code-HFD_lzk / scripts /data /materialize_hfd_ami_mirror.py
Tinghong-Ye22's picture
Add files using upload-large-folder tool
8562e9b verified
Raw History Blame Contribute Delete
10 kB
#!/usr/bin/env python3
"""Verified IHM mirror fallback; keep only fully contained annotated sentences."""
from __future__ import annotations
import argparse
from collections import Counter, defaultdict
from concurrent.futures import ThreadPoolExecutor
import hashlib
import io
import json
from pathlib import Path
import struct
import subprocess
import sys
import wave
import xml.etree.ElementTree as ET
import zipfile
import numpy as np
import pyarrow.parquet as pq
ROOT = Path(__file__).resolve().parents[2]
if str(ROOT) not in sys.path:
sys.path.insert(0, str(ROOT))
from scripts.data.hfd_third_party_common import read_jsonl, sha256_file, write_json_atomic, write_jsonl_atomic
REVISION = '46f28f2503e2ec48f8867a84eef356c70476beab'
BASE = 'https://huggingface.co/datasets/edinburghcstr/ami'
def decode_mirror_wav(payload):
"""The mirror stores original PCM16 values in IEEE float WAV containers."""
if payload[:4] != b'RIFF' or payload[8:12] != b'WAVE':
raise ValueError('Mirror audio is not a WAV')
pos = 12
fmt = pcm = None
while pos + 8 <= len(payload):
kind, size = struct.unpack_from('<4sI', payload, pos)
body = payload[pos + 8:pos + 8 + size]
if len(body) != size:
raise ValueError('Truncated mirror WAV')
if kind == b'fmt ':
fmt = struct.unpack_from('<HHIIHH', body)
if kind == b'data':
pcm = body
pos += 8 + size + size % 2
if fmt == (3, 1, 16000, 64000, 4, 32):
floats = np.frombuffer(pcm, dtype='<f4')
values = floats.astype(np.float64) * 32768
if not np.isfinite(values).all() or np.max(np.abs(values - np.rint(values)), initial=0) > 1e-4:
raise ValueError('Mirror audio is not losslessly representable as PCM16')
if np.min(values, initial=0) < -32768 or np.max(values, initial=0) > 32767:
raise ValueError('Mirror PCM16 range exceeded')
return np.rint(values).astype('<i2')
if fmt == (1, 1, 16000, 32000, 2, 16):
return np.frombuffer(pcm, dtype='<i2')
raise ValueError(f'Unexpected mirror WAV format: {fmt}')
def mirror_bounds(row):
# Avoid float32 timestamp rounding: IDs encode centiseconds exactly.
first, last = row['audio_id'].rsplit('_', 2)[-2:]
start, end = int(first) * 160, int(last) * 160
if abs(row['begin_time'] * 16000 - start) > 16 or abs(row['end_time'] * 16000 - end) > 16:
raise ValueError('Mirror timestamps disagree with segment ID')
return start, end
def annotated_spoken_bounds(planned, archive):
"""Drop optional context padding, never any annotated spoken sample."""
cache = {}
with zipfile.ZipFile(archive) as source:
for row in planned:
member = row['words_member']
if member not in cache:
words = list(ET.fromstring(source.read(member)))
indices = {w.get('{http://nite.sourceforge.net/}id'): i for i, w in enumerate(words)}
cache[member] = words, indices
words, indices = cache[member]
selected = words[indices[row['word_id_first']]:indices[row['word_id_last']] + 1]
spoken = [w for w in selected if w.tag == 'w' and w.get('punc') != 'true']
first = min(float(w.get('starttime')) for w in spoken)
last = max(float(w.get('endtime')) for w in spoken)
if first < row['start_seconds'] - .0001 or last > row['end_seconds'] + .0001:
raise ValueError('Spoken words lie outside planned sentence')
row['planned_context_start_seconds'] = row['start_seconds']
row['planned_context_end_seconds'] = row['end_seconds']
row['start_seconds'], row['end_seconds'] = first, last
row['boundary_policy'] = 'all_annotated_words_no_optional_silence_margin'
return planned
def download(raw, workers):
revision = json.loads((raw / 'ami_hf_revision.json').read_text())['sha']
if revision != REVISION:
raise ValueError('Unexpected mirror revision; review before changing the pin')
files = json.loads((raw / 'ami_hf_tree.json').read_text())
directory = raw / 'ami_hf'
directory.mkdir(exist_ok=True)
def fetch(row):
path = directory / Path(row['path']).name
expected = row['lfs']['oid']
if not path.exists() or path.stat().st_size != row['size'] or sha256_file(path) != expected:
subprocess.run(['curl', '-fsSL', '--retry', '3', '--retry-all-errors',
'--connect-timeout', '20', '--max-time', '900', '-o', str(path),
f"{BASE}/resolve/{REVISION}/{row['path']}?download=true"], check=True)
if path.stat().st_size != row['size'] or sha256_file(path) != expected:
raise ValueError(f'Mirror LFS hash mismatch: {path.name}')
print(f'verified {path.name}', flush=True)
with ThreadPoolExecutor(max_workers=workers) as pool:
list(pool.map(fetch, [r for r in files if r['path'].endswith('.parquet')]))
def materialize(raw):
planned = annotated_spoken_bounds(read_jsonl(raw / 'ami_plan.jsonl'), raw / 'ami_public_manual_1.6.2.zip')
groups = defaultdict(list)
for row in planned:
groups[row['meeting'], row['raw_speaker'], f"H{int(row['channel']):02d}"].append(row)
accepted = {}
rejected = {}
directory = raw / 'ami_mirror_clips'
directory.mkdir(exist_ok=True)
for info in json.loads((raw / 'ami_hf_tree.json').read_text()):
if not info['path'].endswith('.parquet'):
continue
path = raw / 'ami_hf' / Path(info['path']).name
digest = sha256_file(path)
if digest != info['lfs']['oid']:
raise ValueError(f'Mirror changed: {path}')
file = pq.ParquetFile(path)
row_index = 0
for batch in file.iter_batches(batch_size=128):
for sample in batch.to_pylist():
key = sample['meeting_id'], sample['speaker_id'], sample['microphone_id']
candidates = groups.get(key, [])
begin, end = mirror_bounds(sample)
fits = [r for r in candidates if r['asset_id'] not in accepted
and begin <= round(r['start_seconds'] * 16000)
and round(r['end_seconds'] * 16000) <= end]
if fits:
try:
audio = decode_mirror_wav(sample['audio']['bytes'])
except ValueError as exc:
for row in fits:
rejected[row['asset_id']] = 'unsupported_source_audio: ' + str(exc)
row_index += 1
continue
if abs(len(audio) - (end - begin)) > 1:
raise ValueError('Mirror audio length disagrees with time range')
for row in fits:
first, last = round(row['start_seconds'] * 16000), round(row['end_seconds'] * 16000)
if last - begin > len(audio):
continue
clip = directory / (row['asset_id'].replace(':', '_') + '.wav')
pcm = audio[first - begin:last - begin].tobytes()
with wave.open(str(clip), 'wb') as wav:
wav.setnchannels(1); wav.setsampwidth(2); wav.setframerate(16000); wav.writeframes(pcm)
record = {**row, 'raw_path': str(clip.resolve()),
'raw_sha256': sha256_file(clip),
'raw_storage_policy': 'lossless_contained_sentence_from_verified_ihm_mirror',
'mirror_url': f"{BASE}/resolve/{REVISION}/{info['path']}",
'mirror_revision': REVISION, 'mirror_path': str(path.resolve()),
'mirror_sha256': digest, 'mirror_row_index': row_index,
'mirror_audio_id': sample['audio_id'], 'mirror_text': sample['text'],
'mirror_begin_sample_16k': begin, 'mirror_end_sample_16k': end,
'original_start_sample_16k': first, 'original_end_sample_16k': last,
'downloaded_pcm_sha256': hashlib.sha256(pcm).hexdigest()}
write_json_atomic(clip.with_suffix('.download.json'), record)
accepted[row['asset_id']] = record
row_index += 1
print(f'{path.name}: {len(accepted)} contained sentences so far', flush=True)
records = [accepted[key] for key in sorted(accepted)]
write_jsonl_atomic(raw / 'ami_materialized.jsonl', records)
write_jsonl_atomic(raw / 'ami_mirror_unavailable.jsonl', [
{'asset_id': r['asset_id'], 'reason': rejected.get(r['asset_id'], 'no_single_mirror_segment_contains_full_annotated_sentence')}
for r in planned if r['asset_id'] not in accepted])
result = {'planned_clips': len(planned), 'materialized_clips': len(records),
'speakers': len({r['speaker_id'] for r in records}),
'splits': dict(Counter(r['split'] for r in records)),
'hours': sum(r['end_seconds'] - r['start_seconds'] for r in records) / 3600,
'mirror_revision': REVISION, 'no_incomplete_sentence_or_silence_padding': True}
write_json_atomic(raw / 'ami_mirror_materialization.json', result)
print(json.dumps(result, indent=2), flush=True)
if __name__ == '__main__':
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument('--raw-root', required=True, type=Path)
parser.add_argument('--mode', choices=('download', 'materialize'), required=True)
parser.add_argument('--workers', type=int, default=6)
args = parser.parse_args()
if args.mode == 'download':
download(args.raw_root.resolve(), args.workers)
else:
materialize(args.raw_root.resolve())