File size: 3,347 Bytes
a72e446
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6955e32
a72e446
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6955e32
 
 
a72e446
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
"""
Convert a parquet-format HuggingFace dataset back to raw image files.

Input:  a local directory of .parquet shards  OR  an HF repo ID
Output: GSD/{split}/{image,mask,reflections,edge}/<image_id>.png  +  metadata.jsonl
"""
import argparse
import json
from pathlib import Path

from datasets import load_dataset, load_from_disk

# split -> (has_reflections, has_edge)
SPLITS = {
    "train": (True,  True),
    "extra": (False, True),
    "test":  (False, False),
}


def save_split(ds, split: str, has_reflections: bool, has_edge: bool, out_dir: Path):
    split_dir = out_dir / split
    img_dir  = split_dir / "image";        img_dir.mkdir(parents=True, exist_ok=True)
    mask_dir = split_dir / "mask";         mask_dir.mkdir(parents=True, exist_ok=True)
    refl_dir = split_dir / "reflections" if has_reflections else None
    edge_dir = split_dir / "edge"        if has_edge        else None
    if refl_dir: refl_dir.mkdir(parents=True, exist_ok=True)
    if edge_dir: edge_dir.mkdir(parents=True, exist_ok=True)

    rows = []
    for i, sample in enumerate(ds):
        stem = sample.get("image_id") or f"{i:06d}"

        img_path  = img_dir  / f"{stem}.jpg"
        mask_path = mask_dir / f"{stem}.png"
        sample["image"].save(img_path)
        sample["mask"].save(mask_path)

        refl_rel = edge_rel = ""
        if has_reflections and sample.get("reflections") is not None:
            p = refl_dir / f"{stem}.png"
            sample["reflections"].save(p)
            refl_rel = f"reflections/{p.name}"
        if has_edge and sample.get("edge") is not None:
            p = edge_dir / f"{stem}.png"
            sample["edge"].save(p)
            edge_rel = f"edge/{p.name}"

        rows.append({
            "file_name":   f"image/{img_path.name}",
            "mask":        f"mask/{mask_path.name}",
            "reflections": refl_rel,
            "edge":        edge_rel,
        })

        if (i + 1) % 100 == 0:
            print(f"  {i + 1}/{len(ds)}")

    # (split_dir / "metadata.jsonl").write_text(
    #     "\n".join(json.dumps(r) for r in rows)
    # )
    print(f"  saved {len(rows)} samples -> {split_dir}")


def main():
    parser = argparse.ArgumentParser(description="Convert parquet dataset to raw images")
    src = parser.add_mutually_exclusive_group(required=True)
    src.add_argument("--repo",  help="HuggingFace repo ID, e.g. garrying/GSD")
    src.add_argument("--local", help="Path to a directory saved with save_to_disk()")
    parser.add_argument("--out", default="GSD", help="Output root directory (default: ./GSD)")
    parser.add_argument("--splits", nargs="+", choices=list(SPLITS), default=list(SPLITS),
                        help="Which splits to convert (default: all)")
    args = parser.parse_args()

    out_dir = Path(args.out)
    out_dir.mkdir(parents=True, exist_ok=True)

    for split in args.splits:
        has_reflections, has_edge = SPLITS[split]
        print(f"\nLoading {split}...")

        if args.repo:
            ds = load_dataset(args.repo, split=split)
        else:
            ds = load_from_disk(str(Path(args.local) / split))

        print(f"  {len(ds)} samples — saving images...")
        save_split(ds, split, has_reflections, has_edge, out_dir)

    print(f"\nDone!  Raw files in: {out_dir.resolve()}")


if __name__ == "__main__":
    main()