| |
| """Unpack the frame archives of this dataset and (optionally) write a parquet with absolute frame paths. |
| |
| python extract_frames.py # unzip frames_zip/*.zip -> ./frames/<frame_dir>/000.jpg ... 031.jpg |
| python extract_frames.py --abs # also write train.abs.parquet (frame paths made absolute; needs pyarrow) |
| python extract_frames.py --out /data # extract somewhere else (frames end up in /data/frames) |
| |
| `train.parquet` stores frame paths relative to the extraction root (`frames/<frame_dir>/NNN.jpg`). |
| """ |
| import argparse |
| import glob |
| import hashlib |
| import json |
| import os |
| import zipfile |
|
|
| HERE = os.path.dirname(os.path.abspath(__file__)) |
|
|
|
|
| def sha256(path): |
| h = hashlib.sha256() |
| with open(path, "rb") as f: |
| for blk in iter(lambda: f.read(1 << 24), b""): |
| h.update(blk) |
| return h.hexdigest() |
|
|
|
|
| def main(): |
| ap = argparse.ArgumentParser() |
| ap.add_argument("--out", default=HERE, help="extraction root (default: the dataset folder)") |
| ap.add_argument("--abs", action="store_true", help="write train.abs.parquet with absolute frame paths") |
| ap.add_argument("--no-verify", action="store_true", help="skip the sha256 check of the zip files") |
| args = ap.parse_args() |
| out = os.path.abspath(args.out) |
| manifest = json.load(open(os.path.join(HERE, "manifest.json"))) |
| want = {z["file"]: z for z in manifest["zips"]} |
| zips = sorted(glob.glob(os.path.join(HERE, "frames_zip", "frames-*.zip"))) |
| assert len(zips) == len(want), f"expected {len(want)} zip files, found {len(zips)}" |
| for zp in zips: |
| rel = "frames_zip/" + os.path.basename(zp) |
| if not args.no_verify: |
| assert sha256(zp) == want[rel]["sha256"], f"sha256 mismatch: {rel} (re-download it)" |
| with zipfile.ZipFile(zp) as z: |
| z.extractall(out) |
| print("extracted", rel, flush=True) |
| n = len(glob.glob(os.path.join(out, "frames", "*", "*.jpg"))) |
| assert n == manifest["frame_files"], f"expected {manifest['frame_files']} frames, found {n}" |
| print(f"ok: {n} frames in {os.path.join(out, 'frames')}") |
| if args.abs: |
| import pyarrow as pa |
| import pyarrow.parquet as pq |
| t = pq.read_table(os.path.join(HERE, "train.parquet")) |
| rows = t.to_pylist() |
| for r in rows: |
| r["videos"][0]["video"] = [os.path.join(out, f) for f in r["videos"][0]["video"]] |
| dst = os.path.join(out, "train.abs.parquet") |
| pq.write_table(pa.Table.from_pylist(rows, schema=t.schema), dst) |
| print("wrote", dst) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
