Datasets:
File size: 19,029 Bytes
0eb47c8 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 | #!/usr/bin/env python3
from __future__ import annotations
import argparse
import importlib.util
import json
import os
import random
import shutil
import time
import urllib.error
import urllib.request
from collections import Counter, defaultdict
from pathlib import Path
from typing import Any, Dict, Iterable, List, Optional, Tuple
REPO = Path(__file__).resolve().parents[1]
QA_SCRIPT = REPO / "scripts" / "run_gpt4o_visual_qa.py"
DIFFICULTIES = ("simple", "medium", "difficult")
SPLITS = ("train", "test")
def load_qa_module():
spec = importlib.util.spec_from_file_location("visual_qa", QA_SCRIPT)
if spec is None or spec.loader is None:
raise RuntimeError(f"Could not load {QA_SCRIPT}")
module = importlib.util.module_from_spec(spec)
spec.loader.exec_module(module)
return module
QA = load_qa_module()
DIFFICULT_EXTRA_PROMPT = """
Additional note for difficult samples:
- This sample is intentionally difficult. Be more tolerant of density, crossings, long routes, small but legible text, and local clutter.
- Still FAIL if the image is genuinely unusable or the graph structure is ambiguous.
- The final selection still requires overall PASS; do not mark a bad image as PASS.
"""
def sample_key(row: Dict[str, Any]) -> str:
value = row.get("unified_sample_id") or row.get("sample_id")
if not value:
raise ValueError(f"Missing sample id in row: {row}")
return str(value)
def read_jsonl(path: Path) -> List[Dict[str, Any]]:
rows: List[Dict[str, Any]] = []
if not path.exists():
return rows
with path.open("r", encoding="utf-8") as handle:
for line in handle:
if line.strip():
rows.append(json.loads(line))
return rows
def write_jsonl(path: Path, rows: Iterable[Dict[str, Any]]) -> None:
path.parent.mkdir(parents=True, exist_ok=True)
with path.open("w", encoding="utf-8") as handle:
for row in rows:
handle.write(json.dumps(row, ensure_ascii=False) + "\n")
def load_manifest(dataset_root: Path, manifest_path: Optional[Path]) -> List[Dict[str, Any]]:
path = manifest_path or dataset_root / "manifest.jsonl"
rows: List[Dict[str, Any]] = []
for row in read_jsonl(path):
row = dict(row)
row["unified_sample_id"] = sample_key(row)
row["image_path"] = str(dataset_root / row["image"])
rows.append(row)
return rows
def allocate_targets(counts: Dict[str, int], total_target: int) -> Dict[str, int]:
total_source = sum(counts.values())
if total_source <= 0:
return {difficulty: 0 for difficulty in DIFFICULTIES}
raw: Dict[str, float] = {
difficulty: total_target * counts.get(difficulty, 0) / total_source
for difficulty in DIFFICULTIES
}
targets = {difficulty: int(raw[difficulty]) for difficulty in DIFFICULTIES}
remaining = total_target - sum(targets.values())
order = sorted(DIFFICULTIES, key=lambda difficulty: (raw[difficulty] - targets[difficulty], difficulty), reverse=True)
for difficulty in order[:remaining]:
targets[difficulty] += 1
return targets
def compute_targets(rows: List[Dict[str, Any]], target_train: int, target_test: int) -> Dict[Tuple[str, str, str], int]:
source = Counter((row["domain"], row["split"], row["difficulty"]) for row in rows)
domains = sorted({str(row["domain"]) for row in rows})
targets: Dict[Tuple[str, str, str], int] = {}
for domain in domains:
for split, total_target in (("train", target_train), ("test", target_test)):
diff_counts = {difficulty: source[(domain, split, difficulty)] for difficulty in DIFFICULTIES}
for difficulty, count in allocate_targets(diff_counts, total_target).items():
targets[(domain, split, difficulty)] = count
return targets
def call_visual_qa(
base_url: str,
api_key: str,
model: str,
image_path: Path,
difficulty: str,
max_tokens: int,
timeout: int,
) -> str:
prompt = QA.VISUAL_QA_PROMPT
if difficulty == "difficult":
prompt += DIFFICULT_EXTRA_PROMPT
payload = {
"model": model,
"messages": [
{"role": "system", "content": "You are a strict but practical visual QA reviewer."},
{
"role": "user",
"content": [
{"type": "text", "text": prompt},
{"type": "image_url", "image_url": {"url": QA.image_data_url(image_path)}},
],
},
],
"temperature": 0,
"max_tokens": max_tokens,
}
request = urllib.request.Request(
base_url.rstrip("/") + "/chat/completions",
data=json.dumps(payload).encode("utf-8"),
headers={"Authorization": f"Bearer {api_key}", "Content-Type": "application/json"},
method="POST",
)
try:
with urllib.request.urlopen(request, timeout=timeout) as response:
body = json.loads(response.read().decode("utf-8"))
except urllib.error.HTTPError as exc:
detail = exc.read().decode("utf-8", errors="replace")
raise RuntimeError(f"HTTP {exc.code}: {detail[:2000]}") from exc
return body["choices"][0]["message"]["content"]
def result_passed(record: Dict[str, Any]) -> bool:
result = record.get("result") if isinstance(record.get("result"), dict) else {}
return record.get("status") == "OK" and result.get("overall") == "PASS"
def link_or_copy(src: Path, dst: Path) -> None:
dst.parent.mkdir(parents=True, exist_ok=True)
if dst.exists():
dst.unlink()
try:
os.link(src, dst)
except OSError:
shutil.copy2(src, dst)
def summarize_progress(
targets: Dict[Tuple[str, str, str], int],
selected: Dict[str, Dict[str, Any]],
reviewed: Dict[str, Dict[str, Any]],
) -> Dict[str, Any]:
selected_counts = Counter()
reviewed_counts = Counter()
failed_counts = Counter()
for row in selected.values():
selected_counts[(row["domain"], row["split"], row["difficulty"])] += 1
for row in reviewed.values():
key = (row.get("domain"), row.get("split"), row.get("difficulty"))
reviewed_counts[key] += 1
if not result_passed(row):
failed_counts[key] += 1
groups: Dict[str, Dict[str, int]] = {}
complete = True
for key, target in sorted(targets.items()):
selected_count = selected_counts[key]
reviewed_count = reviewed_counts[key]
if selected_count < target:
complete = False
groups["/".join(key)] = {
"target": target,
"selected": selected_count,
"reviewed": reviewed_count,
"failed_or_error": failed_counts[key],
"remaining": max(0, target - selected_count),
}
return {
"complete": complete,
"selected_total": len(selected),
"reviewed_total": len(reviewed),
"failed_or_error_total": sum(1 for row in reviewed.values() if not result_passed(row)),
"groups": groups,
}
def write_outputs(
dataset_root: Path,
out_dir: Path,
selected: Dict[str, Dict[str, Any]],
reviewed: Dict[str, Dict[str, Any]],
targets: Dict[Tuple[str, str, str], int],
) -> None:
selected_rows = [row["manifest_row"] for row in sorted(selected.values(), key=lambda item: sample_key(item["manifest_row"]))]
rejected_rows = [row for row in sorted(reviewed.values(), key=sample_key) if not result_passed(row)]
write_jsonl(out_dir / "selected_manifest.jsonl", selected_rows)
write_jsonl(out_dir / "rejected.jsonl", rejected_rows)
(out_dir / "screening_state.json").write_text(
json.dumps(summarize_progress(targets, selected, reviewed), ensure_ascii=False, indent=2),
encoding="utf-8",
)
selected_root = out_dir / "selected_dataset"
for row in selected_rows:
image_rel = Path(row["image"])
ann_rel = Path(row["annotation"])
link_or_copy(dataset_root / image_rel, selected_root / image_rel)
link_or_copy(dataset_root / ann_rel, selected_root / ann_rel)
write_jsonl(selected_root / "manifest.jsonl", selected_rows)
def choose_next_group(
targets: Dict[Tuple[str, str, str], int],
selected_counts: Counter,
available: Dict[Tuple[str, str, str], List[Dict[str, Any]]],
rng: random.Random,
) -> Optional[Tuple[str, str, str]]:
pending = [
key
for key, target in targets.items()
if selected_counts[key] < target and available.get(key)
]
if not pending:
return None
weights = [max(1, targets[key] - selected_counts[key]) for key in pending]
return rng.choices(pending, weights=weights, k=1)[0]
def main() -> int:
parser = argparse.ArgumentParser()
parser.add_argument("--config", type=Path, default=Path("api_config.yaml"))
parser.add_argument("--dataset-root", type=Path, default=Path("outputs/final-before-screen-data"))
parser.add_argument("--manifest", type=Path, default=None)
parser.add_argument("--output-dir", type=Path, default=Path("outputs/final-before-screen-balanced-screening"))
parser.add_argument("--cache-dir", type=Path, default=Path("outputs/gpt4o_visual_qa/final_before_screen_balanced_cache"))
parser.add_argument("--model", default="gpt-4o")
parser.add_argument("--target-train", type=int, default=3000)
parser.add_argument("--target-test", type=int, default=200)
parser.add_argument("--seed", type=int, default=20260610)
parser.add_argument("--max-tokens", type=int, default=700)
parser.add_argument("--timeout", type=int, default=120)
parser.add_argument("--sleep", type=float, default=0.1)
parser.add_argument("--report-every", type=int, default=100)
parser.add_argument("--max-reviews", type=int, default=0, help="Stop after this many new reviews. Use 0 for unlimited.")
parser.add_argument("--max-api-errors", type=int, default=5)
parser.add_argument("--dry-run", action="store_true")
args = parser.parse_args()
rows = load_manifest(args.dataset_root, args.manifest)
targets = compute_targets(rows, args.target_train, args.target_test)
groups: Dict[Tuple[str, str, str], List[Dict[str, Any]]] = defaultdict(list)
for row in rows:
groups[(row["domain"], row["split"], row["difficulty"])].append(row)
rng = random.Random(args.seed)
for key, group_rows in groups.items():
rng.shuffle(group_rows)
args.output_dir.mkdir(parents=True, exist_ok=True)
results_path = args.output_dir / "screening_results.jsonl"
raw_dir = args.output_dir / "raw"
raw_dir.mkdir(parents=True, exist_ok=True)
args.cache_dir.mkdir(parents=True, exist_ok=True)
reviewed: Dict[str, Dict[str, Any]] = {sample_key(row): row for row in read_jsonl(results_path)}
selected: Dict[str, Dict[str, Any]] = {}
for record in reviewed.values():
if result_passed(record):
manifest_row = record.get("manifest_row")
if isinstance(manifest_row, dict):
selected[sample_key(manifest_row)] = record
selected_counts = Counter()
for record in selected.values():
row = record["manifest_row"]
selected_counts[(row["domain"], row["split"], row["difficulty"])] += 1
available: Dict[Tuple[str, str, str], List[Dict[str, Any]]] = {}
for key, group_rows in groups.items():
available[key] = [row for row in group_rows if sample_key(row) not in reviewed]
target_report = {
"/".join(key): {
"target": target,
"source_candidates": len(groups.get(key, [])),
}
for key, target in sorted(targets.items())
}
print(json.dumps({"targets": target_report}, ensure_ascii=False, indent=2), flush=True)
if args.dry_run:
return 0
cfg = QA.parse_simple_yaml(args.config)
api_cfg = cfg.get("api") or cfg
api_key = api_cfg.get("key") or cfg.get("key")
base_url = api_cfg.get("base") or cfg.get("base")
if not api_key or not base_url:
raise ValueError("Could not find api.key and api.base in config")
new_reviews = 0
transient_error_samples: set[str] = set()
consecutive_api_errors = 0
api_error_path = args.output_dir / "api_errors.jsonl"
with results_path.open("a", encoding="utf-8") as out, api_error_path.open("a", encoding="utf-8") as api_error_out:
while True:
filtered_available = {
group_key: [row for row in group_rows if sample_key(row) not in transient_error_samples]
for group_key, group_rows in available.items()
}
key = choose_next_group(targets, selected_counts, filtered_available, rng)
if key is None:
break
domain, split, difficulty = key
while available[key]:
sample = available[key].pop()
if sample_key(sample) not in transient_error_samples:
break
else:
continue
sample_id = sample_key(sample)
image_path = Path(sample["image_path"])
image_hash = QA.image_sha256(image_path)
cache_path = args.cache_dir / f"{sample_id}.json"
record: Dict[str, Any]
used_cache = False
if cache_path.exists():
cached = json.loads(cache_path.read_text(encoding="utf-8"))
if cached.get("image_sha256") == image_hash:
record = cached
used_cache = True
print(f"[cache] {sample_id} {domain}/{split}/{difficulty}", flush=True)
else:
cache_path.unlink()
record = {}
else:
record = {}
if not record:
print(
f"[review] {sample_id} {domain}/{split}/{difficulty} "
f"need={targets[key] - selected_counts[key]}",
flush=True,
)
raw_text = ""
parse_error = None
warnings: List[str] = []
try:
raw_text = call_visual_qa(
str(base_url),
str(api_key),
args.model,
image_path,
difficulty,
args.max_tokens,
args.timeout,
)
(raw_dir / f"{sample_id}.raw.txt").write_text(raw_text, encoding="utf-8")
obj, parse_error = QA.extract_json_object(raw_text)
if obj is None:
status = "PARSE_ERROR"
result = {
"overall": "FAIL",
"checks": {name: "FAIL" for name in QA.CHECK_NAMES},
"failure_reasons": [str(parse_error)],
"quality_score": 1,
}
else:
status = "OK"
result, warnings = QA.normalize_result(obj)
except Exception as exc:
parse_error = f"api_error:{exc}"
error_record = {
"unified_sample_id": sample_id,
"sample_id": sample.get("sample_id", sample_id),
"source_sample_id": sample.get("source_sample_id"),
"domain": domain,
"difficulty": difficulty,
"split": split,
"image": sample.get("image"),
"annotation": sample.get("annotation"),
"image_sha256": image_hash,
"manifest_row": {k: v for k, v in sample.items() if k != "image_path"},
"status": "API_ERROR",
"parse_error": parse_error,
}
api_error_out.write(json.dumps(error_record, ensure_ascii=False) + "\n")
api_error_out.flush()
transient_error_samples.add(sample_id)
consecutive_api_errors += 1
print(f" API_ERROR {parse_error}", flush=True)
if consecutive_api_errors >= args.max_api_errors:
raise RuntimeError(f"Stopped after {consecutive_api_errors} consecutive API errors")
continue
record = {
"unified_sample_id": sample_id,
"sample_id": sample.get("sample_id", sample_id),
"source_sample_id": sample.get("source_sample_id"),
"domain": domain,
"difficulty": difficulty,
"split": split,
"image": sample.get("image"),
"annotation": sample.get("annotation"),
"image_sha256": image_hash,
"manifest_row": {k: v for k, v in sample.items() if k != "image_path"},
"result": result,
"status": status,
"parse_error": parse_error,
"normalization_warnings": warnings,
}
cache_path.write_text(json.dumps(record, ensure_ascii=False, indent=2), encoding="utf-8")
new_reviews += 1
consecutive_api_errors = 0
reviewed[sample_id] = record
out.write(json.dumps(record, ensure_ascii=False) + "\n")
out.flush()
if result_passed(record):
selected[sample_id] = record
selected_counts[key] += 1
print(f" PASS selected={selected_counts[key]}/{targets[key]}", flush=True)
else:
result = record.get("result") or {}
reasons = "; ".join(result.get("failure_reasons") or []) or record.get("parse_error") or "-"
print(f" FAIL/ERROR reasons={reasons}", flush=True)
if args.report_every and len(reviewed) % args.report_every == 0:
summary = summarize_progress(targets, selected, reviewed)
print(json.dumps({k: summary[k] for k in ("selected_total", "reviewed_total", "failed_or_error_total", "complete")}, ensure_ascii=False), flush=True)
write_outputs(args.dataset_root, args.output_dir, selected, reviewed, targets)
if args.max_reviews and new_reviews >= args.max_reviews:
break
if args.sleep and not used_cache:
time.sleep(args.sleep)
write_outputs(args.dataset_root, args.output_dir, selected, reviewed, targets)
summary = summarize_progress(targets, selected, reviewed)
print(json.dumps(summary, ensure_ascii=False, indent=2), flush=True)
return 0 if summary["complete"] else 2
if __name__ == "__main__":
raise SystemExit(main())
|