May-apple commited on
Commit
7b143b1
·
verified ·
1 Parent(s): accb95e

add extract_parquet_to_dir.py: stream parquet shards back to sample-dir layout

Browse files
Files changed (1) hide show
  1. extract_parquet_to_dir.py +176 -0
extract_parquet_to_dir.py ADDED
@@ -0,0 +1,176 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python3
2
+ """Materialize VBVR-Reorganized parquet shards back into the on-disk
3
+ sample-directory layout described in the dataset README.
4
+
5
+ Each row of the parquet contains the full bytes of `first_frame.png`,
6
+ `final_frame.png`, and `ground_truth.mp4` plus the `prompt` /
7
+ `prompt_original` text and the raw `metadata_json` string. This script
8
+ streams the parquet (no full-file load) and writes one directory per row
9
+ under:
10
+
11
+ <out>/<split>/<class>/<task>/<sample_id>/
12
+ first_frame.png
13
+ final_frame.png
14
+ ground_truth.mp4
15
+ prompt.txt
16
+ prompt_original.txt (only if present)
17
+ metadata.json (only if non-empty)
18
+
19
+ `split` is taken from the row's `split` column verbatim (e.g.
20
+ `test/In-Domain_50`, `train`, `train_pr_rewritten`). `sample_id`
21
+ already encodes the per-task wrapper layer for train rows
22
+ (`<task>_task/<task>_<id>`) so paths reproduce the source tree exactly.
23
+
24
+ USAGE
25
+
26
+ # 1) extract a single parquet shard:
27
+ python3 extract_parquet_to_dir.py \\
28
+ --parquet parquet/test_in_domain.parquet \\
29
+ --out ./VBVR-extracted
30
+
31
+ # 2) extract every parquet under parquet/ into one tree:
32
+ python3 extract_parquet_to_dir.py \\
33
+ --parquet parquet/ \\
34
+ --out ./VBVR-extracted
35
+
36
+ # 3) extract only one task from a train shard:
37
+ python3 extract_parquet_to_dir.py \\
38
+ --parquet parquet/train__G-13_grid_number_sequence.parquet \\
39
+ --out ./VBVR-extracted \\
40
+ --task G-13_grid_number_sequence_data-generator
41
+
42
+ The script never loads more than one row-group into memory at a time, so
43
+ it's safe on the multi-GB train shards.
44
+ """
45
+ from __future__ import annotations
46
+
47
+ import argparse
48
+ import sys
49
+ import time
50
+ from pathlib import Path
51
+
52
+ import pyarrow.parquet as pq
53
+
54
+
55
+ def iter_parquet_files(target: Path):
56
+ if target.is_file():
57
+ yield target
58
+ elif target.is_dir():
59
+ yield from sorted(target.glob("*.parquet"))
60
+ else:
61
+ raise FileNotFoundError(target)
62
+
63
+
64
+ def write_row(out_root: Path, row: dict, *, overwrite: bool) -> bool:
65
+ split = row["split"] or ""
66
+ cls = row["class"] or ""
67
+ task = row["task"] or ""
68
+ sid = row["sample_id"] or ""
69
+ if not (split and cls and task and sid):
70
+ print(f" ! row missing identifying fields, skipped: {row.get('sample_id')!r}",
71
+ file=sys.stderr)
72
+ return False
73
+ dst = out_root / split / cls / task / sid
74
+ if dst.exists() and not overwrite:
75
+ # Reuse existing directory but still (re)write files in case bytes changed.
76
+ pass
77
+ dst.mkdir(parents=True, exist_ok=True)
78
+
79
+ ff = row.get("first_frame")
80
+ if ff and ff.get("bytes"):
81
+ (dst / "first_frame.png").write_bytes(ff["bytes"])
82
+ ef = row.get("final_frame")
83
+ if ef and ef.get("bytes"):
84
+ (dst / "final_frame.png").write_bytes(ef["bytes"])
85
+ gv = row.get("ground_truth_video")
86
+ if gv and gv.get("bytes"):
87
+ (dst / "ground_truth.mp4").write_bytes(gv["bytes"])
88
+
89
+ prompt = row.get("prompt")
90
+ if prompt is not None:
91
+ (dst / "prompt.txt").write_text(prompt, encoding="utf-8")
92
+ op = row.get("prompt_original")
93
+ if op is not None and op != "":
94
+ (dst / "prompt_original.txt").write_text(op, encoding="utf-8")
95
+ mj = row.get("metadata_json")
96
+ if mj:
97
+ (dst / "metadata.json").write_text(mj, encoding="utf-8")
98
+ return True
99
+
100
+
101
+ def extract_one_parquet(pq_path: Path, out_root: Path, *,
102
+ only_class: str | None,
103
+ only_task: str | None,
104
+ only_sample: str | None,
105
+ overwrite: bool) -> tuple[int, int]:
106
+ """Stream a single parquet file; return (rows_written, rows_skipped)."""
107
+ pf = pq.ParquetFile(str(pq_path))
108
+ written = 0
109
+ skipped = 0
110
+ t0 = time.time()
111
+ print(f"\n[{pq_path.name}] {pf.metadata.num_rows} rows in "
112
+ f"{pf.num_row_groups} row groups", flush=True)
113
+ for rg_idx in range(pf.num_row_groups):
114
+ table = pf.read_row_group(rg_idx)
115
+ rows = table.to_pylist()
116
+ for row in rows:
117
+ if only_class and row.get("class") != only_class:
118
+ skipped += 1; continue
119
+ if only_task and row.get("task") != only_task:
120
+ skipped += 1; continue
121
+ if only_sample and row.get("sample_id") != only_sample:
122
+ skipped += 1; continue
123
+ if write_row(out_root, row, overwrite=overwrite):
124
+ written += 1
125
+ elapsed = time.time() - t0
126
+ print(f" rg {rg_idx+1}/{pf.num_row_groups} "
127
+ f"+{len(rows)} rows "
128
+ f"(written={written} skipped={skipped} {elapsed:.0f}s)",
129
+ flush=True)
130
+ return written, skipped
131
+
132
+
133
+ def main(argv: list[str]) -> int:
134
+ ap = argparse.ArgumentParser(description=__doc__,
135
+ formatter_class=argparse.RawDescriptionHelpFormatter)
136
+ ap.add_argument("--parquet", required=True, type=Path,
137
+ help="Path to a .parquet file OR a directory containing .parquet files.")
138
+ ap.add_argument("--out", required=True, type=Path,
139
+ help="Output root directory. Sample dirs are written under "
140
+ "<out>/<split>/<class>/<task>/<sample_id>/.")
141
+ ap.add_argument("--class", dest="only_class", default=None,
142
+ help="Only extract rows whose `class` column equals this "
143
+ "(e.g. 'Pure_Reasoning' or 'Instruction_Following').")
144
+ ap.add_argument("--task", dest="only_task", default=None,
145
+ help="Only extract rows whose `task` column equals this "
146
+ "(e.g. 'G-13_grid_number_sequence_data-generator').")
147
+ ap.add_argument("--sample", dest="only_sample", default=None,
148
+ help="Only extract the row whose `sample_id` equals this exact string.")
149
+ ap.add_argument("--overwrite", action="store_true",
150
+ help="(Default: also true.) Overwrite files in existing dirs. "
151
+ "This flag is kept for clarity.")
152
+ args = ap.parse_args(argv)
153
+
154
+ args.out.mkdir(parents=True, exist_ok=True)
155
+ files = list(iter_parquet_files(args.parquet))
156
+ if not files:
157
+ print(f"no parquet files under {args.parquet}", file=sys.stderr)
158
+ return 1
159
+ print(f"extracting {len(files)} parquet file(s) into {args.out}", flush=True)
160
+
161
+ total_written = 0; total_skipped = 0
162
+ for f in files:
163
+ w, s = extract_one_parquet(f, args.out,
164
+ only_class=args.only_class,
165
+ only_task=args.only_task,
166
+ only_sample=args.only_sample,
167
+ overwrite=args.overwrite)
168
+ total_written += w; total_skipped += s
169
+
170
+ print(f"\ndone: {total_written} rows written, {total_skipped} rows skipped.")
171
+ print(f"output tree at {args.out}")
172
+ return 0
173
+
174
+
175
+ if __name__ == "__main__":
176
+ sys.exit(main(sys.argv[1:]))