File size: 14,502 Bytes
3dac39e | 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 | #!/usr/bin/env python3
"""Exact-match span F1 evaluation for Arcspan NER models.
Computes both exact-boundary span F1 (CoNLL/seqeval style) and OPF's native
containment-based span F1, printing them side by side for comparison.
Usage:
python3 scripts/eval_exact_match.py \
--checkpoint checkpoints/r8_5class/ \
--test-data data/processed/aptner_5class_test_clean.jsonl \
--device cuda
"""
from __future__ import annotations
import argparse
import json
import sys
import time
from collections import defaultdict
from pathlib import Path
from typing import Sequence
# ---------------------------------------------------------------------------
# Span type: (label: str, start: int, end: int) — character offsets
# ---------------------------------------------------------------------------
Span = tuple[str, int, int]
def parse_gold_spans(record: dict) -> list[Span]:
"""Extract gold spans from a JSONL record with OPF 'spans' format.
Format: {"spans": {"Label: text": [[start, end], ...], ...}}
"""
spans_field = record.get("spans", {})
if not spans_field:
return []
result: list[Span] = []
for key, offsets in spans_field.items():
# key format: "Label: matched_text"
label = key.split(":")[0].strip()
for offset_pair in offsets:
start, end = int(offset_pair[0]), int(offset_pair[1])
result.append((label, start, end))
return result
def predict_spans(redactor, text: str) -> list[Span]:
"""Run OPF inference and return predicted spans as (label, start, end)."""
result = redactor.redact(text)
spans: list[Span] = []
for det in result.detected_spans:
spans.append((det.label, det.start, det.end))
return spans
# ---------------------------------------------------------------------------
# Exact-match metrics
# ---------------------------------------------------------------------------
def _span_set(spans: list[Span]) -> set[Span]:
return set(spans)
def compute_exact_match_metrics(
all_gold: list[list[Span]],
all_pred: list[list[Span]],
) -> dict:
"""Compute exact-match span-level P/R/F1 (micro + macro + per-class)."""
# Per-class accumulators
class_tp: defaultdict[str, int] = defaultdict(int)
class_fp: defaultdict[str, int] = defaultdict(int)
class_fn: defaultdict[str, int] = defaultdict(int)
for gold_spans, pred_spans in zip(all_gold, all_pred):
gold_set = _span_set(gold_spans)
pred_set = _span_set(pred_spans)
# TP: in both gold and pred (exact label + start + end)
tp_spans = gold_set & pred_set
fp_spans = pred_set - gold_set
fn_spans = gold_set - pred_set
for label, _, _ in tp_spans:
class_tp[label] += 1
for label, _, _ in fp_spans:
class_fp[label] += 1
for label, _, _ in fn_spans:
class_fn[label] += 1
all_labels = sorted(set(class_tp) | set(class_fp) | set(class_fn))
# Per-class metrics
per_class = {}
for label in all_labels:
tp = class_tp[label]
fp = class_fp[label]
fn = class_fn[label]
p = tp / (tp + fp) if (tp + fp) > 0 else 0.0
r = tp / (tp + fn) if (tp + fn) > 0 else 0.0
f1 = 2 * p * r / (p + r) if (p + r) > 0 else 0.0
per_class[label] = {"precision": p, "recall": r, "f1": f1,
"tp": tp, "fp": fp, "fn": fn,
"support": tp + fn}
# Micro-average
total_tp = sum(class_tp.values())
total_fp = sum(class_fp.values())
total_fn = sum(class_fn.values())
micro_p = total_tp / (total_tp + total_fp) if (total_tp + total_fp) > 0 else 0.0
micro_r = total_tp / (total_tp + total_fn) if (total_tp + total_fn) > 0 else 0.0
micro_f1 = 2 * micro_p * micro_r / (micro_p + micro_r) if (micro_p + micro_r) > 0 else 0.0
# Macro-average
if all_labels:
macro_p = sum(per_class[l]["precision"] for l in all_labels) / len(all_labels)
macro_r = sum(per_class[l]["recall"] for l in all_labels) / len(all_labels)
macro_f1 = sum(per_class[l]["f1"] for l in all_labels) / len(all_labels)
else:
macro_p = macro_r = macro_f1 = 0.0
return {
"per_class": per_class,
"micro": {"precision": micro_p, "recall": micro_r, "f1": micro_f1},
"macro": {"precision": macro_p, "recall": macro_r, "f1": macro_f1},
"total_tp": total_tp, "total_fp": total_fp, "total_fn": total_fn,
}
# ---------------------------------------------------------------------------
# Containment-match metrics (OPF style)
# ---------------------------------------------------------------------------
def compute_containment_metrics(
all_gold: list[list[Span]],
all_pred: list[list[Span]],
) -> dict:
"""Compute containment-based span P/R/F1.
Precision: predicted span is TP if it is *contained within* a gold span
with the same label.
Recall: gold span is TP if it is *contained within* a predicted span
with the same label.
"""
class_tp_p: defaultdict[str, int] = defaultdict(int) # for precision
class_fp: defaultdict[str, int] = defaultdict(int)
class_tp_r: defaultdict[str, int] = defaultdict(int) # for recall
class_fn: defaultdict[str, int] = defaultdict(int)
for gold_spans, pred_spans in zip(all_gold, all_pred):
# Precision direction: pred contained in gold
for p_label, p_s, p_e in pred_spans:
matched = False
for g_label, g_s, g_e in gold_spans:
if p_label == g_label and g_s <= p_s and g_e >= p_e:
matched = True
break
if matched:
class_tp_p[p_label] += 1
else:
class_fp[p_label] += 1
# Recall direction: gold contained in pred
for g_label, g_s, g_e in gold_spans:
matched = False
for p_label, p_s, p_e in pred_spans:
if g_label == p_label and p_s <= g_s and p_e >= g_e:
matched = True
break
if matched:
class_tp_r[g_label] += 1
else:
class_fn[g_label] += 1
all_labels = sorted(set(class_tp_p) | set(class_fp) | set(class_tp_r) | set(class_fn))
per_class = {}
for label in all_labels:
tp_p = class_tp_p[label]
fp = class_fp[label]
tp_r = class_tp_r[label]
fn = class_fn[label]
p = tp_p / (tp_p + fp) if (tp_p + fp) > 0 else 0.0
r = tp_r / (tp_r + fn) if (tp_r + fn) > 0 else 0.0
f1 = 2 * p * r / (p + r) if (p + r) > 0 else 0.0
per_class[label] = {"precision": p, "recall": r, "f1": f1,
"support": tp_r + fn}
total_tp_p = sum(class_tp_p.values())
total_fp = sum(class_fp.values())
total_tp_r = sum(class_tp_r.values())
total_fn = sum(class_fn.values())
micro_p = total_tp_p / (total_tp_p + total_fp) if (total_tp_p + total_fp) > 0 else 0.0
micro_r = total_tp_r / (total_tp_r + total_fn) if (total_tp_r + total_fn) > 0 else 0.0
micro_f1 = 2 * micro_p * micro_r / (micro_p + micro_r) if (micro_p + micro_r) > 0 else 0.0
if all_labels:
macro_p = sum(per_class[l]["precision"] for l in all_labels) / len(all_labels)
macro_r = sum(per_class[l]["recall"] for l in all_labels) / len(all_labels)
macro_f1 = sum(per_class[l]["f1"] for l in all_labels) / len(all_labels)
else:
macro_p = macro_r = macro_f1 = 0.0
return {
"per_class": per_class,
"micro": {"precision": micro_p, "recall": micro_r, "f1": micro_f1},
"macro": {"precision": macro_p, "recall": macro_r, "f1": macro_f1},
}
# ---------------------------------------------------------------------------
# Printing
# ---------------------------------------------------------------------------
def print_metrics_table(title: str, metrics: dict, show_counts: bool = False) -> None:
print(f"\n{'=' * 72}")
print(f" {title}")
print(f"{'=' * 72}")
per_class = metrics["per_class"]
if per_class:
if show_counts:
header = f" {'Label':<20s} {'Prec':>7s} {'Rec':>7s} {'F1':>7s} {'TP':>5s} {'FP':>5s} {'FN':>5s} {'Sup':>5s}"
else:
header = f" {'Label':<20s} {'Prec':>7s} {'Rec':>7s} {'F1':>7s} {'Sup':>5s}"
print(header)
print(f" {'-' * (len(header) - 2)}")
for label in sorted(per_class):
m = per_class[label]
if show_counts:
print(f" {label:<20s} {m['precision']:7.4f} {m['recall']:7.4f} {m['f1']:7.4f} "
f"{m.get('tp', '-'):>5} {m.get('fp', '-'):>5} {m.get('fn', '-'):>5} {m['support']:>5}")
else:
print(f" {label:<20s} {m['precision']:7.4f} {m['recall']:7.4f} {m['f1']:7.4f} {m['support']:>5}")
print()
micro = metrics["micro"]
macro = metrics["macro"]
print(f" {'micro-avg':<20s} {micro['precision']:7.4f} {micro['recall']:7.4f} {micro['f1']:7.4f}")
print(f" {'macro-avg':<20s} {macro['precision']:7.4f} {macro['recall']:7.4f} {macro['f1']:7.4f}")
def print_comparison(exact: dict, containment: dict) -> None:
print(f"\n{'=' * 72}")
print(" COMPARISON: Exact-Match vs Containment Span F1")
print(f"{'=' * 72}")
print(f" {'Metric':<25s} {'Exact':>10s} {'Contain':>10s} {'Delta':>10s}")
print(f" {'-' * 55}")
for agg in ["micro", "macro"]:
for m in ["precision", "recall", "f1"]:
e = exact[agg][m]
c = containment[agg][m]
delta = c - e
print(f" {agg + '-' + m:<25s} {e:10.4f} {c:10.4f} {delta:+10.4f}")
# Per-class F1 comparison
all_labels = sorted(set(exact["per_class"]) | set(containment["per_class"]))
if all_labels:
print(f"\n {'Label':<20s} {'Exact-F1':>10s} {'Cont-F1':>10s} {'Delta':>10s}")
print(f" {'-' * 50}")
for label in all_labels:
e_f1 = exact["per_class"].get(label, {}).get("f1", 0.0)
c_f1 = containment["per_class"].get(label, {}).get("f1", 0.0)
print(f" {label:<20s} {e_f1:10.4f} {c_f1:10.4f} {c_f1 - e_f1:+10.4f}")
# ---------------------------------------------------------------------------
# Main
# ---------------------------------------------------------------------------
def main() -> None:
parser = argparse.ArgumentParser(
description="Exact-match span F1 evaluation for Arcspan NER models"
)
parser.add_argument("--checkpoint", required=True, help="Path to OPF checkpoint dir")
parser.add_argument("--test-data", required=True, help="Path to test JSONL file")
parser.add_argument("--device", default="cuda", choices=["cuda", "cpu"])
parser.add_argument("--max-examples", type=int, default=None,
help="Limit number of examples (for quick testing)")
parser.add_argument("--decode-mode", default="viterbi", choices=["viterbi", "argmax"])
parser.add_argument("--json-out", default=None, help="Write metrics JSON to this path")
args = parser.parse_args()
# Load model
print(f"Loading model from {args.checkpoint} on {args.device}...")
from opf import OPF
redactor = OPF(
model=args.checkpoint,
device=args.device,
output_mode="typed",
decode_mode=args.decode_mode,
trim_whitespace=True,
discard_overlapping_predicted_spans=False,
)
if args.decode_mode == "viterbi":
redactor.set_viterbi_decoder()
print("Model loaded.")
# Load test data
test_path = Path(args.test_data)
records: list[dict] = []
with open(test_path, "r", encoding="utf-8") as f:
for line in f:
line = line.strip()
if not line:
continue
records.append(json.loads(line))
if args.max_examples is not None:
records = records[:args.max_examples]
print(f"Loaded {len(records)} test examples from {test_path}")
# Run inference
all_gold: list[list[Span]] = []
all_pred: list[list[Span]] = []
n_gold_spans = 0
n_pred_spans = 0
start_time = time.perf_counter()
for i, record in enumerate(records):
text = record["text"]
gold = parse_gold_spans(record)
pred = predict_spans(redactor, text)
all_gold.append(gold)
all_pred.append(pred)
n_gold_spans += len(gold)
n_pred_spans += len(pred)
if (i + 1) % 100 == 0:
elapsed = time.perf_counter() - start_time
print(f" [{i+1}/{len(records)}] {elapsed:.1f}s "
f"({(i+1)/elapsed:.1f} ex/s)", file=sys.stderr)
elapsed = time.perf_counter() - start_time
print(f"\nInference complete: {len(records)} examples, "
f"{n_gold_spans} gold spans, {n_pred_spans} predicted spans, "
f"{elapsed:.1f}s ({len(records)/elapsed:.1f} ex/s)")
# Compute metrics
exact_metrics = compute_exact_match_metrics(all_gold, all_pred)
containment_metrics = compute_containment_metrics(all_gold, all_pred)
print_metrics_table("EXACT-MATCH Span Metrics (CoNLL/seqeval style)",
exact_metrics, show_counts=True)
print_metrics_table("CONTAINMENT Span Metrics (OPF native style)",
containment_metrics)
print_comparison(exact_metrics, containment_metrics)
# Optional JSON output
if args.json_out:
output = {
"exact_match": {
"micro": exact_metrics["micro"],
"macro": exact_metrics["macro"],
"per_class": exact_metrics["per_class"],
},
"containment": {
"micro": containment_metrics["micro"],
"macro": containment_metrics["macro"],
"per_class": containment_metrics["per_class"],
},
"n_examples": len(records),
"n_gold_spans": n_gold_spans,
"n_pred_spans": n_pred_spans,
"checkpoint": args.checkpoint,
"test_data": args.test_data,
"decode_mode": args.decode_mode,
}
out_path = Path(args.json_out)
out_path.parent.mkdir(parents=True, exist_ok=True)
out_path.write_text(json.dumps(output, indent=2), encoding="utf-8")
print(f"\nMetrics written to {args.json_out}")
print()
if __name__ == "__main__":
main()
|