code/src/preprocessing/view_canonicalization.py

#!/usr/bin/env python3
"""View-aware rotation canonicalization (heuristic).

view_ile_hint tablet_view alanı zaten var. Şimdi: aspect ratio + view hint ile
doğru orientation'a rotate et (offline pass).
"""
import json, os
from pathlib import Path
from concurrent.futures import ProcessPoolExecutor
import numpy as np
from PIL import Image

ROOT = Path("/arf/scratch/stakan/hitit-proje")
SOURCES = ROOT / "datasets" / "sources"

def infer_canonical_rotation(img_path, view_hint):
    """
    View hint + aspect ratio ile kanonik rotation tespit et.
    Return: rotation_degrees (0, 90, 180, 270) ve landscape_score
    """
    try:
        with Image.open(img_path) as img:
            w, h = img.size
        aspect = w / max(h, 1)
        # Tablet'ler genelde landscape veya kare. Çok aspect lehinde olan bir view'a göre:
        # - obverse/reverse: landscape (aspect>1) tipik
        # - edges: portrait (aspect<1) veya ince
        # Bu basit heuristic; asıl model-based classifier ayrı.
        if view_hint in ('obverse', 'reverse'):
            return 0 if aspect >= 0.7 else 90  # portrait'i landscape'e çevir
        elif view_hint in ('left_edge', 'right_edge'):
            return 90 if aspect > 1 else 0
        return 0
    except Exception:
        return 0

def worker(item):
    rid, path, view = item
    return (rid, infer_canonical_rotation(path, view))

def main():
    items = []
    for d in sorted(SOURCES.iterdir()):
        if not d.is_dir(): continue
        mp = d / "manifest.jsonl"
        if not mp.exists(): continue
        with open(mp) as f:
            for line in f:
                r = json.loads(line)
                p = r.get('path')
                if p and r.get('storage') == 'fs' and r.get('integrity_ok') is True:
                    items.append((r['id'], p, r.get('view', 'unknown')))
    # Unique by path (daha hızlı)
    seen_paths = set()
    uniq = []
    for rid, path, view in items:
        if path in seen_paths: continue
        seen_paths.add(path)
        uniq.append((rid, path, view))
    print(f"Tarama: {len(uniq):,} unique image")
    
    results = {}
    with ProcessPoolExecutor(max_workers=200) as ex:
        for rid, rot in ex.map(worker, uniq, chunksize=500):
            results[rid] = rot
    
    # Manifest güncelle
    from collections import Counter
    rot_counts = Counter()
    for d in sorted(SOURCES.iterdir()):
        if not d.is_dir(): continue
        for mf in ['manifest.jsonl', 'manifest_classification.jsonl', 'manifest_detection.jsonl']:
            mp = d / mf
            if not mp.exists(): continue
            records = []
            with open(mp) as f:
                for line in f:
                    r = json.loads(line)
                    rot = results.get(r.get('id'))
                    if rot is not None:
                        r['canonical_rotation_deg'] = int(rot)
                        if mf == 'manifest.jsonl':
                            rot_counts[rot] += 1
                    records.append(r)
            with open(mp, 'w') as f:
                for r in records:
                    f.write(json.dumps(r, ensure_ascii=False) + '\n')
    
    print(f"Rotation counts: {dict(rot_counts)}")
    
    with open(ROOT / "datasets" / "processed" / "view_canonicalization_report.json", 'w') as f:
        json.dump({
            "method": "heuristic (aspect + view_hint)",
            "n_images_processed": len(results),
            "rotation_distribution": dict(rot_counts),
            "note": "Heuristic v1. Model-based classifier (MobileNetV3) için ayrı training gerekir."
        }, f, indent=2, ensure_ascii=False)

if __name__ == '__main__':
    main()