#!/usr/bin/env python3 """Export Hub Parquet rows to YOLO without resizing or re-encoding images. pip install datasets pyarrow huggingface_hub python export_yolo.py --repo-id OWNER/DATASET --output ./aircraft-yolo python export_yolo.py --repo-id OWNER/DATASET --output ./plane-yolo --single-class python export_yolo.py --local-dir ./downloaded-hub-repo --output ./aircraft-yolo The default retains size classes 0/1/2. --single-class maps all three to plane=0. The streaming Image column is explicitly decoded as bytes, so even large native images are written without pixel decoding. Some individual native images exceed 1 GiB; sufficient memory and disk space for the encoded row are still required. """ import argparse import collections import hashlib import json import math from pathlib import Path NAMES = {0: 'small_plane', 1: 'middle_plane', 2: 'big_plane'} def annotation_lists(row): objects = row.get('objects') if objects is None: raise ValueError('Missing objects field; expected objects.yolo_bbox and category/category_id.') if isinstance(objects, dict): boxes = objects.get('yolo_bbox') categories = objects.get('category', objects.get('category_id', objects.get('categories'))) elif isinstance(objects, list): boxes = [obj['yolo_bbox'] for obj in objects] categories = [obj.get('category', obj.get('category_id')) for obj in objects] else: raise ValueError('objects must be a dict of lists or a list of object dictionaries.') if boxes is None or categories is None or len(boxes) != len(categories): raise ValueError('Missing or inconsistent object categories and normalized YOLO boxes.') return boxes, categories def export_one_row(row, split, output, single_class=False): image = row.get('image') if isinstance(image, dict): encoded = image.get('bytes') encoded_path = image.get('path') elif isinstance(image, (bytes, bytearray, memoryview)): encoded = image encoded_path = None else: encoded = row.get('image_bytes') encoded_path = None if encoded is None: raise ValueError('No embedded image bytes. Ensure the Image column uses decode=False.') if not isinstance(encoded, (bytes, bytearray, memoryview)): raise ValueError('Image bytes have an unexpected type; pixel decoding is not supported.') filename = row.get('file_name') or row.get('filename') or row.get('original_filename') or encoded_path if not filename: raise ValueError('Missing filename.') # Keep only the basename from a parquet image.path while rejecting user-facing paths. filename = str(filename) if row.get('file_name') or row.get('filename') or row.get('original_filename'): if Path(filename).name != filename or '\\' in filename: raise ValueError('Unsafe filename: ' + filename) else: filename = Path(filename).name if filename in ['', '.', '..']: raise ValueError('Unsafe filename.') image_path = output / 'images' / split / filename label_path = output / 'labels' / split / (Path(filename).stem + '.txt') if image_path.exists() or label_path.exists(): raise FileExistsError('Output filename already exists: ' + filename) digest = hashlib.sha256(encoded).hexdigest() expected = row.get('sha256') or row.get('image_sha256') if expected and digest != expected: raise ValueError('Image SHA256 mismatch: ' + filename) boxes, categories = annotation_lists(row) lines = [] counts = collections.Counter() for box, category in zip(boxes, categories): category = int(category) if category not in NAMES or len(box) != 4: raise ValueError('Invalid category or YOLO bbox: ' + filename) cx, cy, width, height = map(float, box) values = [cx, cy, width, height] if not all(math.isfinite(value) for value in values) or width <= 0 or height <= 0: raise ValueError('Invalid normalized YOLO bbox: ' + filename) if min(cx-width/2, cy-height/2) < -1e-10 or max(cx+width/2, cy+height/2) > 1+1e-10: raise ValueError('YOLO bbox extends beyond image bounds: ' + filename) selected = 0 if single_class else category counts[selected] += 1 lines.append(str(selected) + ' ' + ' '.join(format(value, '.17g') for value in values)) image_path.parent.mkdir(parents=True, exist_ok=True) label_path.parent.mkdir(parents=True, exist_ok=True) # Write the exact encoded bytes, with no Image.open, decode, resize, or save. view = memoryview(encoded) with image_path.open('xb') as handle: for position in range(0, len(view), 8*1024*1024): handle.write(view[position:position+8*1024*1024]) label_path.write_text('\n'.join(lines) + ('\n' if lines else '')) return {'split': split, 'filename': filename, 'image_relative_path': str(image_path.relative_to(output)), 'label_relative_path': str(label_path.relative_to(output)), 'sha256': digest, 'sample_id': row.get('sample_id'), 'source_group_id': row.get('source_group_id'), 'source_dataset': row.get('source_dataset'), 'annotation_count': len(lines), 'class_counts': dict(counts)} def discover_parquet(root, splits): files = list(root.rglob('*.parquet')) result = {} for split in splits: aliases = ['validation', 'val'] if split == 'val' else [split] matched = sorted(str(path) for path in files if any(path.name.startswith(alias+'-') or alias in path.relative_to(root).parts[:-1] for alias in aliases)) if not matched: raise FileNotFoundError('No parquet shards found for split ' + split + ' under ' + str(root)) result['validation' if split == 'val' else split] = matched return result def main(): parser = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter) source = parser.add_mutually_exclusive_group(required=True) source.add_argument('--repo-id', help='Hugging Face dataset ID, OWNER/DATASET.') source.add_argument('--local-dir', type=Path, help='Downloaded repository containing train/val/test Parquet shards.') parser.add_argument('--config', default='default', help='Hub dataset configuration; default is the full-resolution default config. The preview config is not a training source.') parser.add_argument('--revision', help='Optional Hub commit SHA or revision.') parser.add_argument('--output', type=Path, required=True) parser.add_argument('--single-class', action='store_true', help='Map original size class IDs 0/1/2 to one plane=0 class.') parser.add_argument('--splits', nargs='+', choices=['train', 'val', 'test'], default=['train', 'val', 'test']) parser.add_argument('--max-samples', type=int, help='Optional smoke-test limit per split; omitted exports all images.') args = parser.parse_args() if args.max_samples is not None and args.max_samples <= 0: parser.error('--max-samples must be positive.') if args.config == 'preview': parser.error('The preview configuration contains display images and cannot be exported as training data. Use --config default.') if args.output.exists() and any(args.output.iterdir()): parser.error('--output must be a new or empty directory; existing files are never overwritten.') from datasets import Image, load_dataset args.output.mkdir(parents=True, exist_ok=True) if args.local_dir: datasets = load_dataset('parquet', data_files=discover_parquet(args.local_dir, args.splits), streaming=True) else: kwargs = {'path': args.repo_id, 'streaming': True} if args.config: kwargs['name'] = args.config if args.revision: kwargs['revision'] = args.revision datasets = load_dataset(**kwargs) summary = {'mode': 'plane' if args.single_class else 'size3', 'splits': {}, 'image_bytes_reencoded': False} with (args.output/'export_manifest.jsonl').open('w') as manifest: for split in args.splits: source_split = 'validation' if split == 'val' else split if source_split not in datasets and split == 'val' and 'val' in datasets: source_split = 'val' data = datasets[source_split] if 'image' in (data.features or {}): data = data.cast_column('image', Image(decode=False)) images = 0 objects = 0 classes = collections.Counter() for row in data: result = export_one_row(row, split, args.output, args.single_class) manifest.write(json.dumps(result, ensure_ascii=False)+'\n') images += 1 objects += result['annotation_count'] classes.update(result['class_counts']) if images % 100 == 0: print(split+': '+str(images)+' images', flush=True) if args.max_samples is not None and images >= args.max_samples: break summary['splits'][split] = {'images': images, 'objects': objects, 'class_counts': dict(classes)} names = {0: 'plane'} if args.single_class else NAMES (args.output/'data.yaml').write_text('train: images/train\nval: images/val\ntest: images/test\n\nnames:\n'+''.join(' '+str(key)+': '+name+'\n' for key, name in names.items())) (args.output/'export_summary.json').write_text(json.dumps(summary, ensure_ascii=False, indent=2)+'\n') print(json.dumps(summary, ensure_ascii=False, indent=2)) if __name__ == '__main__': main()