Upload scripts/build_clean_dataset.py

scripts/build_clean_dataset.py

@@ -0,0 +1,164 @@
+
+import csv
+import errno
+import os
+import shutil
+import time
+from pathlib import Path
+from collections import Counter
+
+SRC_ROOT = Path('/aifs4su/yezhen/data')
+SRC_CSV = SRC_ROOT / 'unified_all_dataset_more_filtered.csv'
+OUT_ROOT = SRC_ROOT / 'Talker-T2AV-Data-clean'
+META_DIR = OUT_ROOT / 'metadata'
+OUT_CSV = META_DIR / 'train.csv'
+REPORT = OUT_ROOT / 'build_report.txt'
+
+PREFIXES = {
+    'dh_motion': 'apdcephfs_gy2/share_302533218/zhenye/data/DH-FaceVid-1K-lia-x-features/',
+    'dh_audio': 'apdcephfs_gy2/share_302533218/zhenye/data/DH-FaceVid-1K-processed_separated/',
+    'dh_video_dir': 'DH-FaceVid-1K-processed',
+    'ditto': 'apdcephfs_gy2/share_302533218/zhenye/ditto_data_v214/',
+    'seamless_audio': 'apdcephfs_gy2/share_302533218/zhenye/data/seamless_emilia_v2/',
+    'hallo3_audio': 'apdcephfs_gy2/share_302533218/zhenye/data/hallo3_training_data/hallo3/audios_separated/',
+    'ravdess_audio': 'apdcephfs_gy2/share_302533218/zhenye/data/ravadess/',
+    'mead_audio': 'apdcephfs_gy2/share_302533218/zhenye/data/mead/',
+    'talkvid_audio': 'apdcephfs_gy4/share_302533218/zhenye/talkvid_data/',
+}
+
+def require_suffix(path: str, prefix: str) -> str:
+    if not path.startswith(prefix):
+        raise ValueError(f'path does not start with expected prefix: {path} prefix={prefix}')
+    return path[len(prefix):]
+
+def map_motion(src: str, dataset: str) -> str:
+    if dataset == 'dh_facevid':
+        suffix = require_suffix(src, PREFIXES['dh_motion'])
+    else:
+        suffix = require_suffix(src, PREFIXES['ditto'] + dataset + '/')
+    return f'motion/{dataset}/{suffix}'
+
+def map_audio(src: str, dataset: str) -> str:
+    if dataset == 'dh_facevid':
+        suffix = require_suffix(src, PREFIXES['dh_audio'])
+    elif dataset == 'seamless_emilia':
+        suffix = require_suffix(src, PREFIXES['seamless_audio'])
+    elif dataset == 'hallo3':
+        suffix = require_suffix(src, PREFIXES['hallo3_audio'])
+    elif dataset == 'ravdess':
+        suffix = require_suffix(src, PREFIXES['ravdess_audio'])
+    elif dataset == 'mead':
+        suffix = require_suffix(src, PREFIXES['mead_audio'])
+    elif dataset == 'talkvid':
+        suffix = require_suffix(src, PREFIXES['talkvid_audio'])
+    else:
+        raise ValueError(f'unknown dataset_source for audio: {dataset}')
+    return f'audio/{dataset}/{suffix}'
+
+def map_video_source_and_clean(motion_src: str, dataset: str):
+    if dataset == 'dh_facevid':
+        stem = Path(motion_src).stem
+        video_src = f"{PREFIXES['dh_video_dir']}/{stem}.mp4"
+        video_clean = f'video/{dataset}/{stem}.mp4'
+    else:
+        video_src = str(Path(motion_src).with_suffix('.mp4'))
+        motion_clean = map_motion(motion_src, dataset).replace('/seamless_emilia/', '/seamless/')
+        video_clean = str(Path(motion_clean.replace('motion/', 'video/', 1)).with_suffix('.mp4'))
+    return video_src, video_clean
+
+def link_file(src_rel: str, dst_rel: str, missing_counter: Counter, linked_counter: Counter, kind: str):
+    src = SRC_ROOT / src_rel
+    dst = OUT_ROOT / dst_rel
+    if not src.exists():
+        missing_counter[kind] += 1
+        return False
+    dst.parent.mkdir(parents=True, exist_ok=True)
+    if dst.exists():
+        linked_counter[f'{kind}_exists'] += 1
+        return True
+    try:
+        os.link(src, dst)
+    except OSError as e:
+        if e.errno == errno.EEXIST:
+            linked_counter[f'{kind}_exists'] += 1
+            return True
+        if e.errno == errno.EXDEV:
+            os.symlink(src, dst)
+        else:
+            raise
+    linked_counter[kind] += 1
+    return True
+
+def main():
+    start = time.time()
+    META_DIR.mkdir(parents=True, exist_ok=True)
+    tmp_csv = OUT_CSV.with_suffix('.csv.tmp')
+    fields = [
+        'sample_id', 'dataset_source', 'split', 'text',
+        'wav_path', 'motion_pt_path', 'video_path',
+        'pt_length', 'fps',
+    ]
+    counts = Counter()
+    missing = Counter()
+    linked = Counter()
+    errors = []
+
+    with SRC_CSV.open('r', newline='', encoding='utf-8') as fin, tmp_csv.open('w', newline='', encoding='utf-8') as fout:
+        reader = csv.DictReader(fin)
+        writer = csv.DictWriter(fout, fieldnames=fields)
+        writer.writeheader()
+        for idx, row in enumerate(reader):
+            dataset = row['dataset_source']
+            clean_dataset = 'seamless' if dataset == 'seamless_emilia' else dataset
+            counts[clean_dataset] += 1
+            try:
+                motion_src = row['motion_pt_path']
+                wav_src = row['wav_path']
+                motion_clean = map_motion(motion_src, dataset).replace('/seamless_emilia/', '/seamless/')
+                wav_clean = map_audio(wav_src, dataset).replace('/seamless_emilia/', '/seamless/')
+                video_src, video_clean = map_video_source_and_clean(motion_src, dataset)
+
+                ok_motion = link_file(motion_src, motion_clean, missing, linked, 'motion')
+                ok_wav = link_file(wav_src, wav_clean, missing, linked, 'audio')
+                ok_video = link_file(video_src, video_clean, missing, linked, 'video')
+
+                if not (ok_motion and ok_wav):
+                    errors.append((idx, dataset, motion_src, wav_src, video_src))
+
+                writer.writerow({
+                    'sample_id': f'{clean_dataset}_{idx:09d}',
+                    'dataset_source': clean_dataset,
+                    'split': 'train',
+                    'text': row.get('text', ''),
+                    'wav_path': wav_clean,
+                    'motion_pt_path': motion_clean,
+                    'video_path': video_clean if ok_video else '',
+                    'pt_length': row.get('pt_length', ''),
+                    'fps': '25',
+                })
+            except Exception as e:
+                errors.append((idx, dataset, str(e)))
+                missing['row_errors'] += 1
+
+            if (idx + 1) % 10000 == 0:
+                elapsed = time.time() - start
+                print(f'processed={idx+1} elapsed={elapsed:.1f}s missing={dict(missing)} linked={sum(v for k,v in linked.items() if not k.endswith("_exists"))}', flush=True)
+
+    os.replace(tmp_csv, OUT_CSV)
+    elapsed = time.time() - start
+    with REPORT.open('w', encoding='utf-8') as f:
+        f.write(f'source_csv={SRC_CSV}\n')
+        f.write(f'output_csv={OUT_CSV}\n')
+        f.write(f'elapsed_sec={elapsed:.1f}\n')
+        f.write(f'counts={dict(counts)}\n')
+        f.write(f'missing={dict(missing)}\n')
+        f.write(f'linked={dict(linked)}\n')
+        f.write(f'errors_count={len(errors)}\n')
+        for err in errors[:100]:
+            f.write(f'error={err}\n')
+    print(f'DONE elapsed={elapsed:.1f}s counts={dict(counts)} missing={dict(missing)} errors={len(errors)} report={REPORT}', flush=True)
+    if missing.get('motion', 0) or missing.get('audio', 0) or missing.get('row_errors', 0):
+        raise SystemExit(2)
+
+if __name__ == '__main__':
+    main()