Grounding-EvalData / scripts /prepare_data.py
Skywalker0410's picture
Release complete academic grounding evaluation data
772e34b verified
Raw History Blame Contribute Delete
12.9 kB
#!/usr/bin/env python3
"""Verify and unpack Grounding-EvalData, then write local release path configs.
Python 3.9+ standard library only. No downloads, inference, or uploads are made.
"""
from __future__ import annotations
import argparse
import glob
import hashlib
import json
import os
from pathlib import Path, PurePosixPath, PureWindowsPath
import re
import shutil
import sys
import tarfile
import tempfile
DATA_PATHS = {
'images': 'images',
'screenspot_pro': 'ScreenSpot-Pro',
'screenspot_v2': 'ScreenSpot-v2',
'osworld_g': 'OSWorld-G',
'refspatial': 'RefSpatial-Bench',
'refcoco': 'RefCOCO',
'refcocoplus': 'RefCOCOplus',
'refcocog': 'RefCOCOg',
'gam_robospatial_context': 'RoboSpatial-Home/data/context-00000-of-00001.parquet',
'gam_osworld_g': 'OSWorld-G/OSWorld-G.json',
'gam_screenspot_pro': 'ScreenSpot-Pro/annotations/*.json',
'gam_screenspot_v2_1': 'ScreenSpot-v2/screenspot_desktop_v2.json',
'gam_screenspot_v2_2': 'ScreenSpot-v2/screenspot_mobile_v2.json',
'gam_screenspot_v2_3': 'ScreenSpot-v2/screenspot_web_v2.json',
}
ANNOTATIONS = {
'gam_coco': 'box_eval/COCO.jsonl',
'gam_lvis': 'box_eval/LVIS.jsonl',
'gam_refcocog_test': 'box_eval/RefCOCOg_test.jsonl',
'gam_refcocog_val': 'box_eval/RefCOCOg_val.jsonl',
'gam_humanref': 'box_eval/HumanRef.jsonl',
'gam_m6doc': 'box_eval/M6Doc.jsonl',
'gam_doclaynet': 'box_eval/DocLayNet.jsonl',
'gam_hiertext': 'box_eval/HierText.jsonl',
'gam_totaltext': 'box_eval/TotalText.jsonl',
'gam_icdar2015': 'box_eval/IC15.jsonl',
'gam_sroie': 'box_eval/SROIE.jsonl',
'gam_visdrone': 'box_eval/VisDrone.jsonl',
'gam_dense200': 'box_eval/Dense200.jsonl',
'gam_rex_point_coco': 'point_eval/COCO.jsonl',
'gam_rex_point_refcocog_test': 'point_eval/RefCOCOg_test.jsonl',
'gam_rex_point_lvis': 'point_eval/LVIS.jsonl',
'gam_rex_point_visdrone': 'point_eval/VisDrone.jsonl',
'gam_rex_point_dense200': 'point_eval/Dense200.jsonl',
'gam_rex_point_refcocog_val': 'point_eval/RefCOCOg_val.jsonl',
'gam_rex_point_humanref': 'point_eval/HumanRef.jsonl',
'gam_visual_coco': 'visual_prompt_eval/COCO.jsonl',
'gam_visual_lvis': 'visual_prompt_eval/LVIS.jsonl',
'gam_visual_dense200': 'visual_prompt_eval/Dense200.jsonl',
'gam_fsc147': 'visual_prompt_eval/FSCD_test.jsonl',
}
def relative_path(value, allow_dot=False):
if not isinstance(value, str) or not value or '\x00' in value or '\\' in value:
raise ValueError(f'Unsafe relative path: {value!r}')
path = PurePosixPath(value)
if path.is_absolute() or PureWindowsPath(value).drive or '..' in path.parts:
raise ValueError(f'Unsafe relative path: {value!r}')
if str(path) == '.' and not allow_dot:
raise ValueError(f'Empty relative path: {value!r}')
return Path(*path.parts)
def sha256(path):
digest = hashlib.sha256()
with path.open('rb') as handle:
for block in iter(lambda: handle.read(8 * 1024 * 1024), b''):
digest.update(block)
return digest.hexdigest()
def safe_destination(root, relative):
destination = root / relative
for part in (root, *destination.parents, destination):
if part.is_symlink():
raise ValueError(f'Refusing symlink destination: {part}')
if not destination.resolve().is_relative_to(root.resolve()):
raise ValueError(f'Path escapes output root: {relative}')
return destination
def verify_bundle(repo, manifest):
"""Verify every checksum entry and require checksums for loose evaluation data."""
checksums = {}
checksum_path = repo / 'SHA256SUMS'
if not checksum_path.is_file():
raise FileNotFoundError(f'Missing checksum inventory: {checksum_path}')
for number, line in enumerate(checksum_path.read_text(encoding='utf-8').splitlines(), 1):
if not line.strip():
continue
match = re.fullmatch(r'([0-9a-fA-F]{64}) [ *](.+)', line)
if not match:
raise ValueError(f'Malformed SHA256SUMS line {number}')
name = relative_path(match.group(2)).as_posix()
if name in checksums:
raise ValueError(f'Duplicate checksum entry: {name}')
checksums[name] = match.group(1).lower()
archives = manifest.get('archives')
if not isinstance(archives, list) or not archives:
raise ValueError('package_manifest.json must contain a nonempty archives list')
archive_names = set()
missing = []
for item in archives:
name = relative_path(item['path']).as_posix()
if name in archive_names:
raise ValueError(f'Duplicate archive manifest entry: {name}')
archive_names.add(name)
relative_path(item.get('extract_to', '.'), allow_dot=True)
digest = item.get('sha256', '').lower()
if not re.fullmatch(r'[0-9a-f]{64}', digest):
raise ValueError(f'Missing or invalid archive SHA256: {name}')
if name not in checksums or checksums[name] != digest:
raise ValueError(f'Archive checksum absent or inconsistent: {name}')
source = repo / name
if not source.is_file():
missing.append(f"{name}; tasks={item.get('tasks', [])}; source={item.get('source_url', '')}")
elif item.get('bytes') is not None and source.stat().st_size != item['bytes']:
raise ValueError(f'Archive size mismatch: {name}')
if missing:
raise FileNotFoundError('Missing required archives:\n ' + '\n '.join(missing))
for name in [*(f'_annotations/{name}' for name in ANNOTATIONS.values()),
'support/OCR/Challenge4_Test_Task1_GT.zip']:
if name not in checksums:
raise ValueError(f'Evaluation input is absent from SHA256SUMS: {name}')
for index, (name, expected) in enumerate(checksums.items(), 1):
path = repo / relative_path(name)
if not path.is_file():
raise FileNotFoundError(f'Missing checksummed file: {name}')
if path.is_symlink() or not path.resolve().is_relative_to(repo):
raise ValueError(f'Refusing checksum path outside bundle or symlink: {name}')
if sha256(path) != expected:
raise ValueError(f'SHA256 mismatch: {name}')
if name in archive_names:
print(f'Verified {name} ({index}/{len(checksums)} checksum entries)', flush=True)
print(f'All {len(checksums)} checksums verified.', flush=True)
return archives
def extract_archive(source, root):
"""Extract regular files/directories only; never replace different existing data."""
root.mkdir(parents=True, exist_ok=True)
seen = set()
written = skipped = 0
with tarfile.open(source, mode='r|*') as archive:
for member in archive:
relative = relative_path(member.name, allow_dot=member.isdir())
normalized = str(relative)
if normalized in seen and not member.isdir():
raise ValueError(f'Duplicate archive member: {member.name}')
seen.add(normalized)
if not (member.isdir() or member.isfile()) or member.issym() or member.islnk():
raise ValueError(f'Unsupported archive member type: {member.name}')
destination = safe_destination(root, relative)
if member.isdir():
destination.mkdir(parents=True, exist_ok=True)
continue
destination.parent.mkdir(parents=True, exist_ok=True)
if destination.exists() and not destination.is_file():
raise ValueError(f'Existing path is not a regular file: {destination}')
content = archive.extractfile(member)
if content is None:
raise ValueError(f'Cannot read archive member: {member.name}')
fd, temporary_name = tempfile.mkstemp(prefix='.grounding-extract-', dir=destination.parent)
temporary = Path(temporary_name)
try:
with os.fdopen(fd, 'wb') as output, content:
shutil.copyfileobj(content, output, length=1024 * 1024)
if temporary.stat().st_size != member.size:
raise ValueError(f'Truncated archive member: {member.name}')
if destination.exists():
if destination.stat().st_size != member.size or sha256(destination) != sha256(temporary):
raise FileExistsError(f'Refusing to replace different existing file: {destination}')
skipped += 1
else:
temporary.chmod(0o644)
os.replace(temporary, destination)
written += 1
finally:
temporary.unlink(missing_ok=True)
return written, skipped
def make_data_paths(repo, data_root):
paths = {key: str(data_root / value) for key, value in DATA_PATHS.items()}
paths.update({key: str(repo / '_annotations' / value) for key, value in ANNOTATIONS.items()})
paths['icdar2015_dontcare'] = str(repo / 'support/OCR/Challenge4_Test_Task1_GT.zip')
return paths
def write_config(path, paths):
text = '# Generated by Grounding-EvalData prepare_data.py; paths refer to this machine.\n'
text += '\n'.join(f'{key}: {json.dumps(value, ensure_ascii=False)}' for key, value in paths.items()) + '\n'
if path.is_symlink():
raise ValueError(f'Refusing symlink config: {path}')
if path.exists():
if path.read_text(encoding='utf-8') != text:
raise FileExistsError(f'Existing config differs; preserve or rename it first: {path}')
else:
with path.open('x', encoding='utf-8') as handle:
handle.write(text)
print(f'Path configuration: {path}', flush=True)
def main(argv=None):
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument('--archives-dir', type=Path, default=Path(__file__).resolve().parents[1],
help='HF repository root containing package_manifest.json and SHA256SUMS')
parser.add_argument('--data-root', type=Path, help='Directory for extracted evaluation data')
parser.add_argument('--code-root', action='append', type=Path, default=[],
help='GroundAnything or GroundingPI checkout; repeat for both releases')
parser.add_argument('--verify-only', action='store_true', help='Verify all package hashes without extraction')
args = parser.parse_args(argv)
if not args.verify_only and args.data_root is None:
parser.error('--data-root is required unless --verify-only is selected')
repo = args.archives_dir.expanduser().resolve(strict=True)
manifest = json.loads((repo / 'package_manifest.json').read_text(encoding='utf-8'))
archives = verify_bundle(repo, manifest)
if args.verify_only:
return 0
data_root = args.data_root.expanduser().resolve()
code_roots = [path.expanduser().resolve(strict=True) for path in args.code_root]
for code_root in code_roots:
if not (code_root / 'configs/datasets.yaml').is_file() or not (code_root / 'scripts/evaluate.py').is_file():
raise ValueError(f'Not a supported release checkout: {code_root}')
data_root.mkdir(parents=True, exist_ok=True)
for item in archives:
destination = safe_destination(data_root, relative_path(item.get('extract_to', '.'), allow_dot=True))
print(f'Extracting {item["path"]} -> {destination}', flush=True)
written, skipped = extract_archive(repo / relative_path(item['path']), destination)
print(f' {written} files written, {skipped} identical files reused', flush=True)
paths = make_data_paths(repo, data_root)
missing = [f'{key}: {value}' for key, value in paths.items() if not glob.glob(value)]
if missing:
raise FileNotFoundError('Unpacked data does not satisfy release paths:\n ' + '\n '.join(missing))
outputs = [root / 'configs/datasets.local.yaml' for root in code_roots]
if not outputs:
outputs = [data_root / 'datasets.local.yaml']
for output in outputs:
write_config(output, paths)
print('\nIn each full-suite evaluation YAML, set:')
print(' data_root: ' + json.dumps(str(data_root)))
print(' datasets: ' + ('configs/datasets.local.yaml' if code_roots else json.dumps(str(outputs[0]))))
print(' mode: GAM')
print(' limit: null')
print('Keep model/service settings from the corresponding release and select its documented 30 or 34 tasks.')
print('Then run: .venv-eval/bin/python scripts/evaluate.py <your-full-suite.yaml> --dry-run')
print('Path checks passed; this does not execute model inference or establish evaluation scores.')
return 0
if __name__ == '__main__':
try:
raise SystemExit(main())
except (OSError, ValueError, KeyError, tarfile.TarError) as exc:
print(f'ERROR: {exc}', file=sys.stderr)
raise SystemExit(2)