File size: 7,571 Bytes
758421f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Export frozen selected units as test-only data; keep the source intact."""
import argparse
from collections import Counter
import copy
import hashlib
import json
from pathlib import Path
import shutil


def sha256(path):
    h = hashlib.sha256()
    with path.open('rb') as stream:
        for chunk in iter(lambda: stream.read(1024 * 1024), b''):
            h.update(chunk)
    return h.hexdigest()


def save_json(path, value):
    path.parent.mkdir(parents=True, exist_ok=True)
    path.write_text(json.dumps(value, ensure_ascii=False, indent=2) + '\n')


def asset_paths(row):
    yield row['image']['path']
    for region in row['reference']['regions'] + row.get('intervention_regions', []):
        mask = region.get('mask')
        if mask and mask['encoding'] == 'binary_png':
            yield mask['path']


def export(source, destination):
    source, destination = source.resolve(), destination.resolve()
    if destination.exists():
        raise FileExistsError(destination)
    if destination.is_relative_to(source) or source.is_relative_to(destination):
        raise ValueError('Source and destination must be separate')
    original = json.loads((source / 'build_report.json').read_text())
    if original['status'] != 'complete':
        raise ValueError('Source preparation is incomplete')
    report = copy.deepcopy(original)
    report['status'] = 'building'
    report['catalog_scope'] = 'frozen selected units only; all assigned to test'
    report['config'].pop('development_fraction', None)
    report['config']['hallusegbench'].pop('development_source', None)
    report['config']['split_policy'] = 'all_selected_are_test'
    report['split_policy'] = 'test_only'
    report['source_build_report_sha256'] = sha256(source / 'build_report.json')
    save_json(destination / 'build_report.json', report)
    assets = set()
    split_sources, split_images = {}, {}
    all_ids = set()
    export_counts = {}
    for name, counts in report['datasets'].items():
        selection_path = source / 'selections' / f'{name}.json'
        selection = json.loads(selection_path.read_text())
        allowed = set(selection['selected_unit_ids'])
        src_manifest = source / 'manifests' / f'{name}.jsonl'
        if sha256(src_manifest) != counts['manifest_sha256'] or sha256(selection_path) != counts['selection_sha256']:
            raise ValueError(f'Source checksums changed: {name}')
        counts['source_selection_sha256'] = counts['selection_sha256']
        selection['test_unit_ids'] = list(selection['selected_unit_ids'])
        selection['development_unit_ids'] = []
        selection['split_policy'] = 'test_only'
        selection['policy'] = report['config'][name]
        lines, rows = [], []
        for line in src_manifest.read_text().splitlines(keepends=True):
            row = json.loads(line)
            if row['unit_id'] in allowed:
                if not row['selection_eligible'] or row['sample_id'] in all_ids:
                    raise ValueError(f'Invalid selected sample: {row["sample_id"]}')
                all_ids.add(row['sample_id'])
                rows.append(row)
                lines.append(line)
                assets.update(asset_paths(row))
        if {r['unit_id'] for r in rows} != allowed:
            raise ValueError(f'Missing selected units: {name}')
        if name == 'hallusegbench':
            pairs = {}
            for row in rows:
                pairs.setdefault(row['unit_id'], set()).add(row['condition'])
            expected = {'factual_original', 'factual_replacement',
                        'counterfactual_original', 'counterfactual_replacement'}
            if any(conditions != expected for conditions in pairs.values()):
                raise ValueError('Incomplete counterfactual pair')
        out_manifest = destination / 'manifests' / f'{name}.jsonl'
        out_manifest.parent.mkdir(parents=True, exist_ok=True)
        out_manifest.write_text(''.join(lines))
        out_selection = destination / 'selections' / selection_path.name
        out_selection.parent.mkdir(parents=True, exist_ok=True)
        save_json(out_selection, selection)
        counts['selection_sha256'] = sha256(out_selection)
        counts['source_catalog_units'] = counts['catalog_units']
        counts['source_catalog_records'] = counts['catalog_records']
        counts['catalog_units'] = len(allowed)
        counts['catalog_records'] = len(rows)
        counts['manifest_sha256'] = sha256(out_manifest)
        counts['source_normalization'] = counts.pop('normalization')
        counts['normalization'] = dict(
            samples_with_annotation_canvas_scaling=sum('annotation_transform' in r['provenance'] for r in rows),
            masks_with_annotation_canvas_scaling=len({x['mask']['path'] for r in rows
                for x in r.get('intervention_regions', []) if 'annotation_transform' in x}),
            areas_recomputed_from_mask=sum(x.get('normalization_note') == 'area_recomputed_from_published_mask'
                for r in rows for x in r['reference']['regions']),
            excluded_invalid_units=[])
        for split in ('selected', 'test', 'development'):
            subset = [r for r in rows if r['unit_id'] in set(selection[f'{split}_unit_ids'])]
            measured = dict(units=len({r['unit_id'] for r in subset}), image_query_records=len(subset),
                distinct_source_images=len({r['source_group_id'] for r in subset}),
                distinct_images=len({r['image']['pixel_sha256'] for r in subset}),
                task_counts=dict(sorted(Counter(r['task_type'] for r in subset if r['is_primary']).items())))
            if split == 'selected' and measured != counts[split]:
                raise ValueError(f'Frozen selected statistics changed: {name}')
            counts[split] = measured
            split_sources.setdefault(split, set()).update(r['source_group_id'] for r in subset)
            split_images.setdefault(split, set()).update(r['image']['pixel_sha256'] for r in subset)
        export_counts[name] = dict(units=len(allowed), records=len(rows),
                                  removed_records=counts['source_catalog_records'] - len(rows))
        print(name, export_counts[name], flush=True)
    if split_sources['test'] & split_sources['development'] or split_images['test'] & split_images['development']:
        raise ValueError('Development/test leakage')
    for relative in sorted(assets):
        src = (source / relative).resolve()
        dst = (destination / relative).resolve()
        if not src.is_relative_to(source) or not dst.is_relative_to(destination):
            raise ValueError(f'Unsafe asset path: {relative}')
        dst.parent.mkdir(parents=True, exist_ok=True)
        shutil.copyfile(src, dst)
    report['status'] = 'complete'
    report['export'] = dict(datasets=export_counts, asset_files=len(assets),
                           asset_bytes=sum((destination / p).stat().st_size for p in assets),
                           selected_unit_ids_unchanged=True, all_selected_units_are_test=True,
                           development_units=0, sample_records_unchanged=True)
    save_json(destination / 'build_report.json', report)
    print(json.dumps(report['export'], ensure_ascii=False, indent=2))


if __name__ == '__main__':
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument('--source', type=Path, default=Path('prepared_data'))
    parser.add_argument('--output', type=Path, required=True)
    args = parser.parse_args()
    export(args.source, args.output)