File size: 15,217 Bytes
d868d2e |
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 |
import logging
import warnings
from collections import defaultdict
from functools import partial
from typing import Callable, Iterable, List, Optional, Set, Tuple
import numpy as np
import pandas as pd
from pytorch_ie import DocumentMetric
from pytorch_ie.annotations import BinaryRelation
from sklearn.metrics import average_precision_score, ndcg_score
logger = logging.getLogger(__name__)
NEG_INF = -1e9 # smaller than any real score
# metrics
def true_mrr(y_true: np.ndarray, y_score: np.ndarray, k: int | None = None) -> float:
"""
Macro MRR over *all* queries.
β’ Reciprocal rank is 0 when a query has no relevant item.
β’ If k is given, restrict the search to the top-k list.
"""
if y_true.size == 0:
return np.nan
rr = []
for t, s in zip(y_true, y_score):
if t.sum() == 0:
rr.append(0.0)
continue
order = np.argsort(-s)
if k is not None:
order = order[:k]
# first position where t == 1, +1 for 1-based rank
first_hit = np.flatnonzero(t[order] > 0)
rank = first_hit[0] + 1 if first_hit.size else np.inf
rr.append(0.0 if np.isinf(rank) else 1.0 / rank)
return np.mean(rr)
def macro_ndcg(y_true: np.ndarray, y_score: np.ndarray, k: int | None = None) -> float:
"""
Macro NDCG@k over all queries.
ndcg_score returns 0 when a query has no positives, so no masking is required.
"""
if y_true.size == 0:
return np.nan
return ndcg_score(y_true, y_score, k=k)
def macro_map(y_true: np.ndarray, y_score: np.ndarray) -> float:
"""
Macro MAP: mean of Average-Precision per query.
Queries without positives contribute AP = 0.
"""
if y_true.size == 0:
return np.nan
ap = []
for t, s in zip(y_true, y_score):
if t.sum() == 0:
ap.append(0.0)
else:
ap.append(average_precision_score(t, s))
return np.mean(ap)
def ap_micro(y_true: np.ndarray, y_score: np.ndarray) -> float:
"""
Micro AP over the entire pool (unchanged).
"""
with warnings.catch_warnings():
warnings.filterwarnings("ignore", message="No positive class found in y_true")
return average_precision_score(y_true.ravel(), y_score.ravel())
# ---------------------------
# Recall@k
# ---------------------------
def recall_at_k_micro(y_true: np.ndarray, y_score: np.ndarray, k: int = 5) -> float:
"""
Micro Recall@k (a.k.a. instance-level recall)
β Each *positive instance* counts once, regardless of which query it belongs to.
β Denominator = total #positives across the whole pool.
"""
total_pos = y_true.sum()
if total_pos == 0:
return np.nan
topk = np.argsort(-y_score, axis=1)[:, :k] # indices of top-k per query
rows = np.arange(topk.shape[0])[:, None]
hits = (y_true[rows, topk] > 0).sum() # total #hits (instances)
return hits / total_pos
def recall_at_k_macro(y_true: np.ndarray, y_score: np.ndarray, k: int = 5) -> float:
"""
Macro Recall@k (query-level recall)
β First compute recall per *query* (#hits / #positives in that query).
β Then average across all queries that actually contain β₯1 positive.
"""
mask = y_true.sum(axis=1) > 0 # keep only valid queries
if not mask.any():
return np.nan
Yt, Ys = y_true[mask], y_score[mask]
topk = np.argsort(-Ys, axis=1)[:, :k]
rows = np.arange(Yt.shape[0])[:, None]
hits_per_q = (Yt[rows, topk] > 0).sum(axis=1) # shape: (n_queries,)
pos_per_q = Yt.sum(axis=1)
return np.mean(hits_per_q / pos_per_q) # average of query recalls
# ---------------------------
# Precision@k
# ---------------------------
def precision_at_k_micro(y_true: np.ndarray, y_score: np.ndarray, k: int = 5) -> float:
"""
Micro Precision@k (pool-level precision)
β Numerator = total #hits across all queries.
β Denominator = total #predictions considered (n_queries Β· k).
"""
if y_true.size == 0:
return np.nan
topk = np.argsort(-y_score, axis=1)[:, :k]
rows = np.arange(topk.shape[0])[:, None]
hits = (y_true[rows, topk] > 0).sum()
total_pred = y_true.shape[0] * k
return hits / total_pred
def precision_at_k_macro(y_true: np.ndarray, y_score: np.ndarray, k: int = 5) -> float:
"""
Macro Precision@k (query-level precision)
β Compute precision = (#hits / k) for each query, **including those with zero positives**,
then average.
"""
if y_true.size == 0:
return np.nan
topk = np.argsort(-y_score, axis=1)[:, :k]
rows = np.arange(topk.shape[0])[:, None]
rel = y_true[rows, topk] > 0 # shape: (n_queries, k)
precision_per_q = rel.mean(axis=1) # mean over k positions
return precision_per_q.mean()
# helper methods
def bootstrap(
metric_fn: Callable[[np.ndarray, np.ndarray], float],
y_true: np.ndarray,
y_score: np.ndarray,
n: int = 1000,
rng=None,
) -> dict[str, float]:
rng = np.random.default_rng(rng)
idx = np.arange(len(y_true))
vals: list[float] = []
while len(vals) < n:
sample = rng.choice(idx, size=len(idx), replace=True)
t = y_true[sample]
s = y_score[sample]
if t.sum() == 0: # no positive at all β resample
continue
vals.append(metric_fn(t, s))
result = np.asarray(vals)
# get 95% confidence interval
lo, hi = np.percentile(result, [2.5, 97.5])
return {"mean": result.mean(), "low": lo, "high": hi}
def evaluate_with_ranx(
pred_rels: set[BinaryRelation],
target_rels: set[BinaryRelation],
metrics: list[str],
include_queries_without_gold: bool = True,
) -> dict[str, float]:
# lazy import to not require ranx via requirements.txt
import ranx
all_rels = set(pred_rels) | set(target_rels)
all_heads = {rel.head for rel in all_rels}
head2id = {head: f"q_{idx}" for idx, head in enumerate(sorted(all_heads))}
tail_and_label2id = {(ann.tail, ann.label): f"d_{idx}" for idx, ann in enumerate(all_rels)}
qrels_dict: dict[str, dict[str, int]] = defaultdict(dict) # {query_id: {doc_id: 1}}
run_dict: dict[str, dict[str, float]] = defaultdict(dict) # {query_id: {doc_id: score}}
for target_rel in target_rels:
query_id = head2id[target_rel.head]
doc_id = tail_and_label2id[(target_rel.tail, target_rel.label)]
if target_rel.score != 1.0:
raise ValueError(
f"target score must be 1.0, but got {target_rel.score} for {target_rel}"
)
qrels_dict[query_id][doc_id] = 1
for pred_rel in pred_rels:
query_id = head2id[pred_rel.head]
doc_id = tail_and_label2id[(pred_rel.tail, pred_rel.label)]
run_dict[query_id][doc_id] = pred_rel.score
if include_queries_without_gold:
# add missing query ids to rund_dict and qrels_dict
for query_id in set(head2id.values()) - set(qrels_dict):
qrels_dict[query_id] = {}
# evaluate
qrels = ranx.Qrels(qrels_dict)
run = ranx.Run(run_dict)
results = ranx.evaluate(qrels, run, metrics, make_comparable=True)
return results
def deduplicate_relations(
relations: Iterable[BinaryRelation], caption: str
) -> Set[BinaryRelation]:
pred2scores = defaultdict(set)
for ann in relations:
pred2scores[ann].add(round(ann.score, 4))
# warning for duplicates
preds_with_duplicates = [ann for ann, scores in pred2scores.items() if len(scores) > 1]
if len(preds_with_duplicates) > 0:
logger.warning(
f"there are {len(preds_with_duplicates)} {caption} with duplicates: "
f"{preds_with_duplicates}. We will take the max score for each annotation."
)
# take the max score for each annotation
result = {ann.copy(score=max(scores)) for ann, scores in pred2scores.items()}
return result
def construct_y_true_and_score(
preds: Iterable[BinaryRelation], targets: Iterable[BinaryRelation]
) -> Tuple[np.ndarray, np.ndarray]:
# helper constructs
all_anns = set(preds) | set(targets)
head2relations = defaultdict(list)
for ann in all_anns:
head2relations[ann.head].append(ann)
target2score = {rel: rel.score for rel in targets}
pred2score = {rel: rel.score for rel in preds}
max_len = max(len(relations) for relations in head2relations.values())
target_rows, pred_rows = [], []
for query in head2relations:
relations = head2relations[query]
# get a very small, random score for missing predictions. Or should we use 0.0 as before? or NEG_INF?
missing_pred_score = NEG_INF # np.random.uniform(0.0, 0.001) #0.0 #
missing_target_score = 0
query_scores = [
(target2score.get(ann, missing_target_score), pred2score.get(ann, missing_pred_score))
for ann in relations
]
# sort by descending order of prediction score
query_scores_sorted = np.array(sorted(query_scores, key=lambda x: x[1], reverse=True))
# pad with zeros so every row has the same length
pad_width = max_len - len(query_scores)
query_target = np.pad(
query_scores_sorted[:, 0], (0, pad_width), constant_values=missing_target_score
)
query_pred = np.pad(
query_scores_sorted[:, 1], (0, pad_width), constant_values=missing_pred_score
)
target_rows.append(query_target)
pred_rows.append(query_pred)
y_true = np.vstack(target_rows) # shape (n_queries, max_len)
y_score = np.vstack(pred_rows)
return y_true, y_score
class SemanticallySameRankingMetric(DocumentMetric):
def __init__(
self,
layer: str,
label: Optional[str] = None,
add_reversed: bool = False,
require_positive_gold: bool = False,
bootstrap_n: Optional[int] = None,
k_values: Optional[List[int]] = None,
return_coverage: bool = True,
show_as_markdown: bool = False,
use_ranx: bool = False,
add_stats_to_result: bool = False,
) -> None:
super().__init__()
self.layer = layer
self.label = label
self.add_reversed = add_reversed
self.require_positive_gold = require_positive_gold
self.bootstrap_n = bootstrap_n
self.k_values = k_values if k_values is not None else [1, 5, 10]
self.return_coverage = return_coverage
self.show_as_markdown = show_as_markdown
self.use_ranx = use_ranx
self.add_stats_to_result = add_stats_to_result
self.metrics = {
"macro_ndcg": macro_ndcg,
"macro_mrr": true_mrr,
"macro_map": macro_map,
"micro_ap": ap_micro,
}
for name, func in [
("macro_ndcg", macro_ndcg),
("micro_recall", recall_at_k_micro),
("micro_precision", precision_at_k_micro),
("macro_recall", recall_at_k_macro),
("macro_precision", precision_at_k_macro),
]:
for k in self.k_values:
self.metrics[f"{name}@{k}"] = partial(func, k=k) # type: ignore
self.ranx_metrics = ["map", "mrr", "ndcg"]
for name in ["recall", "precision", "ndcg"]:
for k in self.k_values:
self.ranx_metrics.append(f"{name}@{k}")
def reset(self) -> None:
"""
Reset the metric to its initial state.
"""
self._preds: List[BinaryRelation] = []
self._targets: List[BinaryRelation] = []
def _update(self, document):
layer = document[self.layer]
ann: BinaryRelation
for ann in layer:
if self.label is None or ann.label == self.label:
if ann.score > 0.0:
self._targets.append(ann.copy())
if self.add_reversed:
self._targets.append(ann.copy(head=ann.tail, tail=ann.head))
for ann in layer.predictions:
if self.label is None or ann.label == self.label:
if ann.score > 0.0:
self._preds.append(ann.copy())
if self.add_reversed:
self._preds.append(ann.copy(head=ann.tail, tail=ann.head))
def _compute(self):
# take the max score for each annotation
preds_deduplicated = deduplicate_relations(self._preds, "predictions")
targets_deduplicated = deduplicate_relations(self._targets, "targets")
stats = {
"gold": len(targets_deduplicated),
"preds": len(preds_deduplicated),
"queries": len(
set(ann.head for ann in targets_deduplicated)
| set(ann.head for ann in preds_deduplicated)
),
}
if self.use_ranx:
if self.bootstrap_n is not None:
raise ValueError(
"Ranx does not support bootstrapping. Please set bootstrap_n=None."
)
scores = evaluate_with_ranx(
preds_deduplicated,
targets_deduplicated,
metrics=self.ranx_metrics,
include_queries_without_gold=not self.require_positive_gold,
)
if self.add_stats_to_result:
scores.update(stats)
# logger.info(f"results via ranx:\n{pd.Series(ranx_result).sort_index().round(3).to_markdown()}")
df = pd.DataFrame.from_records([scores], index=["score"])
else:
y_true, y_score = construct_y_true_and_score(
preds=preds_deduplicated, targets=targets_deduplicated
)
# original definition β share of queries with β₯1 positive
coverage = (y_true.sum(axis=1) > 0).mean()
# keep only queries that actually have at least one gold positive
if self.require_positive_gold:
mask = y_true.sum(axis=1) > 0 # shape: (n_queries,)
y_true = y_true[mask]
y_score = y_score[mask]
if self.bootstrap_n is not None:
scores = {
name: bootstrap(fn, y_true, y_score, n=self.bootstrap_n)
for name, fn in self.metrics.items()
}
if self.add_stats_to_result:
scores["stats"] = stats
df = pd.DataFrame(scores)
else:
scores = {name: fn(y_true, y_score) for name, fn in self.metrics.items()}
if self.add_stats_to_result:
scores.update(stats)
df = pd.DataFrame.from_records([scores], index=["score"])
if self.return_coverage:
scores["coverage"] = coverage
if self.show_as_markdown:
if not self.add_stats_to_result:
logger.info(
logger.info(
f'\nstatistics ({self.layer}):\n{pd.Series(stats, name="value").to_markdown()}'
)
)
logger.info(f"\n{self.layer}:\n{df.round(4).T.to_markdown()}")
return scores
|