Datasets:
File size: 14,843 Bytes
7825400 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 | """Build dataset evaluation from explicitly pinned release inputs."""
from __future__ import annotations
from utils.paths import PROJECT_ROOT
from utils.io import sha256_file as _sha256
import json
from collections import Counter
from datetime import datetime, timezone
from pathlib import Path
from utils.dataset import resolve_local_images
from statistics import mean, median
from scripts.datasets.dataset_quality import build_quality_slices
from utils.release import release_config, release_identity, release_paths, validate_population
ROOT = PROJECT_ROOT
FULL_DATASET = ROOT / "data" / "full_dataset"
MANIFEST_PATH = FULL_DATASET / "dataset_manifest.jsonl"
GOLD_DIR = FULL_DATASET / "ground_truth" / "gold"
IAA_DIR = FULL_DATASET / "ground_truth" / "iaa"
IMAGE_DIR = FULL_DATASET / "images"
DOMAIN_PATH = ROOT / "data" / "violation_domains.json"
DEFAULT_OUTPUT = ROOT / "frontend" / ".generated" / "dataset_eda.json"
def _latest(directory: Path, pattern: str) -> Path:
candidates = sorted(directory.glob(pattern))
if not candidates:
raise FileNotFoundError(f"No {pattern} exports found in {directory}")
return candidates[-1]
def _source(path: Path, records: int) -> dict:
return {"filename": path.name, "records": records, "sha256": _sha256(path)}
def _valid_bbox(bbox: object) -> bool:
if not isinstance(bbox, list) or len(bbox) != 4 or not all(isinstance(value, (int, float)) for value in bbox):
return False
x, y, width, height = bbox
return width > 0 and height > 0 and x >= 0 and y >= 0 and x + width <= 1 and y + height <= 1
def _rating(values: list[int | None]) -> dict:
clean = [value for value in values if value is not None]
counts = Counter(clean)
return {
"eligible": len(clean),
"missing": len(values) - len(clean),
"counts": {str(score): counts[score] for score in range(1, 6)},
"mean": round(mean(clean), 3) if clean else None,
"median": median(clean) if clean else None,
"low": sum(value <= 2 for value in clean),
"high": sum(value >= 4 for value in clean),
}
def _distribution(values: list[str], order: list[str] | None = None) -> list[dict]:
counts = Counter(values)
keys = order or sorted(counts)
return [{"name": key, "count": counts[key]} for key in keys]
def summarize_gold(records: list[dict], domains: list[dict]) -> dict:
domain_ids = {item["name"]: set(item["ids"]) for item in domains}
image_classes = [set(item["use_case_id"] for item in row["violations"]) for row in records]
instance_counts = Counter(
item["use_case_id"] for row in records for item in row["violations"]
)
image_counts = Counter(use_case_id for values in image_classes for use_case_id in values)
class_names = {}
for row in records:
for item in row["violations"]:
class_names.setdefault(item["use_case_id"], item["name"])
violations = []
for domain in domains:
for use_case_id in domain["ids"]:
violations.append({
"id": use_case_id,
"name": class_names.get(use_case_id, f"Violation {use_case_id}"),
"domain": domain["name"],
"images": image_counts[use_case_id],
"instances": instance_counts[use_case_id],
"represented": image_counts[use_case_id] > 0,
})
domain_rows = []
for name, ids in domain_ids.items():
domain_rows.append({
"name": name,
"class_count": len(ids),
"represented_classes": sum(image_counts[use_case_id] > 0 for use_case_id in ids),
"images": sum(bool(values & ids) for values in image_classes),
"instances": sum(instance_counts[use_case_id] for use_case_id in ids),
})
return {
"records": len(records),
"categories": _distribution([row["category"] for row in records]),
"gold_kinds": _distribution([row["kind"] for row in records]),
"gold_methods": _distribution([row.get("method", row["kind"]) for row in records]),
"label_states": [
{"name": "Violation present", "count": sum(bool(row["violations"]) for row in records)},
{"name": "No violation", "count": sum(row["no_violation_detected"] for row in records)},
{"name": "Reported / unusable", "count": sum(row["reported"] for row in records)},
],
"plausibility": _rating([row["plausibility"] for row in records]),
"naturalness": _rating([row["naturalness"] for row in records]),
"domains": domain_rows,
"violations": violations,
"violation_instances": sum(instance_counts.values()),
"represented_classes": sum(item["represented"] for item in violations),
}
def _agreement_summary(records: list[dict]) -> dict:
def values(key: str) -> list[float]:
return [row["agreement"][key] for row in records if row["agreement"].get(key) is not None]
plausibility = values("plausibility_delta")
naturalness = values("naturalness_delta")
jaccard = values("violation_jaccard")
bbox = values("mean_bbox_iou")
scores = values("disagreement_score")
return {
"records": len(records),
"first_pass_submissions": sum(row["n_first_pass"] for row in records),
"plausibility": {
"eligible": len(plausibility), "exact": sum(value == 0 for value in plausibility),
"within_one": sum(value <= 1 for value in plausibility),
"counts": {str(delta): sum(value == delta for value in plausibility) for delta in range(5)},
},
"naturalness": {
"eligible": len(naturalness), "exact": sum(value == 0 for value in naturalness),
"within_one": sum(value <= 1 for value in naturalness),
"counts": {str(delta): sum(value == delta for value in naturalness) for delta in range(5)},
},
"violation_jaccard": {
"eligible": len(jaccard), "exact": sum(value == 1 for value in jaccard),
"none": sum(value == 0 for value in jaccard),
"partial": sum(0 < value < 1 for value in jaccard),
"mean": round(mean(jaccard), 3) if jaccard else None,
},
"bbox_iou": {
"eligible": len(bbox), "mean": round(mean(bbox), 3) if bbox else None,
"median": round(median(bbox), 3) if bbox else None,
"at_least_50": sum(value >= 0.50 for value in bbox),
"at_least_75": sum(value >= 0.75 for value in bbox),
},
"reported_mismatches": sum(bool(row["agreement"].get("reported_mismatch")) for row in records),
"disagreement_score": {
"eligible": len(scores), "mean": round(mean(scores), 3) if scores else None,
"median": round(median(scores), 3) if scores else None,
"maximum": max(scores) if scores else None,
},
}
def _timeline(iaa_records: list[dict], gold_records: list[dict]) -> list[dict]:
iaa_dates = Counter(row["completed_at"] for row in iaa_records if row["completed_at"])
gold_dates = Counter(row["adjudicated_at"] for row in gold_records if row["adjudicated_at"])
dates = sorted(set(iaa_dates) | set(gold_dates))
rows, iaa_total, gold_total = [], 0, 0
for date in dates:
iaa_total += iaa_dates[date]
gold_total += gold_dates[date]
rows.append({"date": date, "iaa": iaa_total, "gold": gold_total})
return rows
def build_dataset_eda(output: Path = DEFAULT_OUTPUT, *, gold_path: Path | None = None,
iaa_path: Path | None = None, manifest_path: Path | None = None) -> dict:
pinned = gold_path is None and iaa_path is None and manifest_path is None
paths = release_paths(verify=pinned)
gold_path = gold_path or paths["gold"]
iaa_path = iaa_path or paths["iaa"]
manifest_path = manifest_path or paths["manifest"]
manifest = [json.loads(line) for line in manifest_path.open(encoding="utf-8")]
if len({row["image_id"] for row in manifest}) != len(manifest):
raise ValueError("Full-dataset manifest contains duplicate image IDs")
manifest_by_id = {row["image_id"]: row for row in manifest}
domains_payload = json.loads(DOMAIN_PATH.read_text(encoding="utf-8"))
domains = [{"name": name, "ids": ids} for name, ids in domains_payload.items()]
domain_by_id = {use_case_id: name for name, ids in domains_payload.items() for use_case_id in ids}
local_images = resolve_local_images(manifest, IMAGE_DIR)
gold_source = json.loads(gold_path.read_text(encoding="utf-8"))
if not isinstance(gold_source, list) or len({row["image_id"] for row in gold_source}) != len(gold_source):
raise ValueError("Gold export must be an array with unique image IDs")
invalid_gold_boxes = 0
gold_records = []
for record in gold_source:
gold = record["gold"]
for item in gold.get("violations", []):
invalid_gold_boxes += int(not _valid_bbox(item.get("bbox")))
gold_records.append({
"image_id": record["image_id"],
"category": manifest_by_id.get(record["image_id"], {}).get("category", "unmapped"),
"kind": record["kind"],
"method": record.get("edit_summary", {}).get("resolution", {}).get("method_id", record["kind"]),
"adjudicated_at": (record.get("adjudicated_at") or "")[:10] or None,
"plausibility": gold.get("plausibility"),
"naturalness": gold.get("naturalness"),
"no_violation_detected": bool(gold.get("no_violation_detected")),
"reported": bool(gold.get("reported", False)),
"violations": [{
"use_case_id": item["use_case_id"],
"name": item.get("use_case_name", f"Violation {item['use_case_id']}"),
"domain": domain_by_id.get(item["use_case_id"], "Unmapped"),
"bbox": item.get("bbox") if _valid_bbox(item.get("bbox")) else None,
} for item in gold.get("violations", [])],
})
raw_iaa = json.loads(iaa_path.read_text(encoding="utf-8"))
if pinned:
validate_population(gold_source, raw_iaa, manifest, release_config()["expected"])
if not isinstance(raw_iaa, list) or len({row["image_id"] for row in raw_iaa}) != len(raw_iaa):
raise ValueError("IAA export must be an array with unique image IDs")
iaa_records = []
invalid_iaa_boxes = 0
for row in raw_iaa:
labels = []
completed = []
for index, label in enumerate(row.get("first_pass", [])):
submitted_at = label.get("submitted_at")
if submitted_at:
completed.append(submitted_at[:10])
violations = []
for item in label.get("violations", []):
if not _valid_bbox(item.get("bbox")):
invalid_iaa_boxes += 1
continue
violations.append({
"use_case_id": item["use_case_id"],
"name": item.get("use_case_name", f"Violation {item['use_case_id']}"),
"bbox": item["bbox"],
})
labels.append({
"label": f"Label {chr(65 + index)}",
"plausibility": label.get("plausibility"),
"naturalness": label.get("naturalness"),
"no_violation_detected": bool(label.get("no_violation_detected")),
"reported": bool(label.get("reported", False)),
"violations": violations,
})
agreement = {
key: row.get("agreement", {}).get(key)
for key in ("plausibility_delta", "naturalness_delta", "violation_jaccard", "mean_bbox_iou", "reported_mismatch", "disagreement_score")
}
iaa_records.append({
"image_id": row["image_id"],
"category": manifest_by_id.get(row["image_id"], {}).get("category", "unmapped"),
"l2_code": row.get("l2_code"),
"n_first_pass": len(labels),
"completed_at": max(completed) if completed else None,
"agreement": agreement,
"labels": labels,
"image_available": row["image_id"] in local_images,
})
manifest_categories = _distribution([row.get("category", "unmapped") for row in manifest])
artifact = {
"schema_version": 1,
"generated_at": datetime.now(timezone.utc).isoformat(),
"release": release_identity() if pinned else None,
"sources": {
"manifest": _source(manifest_path, len(manifest)),
"gold": _source(gold_path, len(gold_records)),
"iaa": _source(iaa_path, len(iaa_records)),
},
"progress": {
"target_images": len(manifest),
"iaa_ready_images": len(iaa_records),
"gold_ready_images": len(gold_records),
"local_images": len(local_images),
"gold_in_iaa": len({row["image_id"] for row in gold_records} & {row["image_id"] for row in iaa_records}),
"timeline": _timeline(iaa_records, gold_records),
},
"manifest_categories": manifest_categories,
"domains": domains,
"gold_summary": summarize_gold(gold_records, domains),
"iaa_summary": _agreement_summary(iaa_records),
"quality_audit": {
"invalid_gold_boxes": invalid_gold_boxes,
"invalid_iaa_boxes": invalid_iaa_boxes,
"gold_missing_manifest": sum(row["category"] == "unmapped" for row in gold_records),
"iaa_missing_manifest": sum(row["category"] == "unmapped" for row in iaa_records),
},
"filter_options": {
"categories": [row["name"] for row in manifest_categories if row["count"]],
"gold_kinds": sorted({row["kind"] for row in gold_records}),
},
"gold_records": gold_records,
"iaa_records": iaa_records,
}
output.parent.mkdir(parents=True, exist_ok=True)
artifact["quality_slices"] = build_quality_slices(gold_records, domains, artifact["gold_summary"]["violations"])
output.write_text(json.dumps(artifact, indent=2), encoding="utf-8")
return artifact
def main() -> None:
import argparse
parser = argparse.ArgumentParser()
parser.add_argument("--output", type=Path, default=DEFAULT_OUTPUT)
parser.add_argument("--gold", type=Path)
parser.add_argument("--iaa", type=Path)
parser.add_argument("--manifest", type=Path)
args = parser.parse_args()
artifact = build_dataset_eda(args.output, gold_path=args.gold, iaa_path=args.iaa, manifest_path=args.manifest)
print(
f"Wrote {args.output} from {artifact['sources']['gold']['filename']} and "
f"{artifact['sources']['iaa']['filename']}"
)
if __name__ == "__main__":
main()
|