#!/usr/bin/env python3 """Add validated SMPL-X sources to the EXISTING HFD tar/index namespace. Builds only new shards in an external staging directory; never repacks audio. Upload is a single parent-revision-guarded commit to a private repository. """ from __future__ import annotations import argparse from concurrent.futures import ThreadPoolExecutor import hashlib import json import os from pathlib import Path import shutil import sqlite3 import time import pandas as pd from huggingface_hub import HfApi, hf_hub_download from package_hf_audio_motion import digest, pack_shard, plan_shards, verify_package_plan, write_index REPO = 'Tinghong-Ye22/hfd_data' LOGICAL = 'motion/seed_all/smplx_amass' ROOT = Path(__file__).resolve().parents[2] def save(path, value): path.parent.mkdir(parents=True, exist_ok=True) tmp = path.with_suffix(path.suffix + '.partial') tmp.write_text(json.dumps(value, ensure_ascii=False, indent=2) + '\n') os.replace(tmp, path) def pair_sources(source, base, metadata, tts_path): """Join by identity, NEVER by CSV/parquet row order or caption alone.""" table = pd.read_parquet(source / 'manifest.parquet') seed = pd.read_csv(metadata, keep_default_na=False) if table.filename.duplicated().any() or seed.filename.duplicated().any(): raise ValueError('Non-unique motion identities') if set(table.filename) != set(seed.filename): raise ValueError('SMPL-X / G1 metadata identities differ') tts = {} for line in tts_path.read_text().splitlines(): row = json.loads(line) name = row['filename'] if name in tts: raise ValueError(f'Duplicate TTS identity: {name}') tts[name] = row with sqlite3.connect(f'file:{base / "index.sqlite"}?mode=ro', uri=True) as conn: assets = {r[0] for r in conn.execute('SELECT path FROM assets')} descriptions = seed.set_index('filename') rows = [] for r in table.to_dict('records'): name = r['filename'] original = descriptions.loc[name] if original.move_g1_path != r['move_g1_path']: raise ValueError(f'G1 path mismatch: {name}') relative = r['move_g1_path'].replace('g1/csv/', 'g1/csv_sonic/').removesuffix('.csv') folder = 'motion/seed_all/' + relative for field in ('joint_pos.csv', 'body_pos.csv', 'body_quat.csv'): if folder + '/' + field not in assets: raise ValueError(f'G1 not in existing HF index: {folder}/{field}') voice = tts.get(name) audio_dir = ('audio/seed_tts/seed_audio/' + Path(voice['audio_path']).name) if voice else None # Build the audio folder map once below (avoid 142k full-index scans). rows.append(dict(motion_id=name, g1_folder=folder, smplx_path=LOGICAL + '/' + r['output_path'], smplx_sha256=r['output_sha256'], source_bvh_sha256=r['source_bvh_sha256'], smplx_frames=int(r['output_frames']), smplx_fps=float(r['fps']), source_frames=int(r['source_frames']), source_frame_time=float(r['source_frame_time']), instruction_en=voice.get('content_short_description') if voice else None, instruction_zh=voice.get('content_short_description_zh') if voice else None, audio_dir=audio_dir)) audio_folders = {} for path in assets: if path.startswith('audio/seed_tts/seed_audio/') and path.endswith('.wav'): folder, voice = path.rsplit('/', 1) audio_folders.setdefault(folder, []).append(voice.removesuffix('.wav')) for row in rows: row['voices'] = sorted(audio_folders.get(row['audio_dir'], [])) row['has_instruction_audio'] = bool(row['voices']) return pd.DataFrame(rows) def main(): p = argparse.ArgumentParser(description=__doc__) p.add_argument('--source', required=True, type=Path) p.add_argument('--metadata', required=True, type=Path) p.add_argument('--tts-metadata', required=True, type=Path) p.add_argument('--output', required=True, type=Path) p.add_argument('--workers', type=int, default=16) p.add_argument('--upload', action='store_true') a = p.parse_args() api = HfApi() info = api.repo_info(REPO, repo_type='dataset') if not info.private: raise RuntimeError('Refusing gated SMPL-X delivery to a public repository') receipt = json.loads((a.source / 'validation_summary.json').read_text()) if receipt['status'] != 'validated' or receipt['success'] != receipt['expected'] or receipt['failures']: raise RuntimeError('Conversion is not completely validated') out = a.output.resolve() if out == a.source.resolve() or a.source.resolve() in out.parents or ROOT == out or ROOT in out.parents: raise ValueError('Use a separate project-external staging directory') if out.exists() and not (out / '.build/human_export.json').exists(): raise FileExistsError(out) state_path = out / '.build/human_export.json' if state_path.exists(): state = json.loads(state_path.read_text()) if state['parent_revision'] != info.sha: raise RuntimeError('HF changed since export began; review before rebuilding index') else: state = dict(parent_revision=info.sha, source=str(a.source.resolve()), stage='prepare') save(state_path, state) base = out / '.build/base' base.mkdir(parents=True, exist_ok=True) for name in ('index.sqlite', 'MANIFEST.json', 'PATH_MAP.json', 'SHA256SUMS', 'README.md'): src = hf_hub_download(REPO, name, repo_type='dataset', revision=state['parent_revision']) shutil.copyfile(src, base / name) paired = pair_sources(a.source, base, a.metadata, a.tts_metadata) if len(paired) != receipt['expected']: raise ValueError('Pair coverage differs from validated conversion') meta = out / 'metadata' meta.mkdir(exist_ok=True) paired.to_parquet(meta / 'seed_g1_smplx_pairs.parquet', index=False) mapping_summary = dict(motions=len(paired), with_audio=int(paired.has_instruction_audio.sum()), eight_voice_motions=int(paired.voices.map(len).eq(8).sum()), audio_directories=int(paired.audio_dir.nunique()), conversion=receipt, parent_revision=state['parent_revision']) save(meta / 'seed_smplx_summary.json', mapping_summary) sources = [(a.source / 'motions', LOGICAL + '/motions')] for name in ('manifest.parquet', 'validation_summary.json', 'README.md', 'conversion_receipt.json', 'identity_receipt.json', 'source_archive_receipt.json'): sources.append((a.source / name, LOGICAL + '/' + name)) buckets, raw_bytes = plan_shards('seed_smplx', sources, 128) verify_package_plan(out, 'seed_smplx', buckets) state.update(stage='packing', raw_bytes=raw_bytes, shards_done=0) save(state_path, state) with ThreadPoolExecutor(max_workers=a.workers) as pool: futures = [pool.submit(pack_shard, out, 'seed_smplx', i, b) for i, b in enumerate(buckets)] records = [] from concurrent.futures import as_completed for f in as_completed(futures): records.append(f.result()) state['shards_done'] = len(records) save(state_path, state) records.sort(key=lambda r: r['file']) write_index(out, records) # The shared root index retains EVERY old asset, while adding the new namespace. with sqlite3.connect(out / 'index.sqlite') as conn: conn.execute('ATTACH DATABASE ? AS old', (str(base / 'index.sqlite'),)) conn.execute('INSERT INTO assets SELECT * FROM old.assets') conn.commit() count = conn.execute('SELECT count(*) FROM assets').fetchone()[0] paths = json.loads((base / 'PATH_MAP.json').read_text()) paths['prefixes'][str(a.source.resolve())] = LOGICAL save(out / 'PATH_MAP.json', paths) manifest = json.loads((base / 'MANIFEST.json').read_text()) if not isinstance(manifest, list): raise ValueError('Unexpected existing MANIFEST schema; do not overwrite it') if any(r['file'].startswith('shards/seed_smplx/') for r in manifest): raise ValueError('SMPL-X package already exists; explicit version review required') manifest.extend(records) save(out / 'MANIFEST.json', manifest) shutil.copyfile(ROOT / 'scripts/data/read_hf_audio_motion.py', out / 'read_hf_audio_motion.py') shutil.copyfile(ROOT / 'doc/hf_audio_motion/SMPLX_HANDOFF.md', out / 'SMPLX_HANDOFF.md') readme = (base / 'README.md').read_text() readme += '\n\n## Human motion 增量包(2026-10-13)\n\n' + ( '新增 `shards/seed_smplx/`(128 个 tar),与原 G1/音频共用根目录索引;无需重复下载语音。' '142,220 条 AMASS SMPL-X 原始动作,约 19 GiB。对应表:' '`metadata/seed_g1_smplx_pairs.parquet`。详细格式、限制和 session 接入见 ' '[SMPLX_HANDOFF.md](SMPLX_HANDOFF.md)。**源动作交付不等于已验证的完整人体 session 数据集。**\n') (out / 'README.md').write_text(readme) # Exclude overwritten metadata checksums, retain untouched old assets verbatim. replaced = {'index.sqlite','PATH_MAP.json','MANIFEST.json','README.md','read_hf_audio_motion.py'} old_lines = [line for line in (base / 'SHA256SUMS').read_text().splitlines() if line.strip() and line.split(maxsplit=1)[1].lstrip('*').removeprefix('./') not in replaced] new_files = sorted(f for f in out.rglob('*') if f.is_file() and '.build' not in f.relative_to(out).parts and f.name != 'SHA256SUMS') checks = [f'{digest(f)} {f.relative_to(out).as_posix()}' for f in new_files] (out / 'SHA256SUMS').write_text('\n'.join(old_lines + checks) + '\n') # Read each new payload THROUGH the merged index and verify its conversion hash. from read_hf_audio_motion import ShardStore store = ShardStore(out) try: for row in paired.itertuples(): if hashlib.sha256(store.read(row.smplx_path)).hexdigest() != row.smplx_sha256: raise ValueError(f'Indexed payload checksum failed: {row.motion_id}') finally: store.close() state.update(stage='verified', assets=count, motions=len(paired), updated_at=time.time()) save(state_path, state) print(json.dumps(state), flush=True) if a.upload: if not api.repo_info(REPO, repo_type='dataset').private: raise RuntimeError('Repository is no longer private') state['stage'] = 'uploading' save(state_path, state) try: result = api.upload_folder(repo_id=REPO, repo_type='dataset', folder_path=out, ignore_patterns=['.build/**'], parent_commit=state['parent_revision'], commit_message='Add validated paired BONES-SEED SMPL-X source shards') except Exception as exc: state.update(stage='upload_failed', error_type=type(exc).__name__, private_storage_limit=('Private storage limit reached' in str(exc)), updated_at=time.time()) save(state_path, state) raise state.update(stage='uploaded', commit=result.oid, url=result.commit_url, updated_at=time.time()) save(state_path, state) print(json.dumps(state), flush=True) if __name__ == '__main__': main()