poly in (('raw', raw), ('deg', deg), ('enc', enc)):
|
| 93 |
+
m[k] = np.zeros((H, W), dtype=np.uint8)
|
| 94 |
+
if len(poly) >= 3:
|
| 95 |
+
cv2.fillPoly(m[k], [poly], 1)
|
| 96 |
+
lost_deg = (m['raw'] > 0) & (m['deg'] == 0) # dégradation (voulue)
|
| 97 |
+
lost_enc = (m['deg'] > 0) & (m['enc'] == 0) # encodage reco-B (ε)
|
| 98 |
+
img[m['raw'] > 0] = pale
|
| 99 |
+
img[lost_deg & (hatch2 > 0)] = (150, 150, 150)
|
| 100 |
+
img[lost_enc & (hatch > 0)] = color
|
| 101 |
+
if len(enc) >= 3:
|
| 102 |
+
cv2.polylines(img, [enc], True, (0, 0, 0), 2)
|
| 103 |
+
deg_total += float(lost_deg.sum())
|
| 104 |
+
enc_total += float(lost_enc.sum())
|
| 105 |
+
raw_total += float((m['raw'] > 0).sum())
|
| 106 |
+
|
| 107 |
+
out_path = d / 'recob_viz.png'
|
| 108 |
+
Image.fromarray(img).save(out_path)
|
| 109 |
+
print(f"Wrote {out_path} (perte dégradation [hachures grises] : "
|
| 110 |
+
f"{100 * deg_total / max(raw_total, 1):.2f} % ; perte encodage reco-B "
|
| 111 |
+
f"[hachures couleur] : {100 * enc_total / max(raw_total, 1):.2f} %)")
|
| 112 |
+
|
| 113 |
+
|
| 114 |
+
def parse_args():
|
| 115 |
+
p = argparse.ArgumentParser()
|
| 116 |
+
p.add_argument('--dir', default='dataset/mosaic_001')
|
| 117 |
+
p.add_argument('--recob', action='store_true',
|
| 118 |
+
help="écrit aussi recob_viz.png (perte d'encodage reco-B hachurée)")
|
| 119 |
+
return p.parse_args()
|
| 120 |
+
|
| 121 |
+
|
| 122 |
+
if __name__ == '__main__':
|
| 123 |
+
a = parse_args()
|
| 124 |
+
visualize(a.dir)
|
| 125 |
+
if a.recob:
|
| 126 |
+
visualize_recob(a.dir)
|
code/requirements.txt
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Dependencies (Python 3.11+). Install:
|
| 2 |
+
# python3 -m venv .venv && source .venv/bin/activate && pip install -r requirements.txt
|
| 3 |
+
numpy==2.2.6
|
| 4 |
+
Pillow==12.0.0
|
| 5 |
+
opencv-python==4.13.0.92
|
| 6 |
+
shapely==2.1.2 # optionnel (clip de sécurité reco-B) ; sans lui, fallback gracieux
|
| 7 |
+
networkx==3.5
|
| 8 |
+
datasets==4.3.0 # source d'images wikiart en streaming (forge_LAR_2mosaic/wikiart.py)
|
| 9 |
+
huggingface_hub==1.4.1 # upload de dataset vers le Hub (tools/hf_upload.py)
|
code/tools/hf_export_tars.py
ADDED
|
@@ -0,0 +1,103 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
"""Exporte un dataset curriculum (`<src>/L*/`) vers le HF Hub en **tar.gz par palier**.
|
| 3 |
+
|
| 4 |
+
Un dataset de 25 000 instances = ~500 k fichiers → un upload browsable est
|
| 5 |
+
impraticable (HF rame/refuse au-delà de ~100 k fichiers). On archive donc CHAQUE
|
| 6 |
+
palier (`L0_explode/`, …) en un seul `.tar.gz`, on l'uploade, puis on **supprime
|
| 7 |
+
l'archive locale** avant de passer au suivant → pic disque ≈ une seule archive.
|
| 8 |
+
|
| 9 |
+
En bonus on pousse les `gnn_meta.json` (browsables, petits) et, optionnellement,
|
| 10 |
+
on supprime d'anciens dossiers *legacy* du repo.
|
| 11 |
+
|
| 12 |
+
Prérequis : `huggingface-cli login` (ou HF_TOKEN). `HF_HUB_ENABLE_HF_TRANSFER=1`
|
| 13 |
+
recommandé pour la vitesse.
|
| 14 |
+
|
| 15 |
+
Exemple :
|
| 16 |
+
HF_HUB_ENABLE_HF_TRANSFER=1 python3 forge_mosaics/tools/hf_export_tars.py \
|
| 17 |
+
--src output/poly_balanced/k10-15 --repo-id <user>/<dataset-repo> \
|
| 18 |
+
--path-in-repo poly_balanced/k10-15 --card /tmp/hf_card.md
|
| 19 |
+
"""
|
| 20 |
+
import argparse
|
| 21 |
+
import os
|
| 22 |
+
import subprocess
|
| 23 |
+
from pathlib import Path
|
| 24 |
+
|
| 25 |
+
|
| 26 |
+
def parse_args():
|
| 27 |
+
p = argparse.ArgumentParser(description=__doc__,
|
| 28 |
+
formatter_class=argparse.RawDescriptionHelpFormatter)
|
| 29 |
+
p.add_argument("--src", default="output/balanced", help="dossier contenant les paliers L*/")
|
| 30 |
+
p.add_argument("--repo-id", required=True)
|
| 31 |
+
p.add_argument("--path-in-repo", default="balanced", help="préfixe cible dans le repo")
|
| 32 |
+
p.add_argument("--tmp", default="output/_tars", help="dossier temporaire des archives")
|
| 33 |
+
p.add_argument("--delete-legacy", nargs="*", default=[],
|
| 34 |
+
help="dossiers du repo à supprimer après upload (ex: clean degraded)")
|
| 35 |
+
p.add_argument("--card", default=None, help="README.md (carte) à pousser à la racine")
|
| 36 |
+
p.add_argument("--overwrite", action="store_true",
|
| 37 |
+
help="ré-uploade même si le tar existe déjà sur le repo (écrase)")
|
| 38 |
+
p.add_argument("--token", default=None)
|
| 39 |
+
return p.parse_args()
|
| 40 |
+
|
| 41 |
+
|
| 42 |
+
def main():
|
| 43 |
+
a = parse_args()
|
| 44 |
+
from huggingface_hub import HfApi
|
| 45 |
+
api = HfApi(token=a.token)
|
| 46 |
+
api.create_repo(a.repo_id, repo_type="dataset", exist_ok=True)
|
| 47 |
+
src = Path(a.src)
|
| 48 |
+
tmp = Path(a.tmp); tmp.mkdir(parents=True, exist_ok=True)
|
| 49 |
+
paliers = sorted(d.name for d in src.iterdir() if d.is_dir() and d.name.startswith("L"))
|
| 50 |
+
print(f"paliers: {paliers}")
|
| 51 |
+
|
| 52 |
+
if a.card:
|
| 53 |
+
api.upload_file(path_or_fileobj=a.card, path_in_repo="README.md",
|
| 54 |
+
repo_id=a.repo_id, repo_type="dataset",
|
| 55 |
+
commit_message="carte : curriculum 5 paliers (remplace 100mosaics legacy)")
|
| 56 |
+
print("[card] poussée")
|
| 57 |
+
|
| 58 |
+
for L in paliers: # gnn_meta browsables (petits)
|
| 59 |
+
meta = src / L / "gnn_meta.json"
|
| 60 |
+
if meta.exists():
|
| 61 |
+
api.upload_file(path_or_fileobj=str(meta),
|
| 62 |
+
path_in_repo=f"{a.path_in_repo}/{L}.gnn_meta.json",
|
| 63 |
+
repo_id=a.repo_id, repo_type="dataset",
|
| 64 |
+
commit_message=f"meta {L}")
|
| 65 |
+
print(f"[meta] {L}")
|
| 66 |
+
|
| 67 |
+
for L in paliers: # 1 archive par palier
|
| 68 |
+
repo_path = f"{a.path_in_repo}/{L}.tar.gz"
|
| 69 |
+
if not a.overwrite and api.file_exists(a.repo_id, repo_path, repo_type="dataset"):
|
| 70 |
+
print(f"[skip] {L} (déjà sur le repo) — resume") # relance idempotente
|
| 71 |
+
continue
|
| 72 |
+
tar = tmp / f"{L}.tar.gz"
|
| 73 |
+
print(f"[tar] {L} …", flush=True)
|
| 74 |
+
# Les tars du 24-26/07/2026 contenaient un AppleDouble `._<fichier>`
|
| 75 |
+
# (163 o de xattrs macOS) à côté de CHAQUE fichier — bruit inutile dans
|
| 76 |
+
# un dataset public. Ceinture (COPYFILE_DISABLE) + bretelles
|
| 77 |
+
# (--no-mac-metadata, bsdtar) ; l'option est retirée si tar la rejette.
|
| 78 |
+
cmd = ["tar", "--no-mac-metadata", "-czf", str(tar), "-C", str(src), L]
|
| 79 |
+
env = {**os.environ, "COPYFILE_DISABLE": "1"}
|
| 80 |
+
if subprocess.run(cmd, env=env).returncode != 0:
|
| 81 |
+
subprocess.run(["tar", "-czf", str(tar), "-C", str(src), L],
|
| 82 |
+
check=True, env=env)
|
| 83 |
+
gb = tar.stat().st_size / 1e9
|
| 84 |
+
print(f"[upload] {L} ({gb:.1f} Go) …", flush=True)
|
| 85 |
+
api.upload_file(path_or_fileobj=str(tar), path_in_repo=repo_path,
|
| 86 |
+
repo_id=a.repo_id, repo_type="dataset",
|
| 87 |
+
commit_message=f"palier {L} (tar.gz)")
|
| 88 |
+
tar.unlink() # libère le disque avant le suivant
|
| 89 |
+
print(f"[done] {L}", flush=True)
|
| 90 |
+
|
| 91 |
+
for folder in a.delete_legacy:
|
| 92 |
+
try:
|
| 93 |
+
api.delete_folder(path_in_repo=folder, repo_id=a.repo_id, repo_type="dataset",
|
| 94 |
+
commit_message=f"retrait legacy {folder}")
|
| 95 |
+
print(f"[legacy] supprimé : {folder}")
|
| 96 |
+
except Exception as e:
|
| 97 |
+
print(f"[legacy] échec {folder}: {e}")
|
| 98 |
+
|
| 99 |
+
print(f"EXPORT FINI → https://huggingface.co/datasets/{a.repo_id}")
|
| 100 |
+
|
| 101 |
+
|
| 102 |
+
if __name__ == "__main__":
|
| 103 |
+
main()
|
code/tools/hf_upload.py
ADDED
|
@@ -0,0 +1,53 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
"""Upload d'un dossier vers le HuggingFace Hub comme **dataset** — gratuit.
|
| 3 |
+
|
| 4 |
+
Le HF Hub héberge gratuitement les datasets (public ou privé, quotas larges).
|
| 5 |
+
C'est la façon recommandée de partager nos données avec les camarades : le CODE
|
| 6 |
+
reste sur GitHub (léger), les DONNÉES (mosaïques, datasets fragmentés) vont sur le
|
| 7 |
+
Hub — au lieu de gonfler le dépôt git.
|
| 8 |
+
|
| 9 |
+
Prérequis :
|
| 10 |
+
pip install huggingface_hub
|
| 11 |
+
huggingface-cli login # token gratuit : https://huggingface.co/settings/tokens
|
| 12 |
+
# (ou export HF_TOKEN=hf_xxx, ou --token hf_xxx)
|
| 13 |
+
|
| 14 |
+
Exemples :
|
| 15 |
+
# un dataset fragmenté, public
|
| 16 |
+
python3 forge_mosaics/tools/hf_upload.py --folder output/mosaics \
|
| 17 |
+
--repo-id <user>/<dataset-repo>
|
| 18 |
+
|
| 19 |
+
# un dataset privé
|
| 20 |
+
python3 forge_mosaics/tools/hf_upload.py --folder output/dataset \
|
| 21 |
+
--repo-id <user>/<dataset-repo> --private
|
| 22 |
+
"""
|
| 23 |
+
import argparse
|
| 24 |
+
|
| 25 |
+
|
| 26 |
+
def parse_args():
|
| 27 |
+
p = argparse.ArgumentParser(description="Upload d'un dossier vers le HF Hub (dataset).")
|
| 28 |
+
p.add_argument("--folder", required=True, help="dossier local à uploader")
|
| 29 |
+
p.add_argument("--repo-id", required=True, help="ex: <user>/<dataset-repo>")
|
| 30 |
+
p.add_argument("--private", action="store_true", help="dataset privé (défaut : public)")
|
| 31 |
+
p.add_argument("--token", default=None, help="token HF (sinon : huggingface-cli login / HF_TOKEN)")
|
| 32 |
+
p.add_argument("--commit", default="upload via tools/hf_upload.py", help="message de commit Hub")
|
| 33 |
+
p.add_argument("--path-in-repo", default=None,
|
| 34 |
+
help="sous-dossier cible dans le repo (ex: clean) → plusieurs dossiers dans 1 dataset")
|
| 35 |
+
p.add_argument("--allow-patterns", nargs="*", default=None,
|
| 36 |
+
help="ne pousser que ces motifs (ex: '*.json' '*.png')")
|
| 37 |
+
return p.parse_args()
|
| 38 |
+
|
| 39 |
+
|
| 40 |
+
def main():
|
| 41 |
+
a = parse_args()
|
| 42 |
+
from huggingface_hub import HfApi # import tardif : message clair si absent
|
| 43 |
+
api = HfApi(token=a.token)
|
| 44 |
+
api.create_repo(repo_id=a.repo_id, repo_type="dataset",
|
| 45 |
+
private=a.private, exist_ok=True)
|
| 46 |
+
api.upload_folder(folder_path=a.folder, repo_id=a.repo_id, repo_type="dataset",
|
| 47 |
+
path_in_repo=a.path_in_repo,
|
| 48 |
+
commit_message=a.commit, allow_patterns=a.allow_patterns)
|
| 49 |
+
print(f"OK → https://huggingface.co/datasets/{a.repo_id}")
|
| 50 |
+
|
| 51 |
+
|
| 52 |
+
if __name__ == "__main__":
|
| 53 |
+
main()
|
code/tools/qa_contact_sheets.py
ADDED
|
@@ -0,0 +1,95 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
"""Planches de contact pour la revue qualitative de l'échantillon `qa_sample_hf.py`.
|
| 3 |
+
|
| 4 |
+
Deux familles de planches :
|
| 5 |
+
|
| 6 |
+
* `sheet_<dataset>_<palier>.png` — 50 cellules `target | source` côte à côte
|
| 7 |
+
(une par mosaïque échantillonnée) : on balaye d'un coup d'œil les erreurs de
|
| 8 |
+
placement, les mosaïques dégénérées (aplats), la sévérité de la dégradation.
|
| 9 |
+
* `targets_<dataset>.png` — les 50 targets seuls, en grille dense : c'est la vue
|
| 10 |
+
qui fait sauter aux yeux une **répétition** de peinture.
|
| 11 |
+
|
| 12 |
+
Chaque cellule porte les 8 premiers caractères de l'uuid → un doublon visuel se
|
| 13 |
+
retrouve dans `manifest.json`.
|
| 14 |
+
"""
|
| 15 |
+
import argparse
|
| 16 |
+
from pathlib import Path
|
| 17 |
+
|
| 18 |
+
from PIL import Image, ImageDraw
|
| 19 |
+
|
| 20 |
+
# les `source.png` de L0_explode montent à 16128² (260 Mpx) → au-delà du garde-fou
|
| 21 |
+
# anti-« decompression bomb » de PIL, qui lèverait DecompressionBombError.
|
| 22 |
+
Image.MAX_IMAGE_PIXELS = None
|
| 23 |
+
|
| 24 |
+
TH = 190 # côté d'une vignette
|
| 25 |
+
PAD = 6
|
| 26 |
+
LAB = 13 # bande de label sous la cellule
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
def _thumb(p, size=TH):
|
| 30 |
+
im = Image.open(p).convert("RGB")
|
| 31 |
+
im.thumbnail((size, size), Image.LANCZOS)
|
| 32 |
+
bg = Image.new("RGB", (size, size), (245, 245, 245))
|
| 33 |
+
bg.paste(im, ((size - im.width) // 2, (size - im.height) // 2))
|
| 34 |
+
return bg
|
| 35 |
+
|
| 36 |
+
|
| 37 |
+
def sheet(mosaics, out, title, pair=True, cols=5):
|
| 38 |
+
"""`pair=True` → cellule target|source ; sinon target seul (grille dense)."""
|
| 39 |
+
cw = TH * 2 + PAD if pair else TH
|
| 40 |
+
ch = TH + LAB
|
| 41 |
+
rows = (len(mosaics) + cols - 1) // cols
|
| 42 |
+
W = cols * (cw + PAD) + PAD
|
| 43 |
+
H = rows * (ch + PAD) + PAD + 26
|
| 44 |
+
sh = Image.new("RGB", (W, H), (255, 255, 255))
|
| 45 |
+
d = ImageDraw.Draw(sh)
|
| 46 |
+
d.text((PAD, 8), title, fill=(0, 0, 0))
|
| 47 |
+
|
| 48 |
+
for i, m in enumerate(sorted(mosaics)):
|
| 49 |
+
x = PAD + (i % cols) * (cw + PAD)
|
| 50 |
+
y = 26 + PAD + (i // cols) * (ch + PAD)
|
| 51 |
+
try:
|
| 52 |
+
sh.paste(_thumb(m / "target.png"), (x, y))
|
| 53 |
+
if pair and (m / "source.png").exists():
|
| 54 |
+
sh.paste(_thumb(m / "source.png"), (x + TH + PAD, y))
|
| 55 |
+
except (FileNotFoundError, OSError) as e:
|
| 56 |
+
d.text((x + 4, y + 4), f"ERR {e.__class__.__name__}", fill=(200, 0, 0))
|
| 57 |
+
d.text((x + 2, y + TH + 1), m.name.replace("mosaic_", "")[:8], fill=(90, 90, 90))
|
| 58 |
+
out.parent.mkdir(parents=True, exist_ok=True)
|
| 59 |
+
sh.save(out)
|
| 60 |
+
return out
|
| 61 |
+
|
| 62 |
+
|
| 63 |
+
def main():
|
| 64 |
+
p = argparse.ArgumentParser()
|
| 65 |
+
p.add_argument("--sample", default="output/qa_sample")
|
| 66 |
+
p.add_argument("--out", default="output/qa_sheets")
|
| 67 |
+
a = p.parse_args()
|
| 68 |
+
root, out = Path(a.sample), Path(a.out)
|
| 69 |
+
|
| 70 |
+
paliers = sorted({d.name for d in root.glob("*/*/L*") if d.is_dir()})
|
| 71 |
+
datasets = sorted({f"{d.parent.name}/{d.name}" for d in root.glob("*/k*") if d.is_dir()})
|
| 72 |
+
|
| 73 |
+
for ds in datasets:
|
| 74 |
+
tag = ds.replace("/", "_")
|
| 75 |
+
for L in paliers:
|
| 76 |
+
mos = [m for m in (root / ds / L).iterdir() if m.is_dir()] \
|
| 77 |
+
if (root / ds / L).is_dir() else []
|
| 78 |
+
if not mos:
|
| 79 |
+
continue
|
| 80 |
+
o = sheet(mos, out / f"sheet_{tag}_{L}.png", f"{ds} {L} ({len(mos)} mosaïques — gauche: target / droite: source)")
|
| 81 |
+
print("[sheet]", o)
|
| 82 |
+
# vue « répétitions » : targets seuls, pris sur le premier palier dispo
|
| 83 |
+
for L in paliers:
|
| 84 |
+
mos = [m for m in (root / ds / L).iterdir() if m.is_dir()] \
|
| 85 |
+
if (root / ds / L).is_dir() else []
|
| 86 |
+
if mos:
|
| 87 |
+
o = sheet(mos, out / f"targets_{tag}.png",
|
| 88 |
+
f"{ds} — targets seuls ({len(mos)}) — repérage des répétitions",
|
| 89 |
+
pair=False, cols=10)
|
| 90 |
+
print("[targets]", o)
|
| 91 |
+
break
|
| 92 |
+
|
| 93 |
+
|
| 94 |
+
if __name__ == "__main__":
|
| 95 |
+
main()
|
code/tools/qa_overlays.py
ADDED
|
@@ -0,0 +1,55 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
"""Overlays de contrôle sur l'échantillon : YOLO-Seg + reco-B.
|
| 3 |
+
|
| 4 |
+
Pour chaque (dataset, palier) de `output/qa_sample`, prend la **même** mosaïque
|
| 5 |
+
(uuid partagé par tous les paliers → comparaison L0→L4 à contenu constant) et
|
| 6 |
+
écrit dans le dossier de la mosaïque :
|
| 7 |
+
|
| 8 |
+
* `source_yolo_viz.png` — polygones `source_yolo.txt` repeints sur `source.png`
|
| 9 |
+
(contrôle : les masques collent-ils aux fragments ? la normalisation par le
|
| 10 |
+
canvas agrandi est-elle correcte ?)
|
| 11 |
+
* `recob_viz.png` — perte de dégradation (hachures grises) vs perte d'encodage
|
| 12 |
+
reco-B (hachures couleur), en repère target.
|
| 13 |
+
|
| 14 |
+
Les copies sont rassemblées à plat dans `--out` pour la revue.
|
| 15 |
+
"""
|
| 16 |
+
import argparse
|
| 17 |
+
import shutil
|
| 18 |
+
import sys
|
| 19 |
+
from pathlib import Path
|
| 20 |
+
|
| 21 |
+
from PIL import Image
|
| 22 |
+
|
| 23 |
+
sys.path.insert(0, str(Path(__file__).resolve().parent.parent / "mosaic2fragments"))
|
| 24 |
+
Image.MAX_IMAGE_PIXELS = None # L0 monte à 260 Mpx (cf. qa_contact_sheets)
|
| 25 |
+
|
| 26 |
+
import visualize # noqa: E402
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
def main():
|
| 30 |
+
p = argparse.ArgumentParser()
|
| 31 |
+
p.add_argument("--sample", default="output/qa_sample")
|
| 32 |
+
p.add_argument("--out", default="output/qa_overlays")
|
| 33 |
+
p.add_argument("--per-cell", type=int, default=1, help="mosaïques par dataset×palier")
|
| 34 |
+
a = p.parse_args()
|
| 35 |
+
root, out = Path(a.sample), Path(a.out)
|
| 36 |
+
out.mkdir(parents=True, exist_ok=True)
|
| 37 |
+
|
| 38 |
+
cells = sorted(d for d in root.glob("*/k*/L*") if d.is_dir())
|
| 39 |
+
for c in cells:
|
| 40 |
+
tag = "_".join(c.parts[-3:])
|
| 41 |
+
for m in sorted(d for d in c.iterdir() if d.is_dir())[:a.per_cell]:
|
| 42 |
+
try:
|
| 43 |
+
visualize.visualize(m)
|
| 44 |
+
visualize.visualize_recob(m)
|
| 45 |
+
except Exception as e: # noqa: BLE001
|
| 46 |
+
print(f"✗ {tag}/{m.name[:14]} : {type(e).__name__}: {e}", flush=True)
|
| 47 |
+
continue
|
| 48 |
+
for f, suf in (("source_yolo_viz.png", "yolo"), ("recob_viz.png", "recob")):
|
| 49 |
+
if (m / f).exists():
|
| 50 |
+
shutil.copy(m / f, out / f"{suf}_{tag}_{m.name[7:15]}.png")
|
| 51 |
+
print(f"✓ {tag}/{m.name[7:15]}", flush=True)
|
| 52 |
+
|
| 53 |
+
|
| 54 |
+
if __name__ == "__main__":
|
| 55 |
+
main()
|
code/tools/qa_sample_hf.py
ADDED
|
@@ -0,0 +1,119 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
"""Échantillonne ~1 % de chaque dataset publié sur le Hub, **sans télécharger les tars**.
|
| 3 |
+
|
| 4 |
+
Un `.tar.gz` n'est pas seekable : on ne peut pas atteindre un membre au milieu
|
| 5 |
+
sans décompresser tout ce qui précède. On lit donc le flux HTTP au fil de l'eau
|
| 6 |
+
(`tarfile` mode `r|gz`) et on **coupe la connexion** dès que N dossiers
|
| 7 |
+
`mosaic_<uuid>/` ont été vus → on ne transfère que ~N/5000 des octets.
|
| 8 |
+
|
| 9 |
+
⚠️ Biais assumé : l'échantillon est un *préfixe* de l'archive. Les noms étant des
|
| 10 |
+
uuid4, il est non biaisé vis-à-vis du **contenu** des peintures, mais il suit
|
| 11 |
+
l'ordre de création des dossiers (donc l'ordre de génération).
|
| 12 |
+
|
| 13 |
+
Les crops `fragments/*.png` sont ignorés (inutiles à la revue visuelle, ~la
|
| 14 |
+
moitié du volume disque).
|
| 15 |
+
"""
|
| 16 |
+
import argparse
|
| 17 |
+
import concurrent.futures as cf
|
| 18 |
+
import io
|
| 19 |
+
import json
|
| 20 |
+
import tarfile
|
| 21 |
+
from pathlib import Path
|
| 22 |
+
|
| 23 |
+
import requests
|
| 24 |
+
from huggingface_hub import get_token, hf_hub_url
|
| 25 |
+
|
| 26 |
+
REPO = "<user>/<dataset-repo>" # défaut ; surchargeable par --repo-id
|
| 27 |
+
DATASETS = [f"{g}/{k}" for g in ("poly_balanced", "mono_compact")
|
| 28 |
+
for k in ("k02", "k03", "k05", "k10-15")]
|
| 29 |
+
PALIERS = ["L0_explode", "L1_translation", "L2_rotation", "L3_light", "L4_strong"]
|
| 30 |
+
KEEP = {"target.png", "source.png", "source_yolo.txt", "graph_fragments.json",
|
| 31 |
+
"graph_complete.json", "gt_layout.json", "degradation.md", "gnn_ready.npz"}
|
| 32 |
+
|
| 33 |
+
|
| 34 |
+
def sample_one(ds, palier, n, out_root, repo):
|
| 35 |
+
"""Extrait les `n` premières mosaïques de `<ds>/<palier>.tar.gz`. → (dl_bytes, uuids)"""
|
| 36 |
+
url = hf_hub_url(repo, f"{ds}/{palier}.tar.gz", repo_type="dataset")
|
| 37 |
+
tok = get_token()
|
| 38 |
+
hdr = {"Authorization": f"Bearer {tok}"} if tok else {}
|
| 39 |
+
dest = Path(out_root) / ds / palier
|
| 40 |
+
dest.mkdir(parents=True, exist_ok=True)
|
| 41 |
+
|
| 42 |
+
seen, done = [], False
|
| 43 |
+
with requests.get(url, headers=hdr, stream=True, timeout=120) as r:
|
| 44 |
+
r.raise_for_status()
|
| 45 |
+
raw = r.raw # urllib3, décompression gzip faite par tarfile
|
| 46 |
+
counted = _Counter(raw)
|
| 47 |
+
try:
|
| 48 |
+
with tarfile.open(fileobj=counted, mode="r|gz") as tf:
|
| 49 |
+
for m in tf:
|
| 50 |
+
parts = m.name.split("/")
|
| 51 |
+
# `tar` sous macOS embarque des AppleDouble `._x` (xattrs) : à jeter.
|
| 52 |
+
# Et l'archive liste d'abord les 5000 répertoires, PUIS les fichiers
|
| 53 |
+
# groupés par mosaïque → on compte les uuid sur les fichiers utiles.
|
| 54 |
+
if any(c.startswith("._") for c in parts):
|
| 55 |
+
continue
|
| 56 |
+
if len(parts) != 3 or not m.isfile() or parts[-1] not in KEEP:
|
| 57 |
+
continue
|
| 58 |
+
uid = parts[1]
|
| 59 |
+
if uid not in seen:
|
| 60 |
+
if len(seen) >= n: # on entre dans la (n+1)-ème → stop
|
| 61 |
+
done = True
|
| 62 |
+
break
|
| 63 |
+
seen.append(uid)
|
| 64 |
+
tgt = dest / uid / parts[-1]
|
| 65 |
+
tgt.parent.mkdir(parents=True, exist_ok=True)
|
| 66 |
+
tgt.write_bytes(tf.extractfile(m).read())
|
| 67 |
+
except (tarfile.ReadError, OSError) as e:
|
| 68 |
+
if not done:
|
| 69 |
+
raise RuntimeError(f"{ds}/{palier}: {e}") from e
|
| 70 |
+
return counted.n, seen
|
| 71 |
+
|
| 72 |
+
|
| 73 |
+
class _Counter(io.RawIOBase):
|
| 74 |
+
"""Wrapper qui compte les octets réellement transférés."""
|
| 75 |
+
|
| 76 |
+
def __init__(self, raw):
|
| 77 |
+
self.raw, self.n = raw, 0
|
| 78 |
+
|
| 79 |
+
def read(self, size=-1):
|
| 80 |
+
b = self.raw.read(size)
|
| 81 |
+
self.n += len(b)
|
| 82 |
+
return b
|
| 83 |
+
|
| 84 |
+
def readable(self):
|
| 85 |
+
return True
|
| 86 |
+
|
| 87 |
+
|
| 88 |
+
def main():
|
| 89 |
+
p = argparse.ArgumentParser()
|
| 90 |
+
p.add_argument("--n", type=int, default=50, help="mosaïques par palier (1%% de 5000)")
|
| 91 |
+
p.add_argument("--out", default="output/qa_sample")
|
| 92 |
+
p.add_argument("--repo-id", default=REPO)
|
| 93 |
+
p.add_argument("--jobs", type=int, default=6)
|
| 94 |
+
a = p.parse_args()
|
| 95 |
+
|
| 96 |
+
jobs = [(ds, L) for ds in DATASETS for L in PALIERS]
|
| 97 |
+
manifest, total = {}, 0
|
| 98 |
+
with cf.ThreadPoolExecutor(a.jobs) as ex:
|
| 99 |
+
futs = {ex.submit(sample_one, ds, L, a.n, a.out, a.repo_id): (ds, L)
|
| 100 |
+
for ds, L in jobs}
|
| 101 |
+
for i, f in enumerate(cf.as_completed(futs), 1):
|
| 102 |
+
ds, L = futs[f]
|
| 103 |
+
try:
|
| 104 |
+
nb, uids = f.result()
|
| 105 |
+
except Exception as e: # noqa: BLE001 — on veut la suite
|
| 106 |
+
print(f"[{i:2d}/40] ✗ {ds}/{L} : {e}", flush=True)
|
| 107 |
+
continue
|
| 108 |
+
total += nb
|
| 109 |
+
manifest[f"{ds}/{L}"] = uids
|
| 110 |
+
print(f"[{i:2d}/40] ✓ {ds}/{L} : {len(uids)} mosaïques, {nb/1e6:.0f} Mo",
|
| 111 |
+
flush=True)
|
| 112 |
+
|
| 113 |
+
Path(a.out).mkdir(parents=True, exist_ok=True)
|
| 114 |
+
(Path(a.out) / "manifest.json").write_text(json.dumps(manifest, indent=1))
|
| 115 |
+
print(f"\nTotal transféré : {total/1e9:.2f} Go → {a.out}")
|
| 116 |
+
|
| 117 |
+
|
| 118 |
+
if __name__ == "__main__":
|
| 119 |
+
main()
|