File size: 19,113 Bytes
d82db76 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 | """Error analysis utilities for benchmark reports."""
from __future__ import annotations
import json
import math
from dataclasses import dataclass
from pathlib import Path
from typing import Any
from .benchmark import _load_benchmark_rows
from .errors import BenchmarkError
from .manifest import ManifestEntry, load_manifest, write_manifest
from .validation import metric_direction, sample_key
@dataclass(slots=True)
class BenchmarkErrorAnalysis:
benchmark_report: str
system: str
metric: str
direction: str
baselines: list[str]
sample_count: int
top_failures_path: str
analysis_path: str
def to_dict(self) -> dict[str, object]:
return {
"benchmark_report": self.benchmark_report,
"system": self.system,
"metric": self.metric,
"direction": self.direction,
"baselines": self.baselines,
"sample_count": self.sample_count,
"top_failures_path": self.top_failures_path,
"analysis_path": self.analysis_path,
}
@dataclass(slots=True)
class FailureManifestExport:
source_manifest: str
top_failures_path: str
output_manifest: str
selected_count: int
split: str | None
summary_path: str
def to_dict(self) -> dict[str, object]:
return {
"source_manifest": self.source_manifest,
"top_failures_path": self.top_failures_path,
"output_manifest": self.output_manifest,
"selected_count": self.selected_count,
"split": self.split,
"summary_path": self.summary_path,
}
def analyze_benchmark_errors(
report_path: str | Path,
output_dir: str | Path,
*,
system: str,
baselines: list[str] | None = None,
metric: str = "cer",
direction: str | None = None,
top_n: int = 20,
) -> BenchmarkErrorAnalysis:
if top_n < 1:
raise BenchmarkError("top_n must be at least 1")
rows = _load_benchmark_rows(report_path)
if not rows:
raise BenchmarkError("benchmark report has no rows")
target_direction = direction or metric_direction(metric)
available_systems = sorted({str(row.get("baseline") or "") for row in rows if str(row.get("baseline") or "")})
if system not in available_systems:
raise BenchmarkError(f"system {system!r} is not present in benchmark report")
compare_baselines = baselines or []
missing = [name for name in compare_baselines if name not in available_systems]
if missing:
raise BenchmarkError(f"requested baselines missing from benchmark report: {', '.join(missing)}")
rows_by_system = _metric_rows_by_system(rows, metric)
system_rows = rows_by_system.get(system, {})
if not system_rows:
raise BenchmarkError(f"system {system!r} has no ok rows with metric {metric!r}")
out = Path(output_dir)
out.mkdir(parents=True, exist_ok=True)
top_failures = _top_failures(system_rows, rows_by_system, compare_baselines, metric, target_direction, top_n)
slice_summary = _slice_summary(system_rows, rows_by_system, compare_baselines, metric, target_direction)
baseline_summary = _baseline_summary(system_rows, rows_by_system, compare_baselines, metric, target_direction)
top_failure_slice_counts = _top_failure_slice_counts(top_failures)
top_failures_path = out / "top-failures.json"
top_failures_path.write_text(json.dumps(top_failures, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
payload = {
"benchmark_report": str(Path(report_path)),
"system": system,
"metric": metric,
"direction": target_direction,
"baselines": compare_baselines,
"sample_count": len(system_rows),
"baseline_summary": baseline_summary,
"slice_summary": slice_summary,
"top_failure_slice_counts": top_failure_slice_counts,
"top_failures_path": str(top_failures_path),
}
analysis_path = out / "error-analysis.json"
analysis_path.write_text(json.dumps(payload, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
_write_markdown(out / "error-analysis.md", payload, top_failures)
return BenchmarkErrorAnalysis(
benchmark_report=str(Path(report_path)),
system=system,
metric=metric,
direction=target_direction,
baselines=compare_baselines,
sample_count=len(system_rows),
top_failures_path=str(top_failures_path),
analysis_path=str(analysis_path),
)
def export_failure_manifest(
source_manifest: str | Path,
top_failures_path: str | Path,
output_manifest: str | Path,
*,
split: str | None = None,
) -> FailureManifestExport:
entries = load_manifest(source_manifest)
entries_by_sample = {entry.sample_id: entry for entry in entries}
try:
failure_rows = json.loads(Path(top_failures_path).read_text(encoding="utf-8"))
except json.JSONDecodeError as exc:
raise BenchmarkError(f"invalid top failures JSON {top_failures_path}: {exc}") from exc
if not isinstance(failure_rows, list):
raise BenchmarkError("top failures file must contain a JSON list")
selected: list[ManifestEntry] = []
missing: list[str] = []
seen: set[str] = set()
for rank, item in enumerate(failure_rows, start=1):
if not isinstance(item, dict):
raise BenchmarkError(f"top failure row {rank} must be an object")
sample_id = str(item.get("sample_id") or item.get("sample_key") or "")
if not sample_id:
raise BenchmarkError(f"top failure row {rank} missing sample_id/sample_key")
if sample_id in seen:
continue
seen.add(sample_id)
source = entries_by_sample.get(sample_id)
if source is None:
missing.append(sample_id)
continue
selected.append(_failure_entry(source, failure=item, rank=rank, split=split))
if missing:
raise BenchmarkError(f"top failure samples missing from source manifest: {', '.join(missing)}")
if not selected:
raise BenchmarkError("failure manifest export selected zero samples")
out_path = Path(output_manifest)
write_manifest(selected, out_path)
summary_path = out_path.with_name(f"{out_path.stem}-failure-manifest.json")
export = FailureManifestExport(
source_manifest=str(Path(source_manifest)),
top_failures_path=str(Path(top_failures_path)),
output_manifest=str(out_path),
selected_count=len(selected),
split=split,
summary_path=str(summary_path),
)
summary_path.write_text(json.dumps(export.to_dict(), ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
_write_failure_manifest_markdown(out_path.with_name(f"{out_path.stem}-failure-manifest.md"), export, selected)
return export
def _metric_rows_by_system(rows: list[dict[str, Any]], metric: str) -> dict[str, dict[str, dict[str, Any]]]:
by_system: dict[str, dict[str, dict[str, Any]]] = {}
for row in rows:
if row.get("status") != "ok":
continue
baseline = str(row.get("baseline") or "")
if not baseline:
continue
metrics = row.get("metrics")
if not isinstance(metrics, dict):
continue
value = metrics.get(metric)
if not isinstance(value, int | float) or not math.isfinite(float(value)):
continue
by_system.setdefault(baseline, {})[sample_key(row)] = row
return by_system
def _failure_entry(source: ManifestEntry, *, failure: dict[str, Any], rank: int, split: str | None) -> ManifestEntry:
metadata = dict(source.metadata or {})
if split:
existing_slices = metadata.get("slices")
slice_values = list(existing_slices) if isinstance(existing_slices, list) else []
if split not in {str(item) for item in slice_values}:
slice_values.append(split)
metadata["slices"] = slice_values
metric_name, metric_value = _failure_metric(failure)
metadata["failure_analysis"] = {
"rank": rank,
"metric": metric_name,
"metric_value": metric_value,
"worst_gap": failure.get("worst_gap"),
"comparisons": failure.get("comparisons") if isinstance(failure.get("comparisons"), dict) else {},
}
return ManifestEntry(
sample_id=source.sample_id,
dataset=source.dataset,
split=split or source.split,
image_path=source.image_path,
text=source.text,
sha256=source.sha256,
metadata=metadata,
)
def _failure_metric(failure: dict[str, Any]) -> tuple[str, float]:
for key, value in failure.items():
if key in {"badness", "worst_gap"}:
continue
if isinstance(value, int | float) and math.isfinite(float(value)):
return key, float(value)
raise BenchmarkError("top failure row has no finite metric value")
def _write_failure_manifest_markdown(path: Path, export: FailureManifestExport, selected: list[ManifestEntry]) -> None:
lines = [
"# Failure Manifest Export",
"",
f"Source manifest: `{export.source_manifest}`",
f"Top failures: `{export.top_failures_path}`",
f"Output manifest: `{export.output_manifest}`",
f"Selected samples: `{export.selected_count}`",
f"Split override: `{export.split or ''}`",
"",
"| rank | sample_id | metric | value | image_path |",
"| ---: | --- | --- | ---: | --- |",
]
for entry in selected:
failure = (entry.metadata or {}).get("failure_analysis") or {}
lines.append(
"| {rank} | {sample_id} | {metric} | {value:.6f} | {image_path} |".format(
rank=int(failure.get("rank") or 0),
sample_id=entry.sample_id,
metric=str(failure.get("metric") or ""),
value=float(failure.get("metric_value") or 0.0),
image_path=entry.image_path,
)
)
path.write_text("\n".join(lines) + "\n", encoding="utf-8")
def _top_failures(
system_rows: dict[str, dict[str, Any]],
rows_by_system: dict[str, dict[str, dict[str, Any]]],
baselines: list[str],
metric: str,
direction: str,
top_n: int,
) -> list[dict[str, Any]]:
ranked = [_failure_record(sample, row, rows_by_system, baselines, metric, direction) for sample, row in system_rows.items()]
ranked.sort(key=_failure_sort_key, reverse=True)
return ranked[:top_n]
def _top_failure_slice_counts(top_failures: list[dict[str, Any]]) -> list[dict[str, object]]:
counts: dict[str, int] = {}
for item in top_failures:
slices = item.get("slices")
if not isinstance(slices, list):
continue
for slice_name in slices:
text = str(slice_name)
if text:
counts[text] = counts.get(text, 0) + 1
return [{"slice": slice_name, "count": count} for slice_name, count in sorted(counts.items(), key=lambda pair: (-pair[1], pair[0]))]
def _failure_sort_key(item: dict[str, Any]) -> tuple[float, float, float]:
worst_gap = item.get("worst_gap")
badness = float(item["badness"])
if isinstance(worst_gap, int | float) and math.isfinite(float(worst_gap)):
gap = float(worst_gap)
return (1.0 if gap > 0 else 0.0, gap if gap > 0 else badness, badness)
return (0.0, badness, badness)
def _failure_record(
sample: str,
row: dict[str, Any],
rows_by_system: dict[str, dict[str, dict[str, Any]]],
baselines: list[str],
metric: str,
direction: str,
) -> dict[str, Any]:
metrics = row.get("metrics") or {}
value = float(metrics[metric])
comparisons: dict[str, dict[str, float]] = {}
gaps: list[float] = []
for baseline in baselines:
baseline_row = rows_by_system.get(baseline, {}).get(sample)
if baseline_row is None:
continue
baseline_value = float((baseline_row.get("metrics") or {})[metric])
gap = _worse_than_baseline(value, baseline_value, direction)
gaps.append(gap)
comparisons[baseline] = {
"baseline_value": baseline_value,
"worse_than_baseline": gap,
}
return {
"sample_key": sample,
"sample_id": row.get("sample_id"),
"input_path": row.get("input_path"),
"slices": sorted({str(item) for item in row.get("slices") or []}),
metric: value,
"badness": _badness(value, direction),
"worst_gap": max(gaps) if gaps else None,
"comparisons": comparisons,
}
def _slice_summary(
system_rows: dict[str, dict[str, Any]],
rows_by_system: dict[str, dict[str, dict[str, Any]]],
baselines: list[str],
metric: str,
direction: str,
) -> list[dict[str, Any]]:
slices = sorted({slice_name for row in system_rows.values() for slice_name in _row_slices(row)})
summaries = []
for slice_name in slices:
keys = sorted(sample for sample, row in system_rows.items() if slice_name in _row_slices(row))
system_values = [float((system_rows[key].get("metrics") or {})[metric]) for key in keys]
item: dict[str, Any] = {
"slice": slice_name,
"sample_count": len(system_values),
"system_mean": _mean(system_values),
"system_max": max(system_values) if system_values else None,
}
baseline_items = {}
for baseline in baselines:
shared = [key for key in keys if key in rows_by_system.get(baseline, {})]
if not shared:
continue
baseline_values = [float((rows_by_system[baseline][key].get("metrics") or {})[metric]) for key in shared]
shared_system_values = [float((system_rows[key].get("metrics") or {})[metric]) for key in shared]
baseline_items[baseline] = {
"pairs": len(shared),
"baseline_mean": _mean(baseline_values),
"mean_worse_than_baseline": _worse_than_baseline(_mean(shared_system_values), _mean(baseline_values), direction),
}
item["baselines"] = baseline_items
summaries.append(item)
summaries.sort(key=lambda item: _badness(float(item["system_mean"]), direction), reverse=True)
return summaries
def _baseline_summary(
system_rows: dict[str, dict[str, Any]],
rows_by_system: dict[str, dict[str, dict[str, Any]]],
baselines: list[str],
metric: str,
direction: str,
) -> dict[str, dict[str, float | int]]:
summary: dict[str, dict[str, float | int]] = {}
for baseline in baselines:
shared = sorted(set(system_rows) & set(rows_by_system.get(baseline, {})))
if not shared:
continue
system_values = [float((system_rows[key].get("metrics") or {})[metric]) for key in shared]
baseline_values = [float((rows_by_system[baseline][key].get("metrics") or {})[metric]) for key in shared]
gaps = [_worse_than_baseline(system_value, baseline_value, direction) for system_value, baseline_value in zip(system_values, baseline_values, strict=True)]
summary[baseline] = {
"pairs": len(shared),
"system_mean": _mean(system_values),
"baseline_mean": _mean(baseline_values),
"mean_worse_than_baseline": _mean(gaps),
"system_worse_count": sum(1 for gap in gaps if gap > 0),
"system_better_count": sum(1 for gap in gaps if gap < 0),
"tie_count": sum(1 for gap in gaps if gap == 0),
}
return summary
def _row_slices(row: dict[str, Any]) -> set[str]:
values = {str(item) for item in row.get("slices") or [] if str(item)}
values.add("all")
return values
def _badness(value: float, direction: str) -> float:
if direction == "lower":
return value
return -value
def _worse_than_baseline(system_value: float, baseline_value: float, direction: str) -> float:
if direction == "lower":
return system_value - baseline_value
return baseline_value - system_value
def _mean(values: list[float]) -> float:
if not values:
raise BenchmarkError("cannot average an empty metric series")
return sum(values) / len(values)
def _write_markdown(path: Path, payload: dict[str, Any], top_failures: list[dict[str, Any]]) -> None:
metric = str(payload["metric"])
lines = [
"# Benchmark Error Analysis",
"",
f"Report: `{payload['benchmark_report']}`",
f"System: `{payload['system']}`",
f"Metric: `{metric}`",
f"Direction: `{payload['direction']}`",
f"Samples: `{payload['sample_count']}`",
"",
"## Baseline Summary",
"",
"| baseline | pairs | system_mean | baseline_mean | mean_worse_than_baseline | system_worse | system_better | ties |",
"| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |",
]
baseline_summary = payload.get("baseline_summary")
if isinstance(baseline_summary, dict) and baseline_summary:
for baseline, item in sorted(baseline_summary.items()):
lines.append(
"| {baseline} | {pairs} | {system_mean:.6f} | {baseline_mean:.6f} | {gap:.6f} | {worse} | {better} | {ties} |".format(
baseline=baseline,
pairs=int(item["pairs"]),
system_mean=float(item["system_mean"]),
baseline_mean=float(item["baseline_mean"]),
gap=float(item["mean_worse_than_baseline"]),
worse=int(item["system_worse_count"]),
better=int(item["system_better_count"]),
ties=int(item["tie_count"]),
)
)
else:
lines.append("| none | 0 | | | | | | |")
lines.extend(
[
"",
"## Hardest Samples",
"",
f"| sample | slices | {metric} | worst_gap | input |",
"| --- | --- | ---: | ---: | --- |",
]
)
for item in top_failures:
gap = item["worst_gap"]
lines.append(
"| {sample} | {slices} | {value:.6f} | {gap} | {input_path} |".format(
sample=item.get("sample_id") or item["sample_key"],
slices=",".join(item.get("slices") or []),
value=float(item[metric]),
gap=f"{float(gap):.6f}" if gap is not None else "",
input_path=item.get("input_path") or "",
)
)
top_failure_counts = payload.get("top_failure_slice_counts")
lines.extend(["", "## Top Failure Slice Counts", "", "| slice | count |", "| --- | ---: |"])
if isinstance(top_failure_counts, list) and top_failure_counts:
for item in top_failure_counts:
if isinstance(item, dict):
lines.append(f"| {item.get('slice', '')} | {int(item.get('count') or 0)} |")
else:
lines.append("| none | 0 |")
lines.extend(["", f"Top failures JSON: `{payload['top_failures_path']}`"])
path.write_text("\n".join(lines) + "\n", encoding="utf-8")
|