TEMA-Data / tools /materialize.py
bluesky7's picture
Initial release
88d648e
Raw History Blame Contribute Delete
6.05 kB
#!/usr/bin/env python3
"""Resolve the annotation release against legally obtained local source waveforms.
No audio is downloaded. No source waveform is modified. All host-specific paths
are written only beneath --output-root, never into the annotation release.
"""
import argparse
import hashlib
import json
import os
from pathlib import Path
import shutil
def sha256(path):
h = hashlib.sha256()
with path.open('rb') as f:
for block in iter(lambda: f.read(1024 * 1024), b''):
h.update(block)
return h.hexdigest()
def records(path):
with path.open(encoding='utf-8') as f:
for line in f:
if line.strip():
yield json.loads(line)
def main():
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument('--release-root', type=Path, default=Path(__file__).resolve().parents[1])
parser.add_argument('--source-root', action='append', default=[], metavar='DATASET=PATH',
help='Repeat for audioset_strong, audiotime, tacos; root precedes catalog source_relative_path.')
parser.add_argument('--output-root', type=Path, required=True)
parser.add_argument('--include', nargs='+', choices=['temporal_init', 'sft', 'rl', 'benchmark'], default=['sft', 'rl', 'benchmark'])
parser.add_argument('--mode', choices=['symlink', 'copy'], default='symlink')
parser.add_argument('--verify-hashes', action='store_true', help='Check all available saved file SHA256 values; no decoding.')
parser.add_argument('--dry-run', action='store_true', help='Resolve/check source files without writing output.')
args = parser.parse_args()
release = args.release_root.resolve()
output = args.output_root.resolve()
if output == release or release in output.parents:
parser.error('--output-root must be outside the released repository')
roots = {}
for specification in args.source_root:
name, sep, value = specification.partition('=')
if not sep or name not in ('audioset_strong', 'audiotime', 'tacos'):
parser.error('source roots must use audioset_strong=..., audiotime=..., tacos=...')
roots[name] = Path(value).expanduser().resolve()
selected = set(args.include)
required = []
missing = []
bad = []
verified = 0
for r in records(release / 'audio/catalog.jsonl'):
if not selected.intersection(r['used_by']):
continue
root = roots.get(r['source_dataset'])
source = root / r['source_relative_path'] if root else None
if source is None or not source.is_file():
missing.append({'audio_id': r['audio_id'], 'dataset': r['source_dataset'],
'source_relative_path': r['source_relative_path']})
continue
if r.get('size_bytes') is not None and source.stat().st_size != r['size_bytes']:
bad.append({'audio_id': r['audio_id'], 'error': 'size mismatch'})
if args.verify_hashes and r.get('file_sha256'):
verified += 1
if sha256(source) != r['file_sha256']:
bad.append({'audio_id': r['audio_id'], 'error': 'file SHA256 mismatch'})
required.append((r, source))
report = {'selected': sorted(selected), 'found_audio': len(required),
'missing_audio': len(missing), 'invalid_audio': len(bad),
'verified_file_hashes': verified, 'missing_examples': missing[:20], 'invalid_examples': bad[:20]}
print(json.dumps(report, indent=2, ensure_ascii=False))
if missing or bad:
raise SystemExit('Audio validation failed; no dataset was materialized. Obtain the exact source exports listed in the catalog.')
if args.dry_run:
return
output.mkdir(parents=True, exist_ok=True)
for r, source in required:
destination = output / r['path']
destination.parent.mkdir(parents=True, exist_ok=True)
if destination.is_symlink():
if destination.resolve() == source.resolve():
continue
raise SystemExit(f'Refusing to overwrite mismatched symlink: {destination}')
if destination.exists():
if args.mode == 'copy' and destination.stat().st_size == source.stat().st_size and sha256(destination) == sha256(source):
continue
raise SystemExit(f'Refusing to overwrite existing file: {destination}')
if args.mode == 'symlink':
destination.symlink_to(source)
else:
shutil.copy2(source, destination)
def rewrite(value):
if isinstance(value, list):
return [rewrite(x) for x in value]
if isinstance(value, dict):
return {k: rewrite(v) for k, v in value.items()}
if isinstance(value, str) and value.startswith(('audio/audioset_strong/', 'audio/audiotime/', 'audio/tacos/')):
return str(output / value)
return value
for split in sorted(selected):
for p in (release / split).glob('*'):
if not p.is_file():
continue
destination = output / split / p.name
destination.parent.mkdir(parents=True, exist_ok=True)
if p.suffix == '.jsonl':
with destination.open('w', encoding='utf-8') as f:
for row in records(p):
f.write(json.dumps(rewrite(row), ensure_ascii=False, separators=(',', ':')) + '\n')
elif p.suffix == '.json':
destination.write_text(json.dumps(rewrite(json.loads(p.read_text())), ensure_ascii=False, indent=2) + '\n')
else:
shutil.copy2(p, destination)
(output / 'audio').mkdir(exist_ok=True)
shutil.copy2(release / 'audio/catalog.jsonl', output / 'audio/catalog.jsonl')
for name in ['task_families.json']:
shutil.copy2(release / name, output / name)
(output / 'materialization_audit.json').write_text(json.dumps(report, indent=2) + '\n')
print(f'Materialized datasets: {output}')
if __name__ == '__main__':
main()