File size: 7,686 Bytes
f08526f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
#!/usr/bin/env python3
"""Materialize Visual_Agent Parquet tables as path-based JSONL and images."""

from __future__ import annotations

import argparse
import hashlib
import json
import os
import tempfile
from pathlib import Path
from typing import Any, Iterable

import pyarrow.parquet as pq


def sha256_bytes(value: bytes) -> str:
    return hashlib.sha256(value).hexdigest()


def sha256_file(path: Path) -> str:
    digest = hashlib.sha256()
    with path.open("rb") as handle:
        for block in iter(lambda: handle.read(1024 * 1024), b""):
            digest.update(block)
    return digest.hexdigest()


def safe_relative_path(value: str) -> Path:
    path = Path(value)
    if path.is_absolute() or not path.parts or ".." in path.parts:
        raise ValueError(f"unsafe relative path: {value!r}")
    if path.parts[0] != "images":
        raise ValueError(f"image path must be rooted under images/: {value!r}")
    return path


def write_bytes_atomic(path: Path, value: bytes) -> None:
    path.parent.mkdir(parents=True, exist_ok=True)
    with tempfile.NamedTemporaryFile(
        dir=path.parent,
        prefix=path.name + ".",
        suffix=".tmp",
        delete=False,
    ) as handle:
        temporary = Path(handle.name)
        handle.write(value)
        handle.flush()
        os.fsync(handle.fileno())
    os.replace(temporary, path)


def write_jsonl_atomic(path: Path, rows: Iterable[dict[str, Any]]) -> None:
    path.parent.mkdir(parents=True, exist_ok=True)
    temporary = path.with_name(path.name + ".tmp")
    with temporary.open("w", encoding="utf-8") as handle:
        for row in rows:
            handle.write(json.dumps(row, ensure_ascii=False) + "\n")
    os.replace(temporary, path)


def materialize_images(input_root: Path, training_root: Path) -> dict[str, int]:
    image_shards = sorted((input_root / "images").glob("*.parquet"))
    if not image_shards:
        raise ValueError(f"no image Parquet shards under {input_root / 'images'}")

    counts = {"rows": 0, "written": 0, "reused": 0}
    seen_paths: set[str] = set()
    for shard in image_shards:
        parquet = pq.ParquetFile(shard)
        for batch in parquet.iter_batches(batch_size=32):
            for row in batch.to_pylist():
                relative_text = str(row["path"])
                if relative_text in seen_paths:
                    raise ValueError(f"duplicate image path in Parquet: {relative_text}")
                seen_paths.add(relative_text)
                relative = safe_relative_path(relative_text)
                value = bytes(row["bytes"])
                expected_size = int(row["size_bytes"])
                expected_sha256 = str(row["sha256"])
                if len(value) != expected_size:
                    raise ValueError(f"{relative_text}: byte length does not match metadata")
                if sha256_bytes(value) != expected_sha256:
                    raise ValueError(f"{relative_text}: embedded bytes fail SHA-256 validation")

                destination = training_root / relative
                if destination.exists():
                    if destination.stat().st_size != expected_size:
                        raise ValueError(f"{destination}: existing file has a different size")
                    if sha256_file(destination) != expected_sha256:
                        raise ValueError(f"{destination}: existing file has different content")
                    counts["reused"] += 1
                else:
                    write_bytes_atomic(destination, value)
                    counts["written"] += 1
                counts["rows"] += 1
    return counts


def materialize_samples(input_root: Path, training_root: Path) -> dict[str, int]:
    sample_shards = sorted((input_root / "samples").glob("*.parquet"))
    if not sample_shards:
        raise ValueError(f"no sample Parquet shards under {input_root / 'samples'}")

    indexed_rows: list[tuple[int, str, dict[str, Any]]] = []
    seen_row_ids: set[str] = set()
    for shard in sample_shards:
        parquet = pq.ParquetFile(shard)
        for batch in parquet.iter_batches(
            batch_size=256,
            columns=["row_index", "row_id", "record_json"],
        ):
            for value in batch.to_pylist():
                row_index = int(value["row_index"])
                row_id = str(value["row_id"])
                if row_id in seen_row_ids:
                    raise ValueError(f"duplicate row_id in samples table: {row_id}")
                seen_row_ids.add(row_id)
                record = json.loads(str(value["record_json"]))
                if not isinstance(record, dict):
                    raise ValueError(f"{row_id}: record_json is not an object")
                for image in record.get("images") or []:
                    relative = safe_relative_path(str(image))
                    if not (training_root / relative).is_file():
                        raise ValueError(f"{row_id}: materialized image is missing: {image}")
                indexed_rows.append((row_index, row_id, record))

    indexed_rows.sort(key=lambda item: item[0])
    expected_indexes = list(range(len(indexed_rows)))
    actual_indexes = [item[0] for item in indexed_rows]
    if actual_indexes != expected_indexes:
        raise ValueError("samples row_index is not a complete zero-based sequence")

    combined_path = training_root / "all_training_trajectories_with_images.jsonl"
    write_jsonl_atomic(combined_path, (item[2] for item in indexed_rows))

    p2r_rows = [item[2] for item in indexed_rows if str(item[1]).startswith("p2r_")]
    p2r_path = training_root / "p2r_v2/p2r_natural_v2_repaired_with_images.jsonl"
    write_jsonl_atomic(p2r_path, p2r_rows)
    return {
        "rows": len(indexed_rows),
        "p2r_v2_rows": len(p2r_rows),
        "combined_jsonl_sha256": sha256_file(combined_path),
    }


def materialize(input_root: Path, output_root: Path) -> dict[str, Any]:
    input_root = input_root.resolve()
    output_root = output_root.resolve()
    manifest_path = input_root / "dataset_manifest.json"
    if not manifest_path.is_file():
        raise ValueError(f"missing manifest: {manifest_path}")
    manifest = json.loads(manifest_path.read_text(encoding="utf-8"))
    training_root = output_root / "training_trajectories_natural"
    training_root.mkdir(parents=True, exist_ok=True)

    image_summary = materialize_images(input_root, training_root)
    sample_summary = materialize_samples(input_root, training_root)
    if image_summary["rows"] != int(manifest["unique_images"]):
        raise ValueError("materialized image count does not match manifest")
    if sample_summary["rows"] != int(manifest["samples"]):
        raise ValueError("materialized sample count does not match manifest")
    if sample_summary["combined_jsonl_sha256"] != str(
        manifest["combined_jsonl_sha256"]
    ):
        raise ValueError("materialized combined JSONL does not match manifest SHA-256")

    return {
        "output_root": str(output_root),
        "samples": sample_summary,
        "images": image_summary,
        "combined_jsonl": str(
            training_root / "all_training_trajectories_with_images.jsonl"
        ),
    }


def main() -> None:
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument("--input-root", type=Path, default=Path("."))
    parser.add_argument("--output-root", type=Path, required=True)
    args = parser.parse_args()
    print(
        json.dumps(
            materialize(args.input_root, args.output_root),
            ensure_ascii=False,
            indent=2,
            sort_keys=True,
        )
    )


if __name__ == "__main__":
    main()