code-HFD / scripts /data /patch_2264_source_motion_dir.py
Tinghong-Ye22's picture
Add files using upload-large-folder tool
9fc9873 verified
Raw History Blame Contribute Delete
13.9 kB
#!/usr/bin/env python3
from __future__ import annotations
import argparse
import json
import multiprocessing as mp
import os
import re
import shutil
import sys
import time
import traceback
from datetime import datetime
from pathlib import Path
from typing import Any
NAME_RE = re.compile(rb'"name"\s*:\s*"([^"]+)"')
SOURCE_MOTION_RE = re.compile(rb'"source_motion_(?:dir|path)"\s*:')
SOURCE_VIDEO_RE = re.compile(rb'"source_video_path"\s*:')
def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser(
description=(
"Patch 2264_with_audio video_clip metadata.jsonl with source_motion_dir "
"so it matches existing metadata_only video_clip datasets."
)
)
parser.add_argument(
"--clip-root",
type=Path,
default=Path("/cpfs01/thscc/sharestorage/AIGC-06-01/Hithink_VideoGen_Group/user_workspace/yetinghong/videogen/psi0_video_clip_50f_dense_stride1_direct_full/2264_with_audio"),
)
parser.add_argument(
"--source-root",
type=Path,
default=Path("/cpfs01/thscc/sharestorage/AIGC-06-01/Hithink_VideoGen_Group/user_workspace/buwendong/videogen/2264_with_audio"),
)
parser.add_argument("--metadata", default="metadata.jsonl")
parser.add_argument("--source-metadata", default="metadata_convert.jsonl")
parser.add_argument("--workers", type=int, default=min(16, os.cpu_count() or 1))
parser.add_argument("--progress-bytes", type=int, default=64 * 1024 * 1024)
parser.add_argument("--apply", action="store_true", help="Atomically replace metadata.jsonl after successful patch.")
parser.add_argument("--keep-temp", action="store_true")
parser.add_argument("--no-space-check", action="store_true")
return parser.parse_args()
def load_source_index(source_root: Path, source_metadata: Path) -> dict[str, tuple[bytes, bytes]]:
if not source_metadata.exists():
raise FileNotFoundError(f"source metadata not found: {source_metadata}")
index: dict[str, tuple[bytes, bytes]] = {}
missing: list[str] = []
with source_metadata.open("r", encoding="utf-8") as f:
for line_no, line in enumerate(f, start=1):
line = line.strip()
if not line:
continue
row = json.loads(line)
name = str(row.get("name", "")).strip()
motion_rel = row.get("converted_path")
video_rel = row.get("video_path")
if not name or not motion_rel or not video_rel:
missing.append(f"{line_no}: missing name/converted_path/video_path")
continue
motion_path = source_root / str(motion_rel)
video_path = source_root / str(video_rel)
if not (motion_path / "joint_pos.csv").exists():
missing.append(f"{line_no}: missing motion csv dir {motion_path}")
continue
if not video_path.exists():
missing.append(f"{line_no}: missing source video {video_path}")
continue
motion_json = json.dumps(str(motion_path), ensure_ascii=False, separators=(",", ":")).encode("utf-8")
video_json = json.dumps(str(video_path), ensure_ascii=False, separators=(",", ":")).encode("utf-8")
index[name] = (motion_json, video_json)
if missing:
preview = "\n".join(missing[:20])
raise RuntimeError(f"source metadata validation failed ({len(missing)} issue(s)); first issues:\n{preview}")
if not index:
raise RuntimeError(f"empty source index from {source_metadata}")
return index
def compute_chunks(path: Path, workers: int) -> list[tuple[int, int, int]]:
size = path.stat().st_size
workers = max(1, min(int(workers), size or 1))
starts = [0]
with path.open("rb") as f:
for i in range(1, workers):
f.seek(size * i // workers)
f.readline()
pos = f.tell()
if pos < size and pos != starts[-1]:
starts.append(pos)
starts.append(size)
return [(idx, starts[idx], starts[idx + 1]) for idx in range(len(starts) - 1) if starts[idx] < starts[idx + 1]]
def patch_line(line: bytes, source_index: dict[str, tuple[bytes, bytes]]) -> tuple[bytes, str | None]:
match = NAME_RE.search(line)
if match is None:
return line, "missing name"
clip_name = match.group(1).decode("utf-8")
source_name = clip_name.split(".chunk", 1)[0]
source = source_index.get(source_name)
if source is None:
return line, f"missing source mapping for {source_name}"
additions: list[bytes] = []
if SOURCE_MOTION_RE.search(line) is None:
additions.append(b'"source_motion_dir":' + source[0])
if SOURCE_VIDEO_RE.search(line) is None:
additions.append(b'"source_video_path":' + source[1])
if not additions:
return line, None
newline = b"\n" if line.endswith(b"\n") else b""
body = line.rstrip(b"\r\n")
if not body.endswith(b"}"):
return line, "line is not a JSON object"
return body[:-1] + b"," + b",".join(additions) + b"}" + newline, None
def worker_main(
idx: int,
start: int,
end: int,
input_path: str,
output_path: str,
source_index: dict[str, tuple[bytes, bytes]],
queue: mp.Queue,
progress_bytes: int,
) -> None:
stats: dict[str, Any] = {
"idx": idx,
"rows": 0,
"patched": 0,
"already_had_source_motion": 0,
"already_had_source_video": 0,
"errors": 0,
"error_examples": [],
"bytes": 0,
}
try:
next_report = progress_bytes
with open(input_path, "rb") as fin, open(output_path, "wb") as fout:
fin.seek(start)
while fin.tell() < end:
line = fin.readline()
if not line:
break
stats["rows"] += 1
stats["bytes"] += len(line)
had_motion = SOURCE_MOTION_RE.search(line) is not None
had_video = SOURCE_VIDEO_RE.search(line) is not None
if had_motion:
stats["already_had_source_motion"] += 1
if had_video:
stats["already_had_source_video"] += 1
patched, error = patch_line(line, source_index)
if error is not None:
stats["errors"] += 1
if len(stats["error_examples"]) < 10:
stats["error_examples"].append(f"chunk={idx} row={stats['rows']}: {error}")
fout.write(line)
else:
if patched != line:
stats["patched"] += 1
fout.write(patched)
if stats["bytes"] >= next_report:
queue.put(("progress", idx, stats["bytes"]))
next_report = stats["bytes"] + progress_bytes
queue.put(("progress", idx, stats["bytes"]))
queue.put(("done", idx, stats))
except BaseException:
queue.put(("error", idx, traceback.format_exc()))
def progress_bar(total: int):
try:
from tqdm import tqdm
return tqdm(total=total, unit="B", unit_scale=True, desc="patch metadata")
except Exception:
return None
def run_workers(
metadata_path: Path,
tmp_dir: Path,
source_index: dict[str, tuple[bytes, bytes]],
workers: int,
progress_bytes: int,
) -> tuple[list[Path], dict[str, Any]]:
chunks = compute_chunks(metadata_path, workers)
queue: mp.Queue = mp.Queue()
processes: list[mp.Process] = []
part_paths: list[Path] = []
for idx, start, end in chunks:
part_path = tmp_dir / f"metadata.part-{idx:05d}.jsonl"
part_paths.append(part_path)
process = mp.Process(
target=worker_main,
args=(idx, start, end, str(metadata_path), str(part_path), source_index, queue, progress_bytes),
)
process.start()
processes.append(process)
pbar = progress_bar(metadata_path.stat().st_size)
last_progress = {idx: 0 for idx, _, _ in chunks}
totals: dict[str, Any] = {
"rows": 0,
"patched": 0,
"already_had_source_motion": 0,
"already_had_source_video": 0,
"errors": 0,
"error_examples": [],
}
done = 0
last_print = time.time()
try:
while done < len(processes):
kind, idx, payload = queue.get()
if kind == "progress":
delta = int(payload) - last_progress.get(idx, 0)
last_progress[idx] = int(payload)
if pbar is not None:
pbar.update(max(0, delta))
elif time.time() - last_print > 5:
seen = sum(last_progress.values())
print(f"[patch] {seen / (1024 ** 3):.2f}/{metadata_path.stat().st_size / (1024 ** 3):.2f} GB", flush=True)
last_print = time.time()
elif kind == "done":
done += 1
for key in ("rows", "patched", "already_had_source_motion", "already_had_source_video", "errors"):
totals[key] += int(payload[key])
totals["error_examples"].extend(payload["error_examples"])
elif kind == "error":
for process in processes:
if process.is_alive():
process.terminate()
raise RuntimeError(f"worker {idx} failed:\n{payload}")
finally:
if pbar is not None:
pbar.close()
for process in processes:
process.join()
return part_paths, totals
def merge_parts(part_paths: list[Path], output_path: Path) -> None:
try:
from tqdm import tqdm
pbar = tqdm(total=sum(path.stat().st_size for path in part_paths), unit="B", unit_scale=True, desc="merge")
except Exception:
pbar = None
with output_path.open("wb") as fout:
for part in part_paths:
with part.open("rb") as fin:
while True:
chunk = fin.read(64 * 1024 * 1024)
if not chunk:
break
fout.write(chunk)
if pbar is not None:
pbar.update(len(chunk))
if pbar is not None:
pbar.close()
def check_free_space(root: Path, metadata_path: Path) -> None:
usage = shutil.disk_usage(root)
required = int(metadata_path.stat().st_size * 1.35) + 8 * 1024**3
if usage.free < required:
raise RuntimeError(
f"not enough free space under {root}: free={usage.free / (1024 ** 3):.1f}GB, "
f"need about {required / (1024 ** 3):.1f}GB"
)
def main() -> None:
args = parse_args()
clip_root: Path = args.clip_root
source_root: Path = args.source_root
metadata_path = clip_root / args.metadata
source_metadata = source_root / args.source_metadata
if not metadata_path.exists():
raise FileNotFoundError(f"clip metadata not found: {metadata_path}")
if not args.no_space_check:
check_free_space(clip_root, metadata_path)
print(f"[source] loading {source_metadata}", flush=True)
source_index = load_source_index(source_root, source_metadata)
print(f"[source] valid source motions/videos: {len(source_index)}", flush=True)
timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
tmp_dir = clip_root / f".source_motion_patch_tmp.{timestamp}.{os.getpid()}"
tmp_dir.mkdir(parents=False, exist_ok=False)
try:
part_paths, totals = run_workers(
metadata_path=metadata_path,
tmp_dir=tmp_dir,
source_index=source_index,
workers=args.workers,
progress_bytes=args.progress_bytes,
)
if totals["errors"]:
preview = "\n".join(totals["error_examples"][:20])
raise RuntimeError(f"patch produced {totals['errors']} error(s); first issues:\n{preview}")
patched_metadata = tmp_dir / "metadata.jsonl"
merge_parts(part_paths, patched_metadata)
new_lines = sum(1 for _ in patched_metadata.open("rb"))
if new_lines != totals["rows"]:
raise RuntimeError(f"line count mismatch after merge: worker_rows={totals['rows']} merged_lines={new_lines}")
print(
"[summary] "
f"rows={totals['rows']} patched={totals['patched']} "
f"already_source_motion={totals['already_had_source_motion']} "
f"already_source_video={totals['already_had_source_video']} "
f"output={patched_metadata}",
flush=True,
)
if not args.apply:
print("[dry-run] not replacing metadata.jsonl. Re-run with --apply to install the patched file.", flush=True)
return
backup_path = clip_root / f"{args.metadata}.before_source_motion_patch.{timestamp}"
print(f"[apply] backup original -> {backup_path}", flush=True)
os.replace(metadata_path, backup_path)
try:
os.replace(patched_metadata, metadata_path)
except BaseException:
if metadata_path.exists():
failed_path = clip_root / f"{args.metadata}.failed_source_motion_patch.{timestamp}"
os.replace(metadata_path, failed_path)
os.replace(backup_path, metadata_path)
raise
print(f"[apply] installed patched metadata: {metadata_path}", flush=True)
print(f"[apply] backup kept at: {backup_path}", flush=True)
finally:
if not args.keep_temp and tmp_dir.exists():
shutil.rmtree(tmp_dir, ignore_errors=True)
if __name__ == "__main__":
try:
main()
except KeyboardInterrupt:
print("interrupted", file=sys.stderr)
raise