/
metrics.py
509 lines (385 loc) · 14.3 KB
/
metrics.py
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
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
import logging
import math
import random
from collections.abc import Iterable
from typing import List
import evaluate as hf_evaluate
import numpy as np
import sacrebleu
import sklearn.metrics
from lm_eval.api.registry import register_aggregation, register_metric
eval_logger = logging.getLogger("lm-eval")
# Register Aggregations First
@register_aggregation("bypass")
def bypass_agg(arr):
return 999
@register_aggregation("mean")
def mean(arr):
return sum(arr) / len(arr)
@register_aggregation("median")
def median(arr):
return arr[len(arr) // 2]
# Certain metrics must be calculated across all documents in a benchmark.
# We use them as aggregation metrics, paired with no-op passthrough metric fns.
@register_aggregation("perplexity")
def perplexity(items):
return math.exp(-mean(items))
@register_aggregation("weighted_perplexity")
def weighted_perplexity(items):
return math.exp(-weighted_mean(items))
@register_aggregation("bits_per_byte")
def bits_per_byte(items):
return -weighted_mean(items) / math.log(2)
@register_aggregation("f1")
def f1_score(items):
unzipped_list = list(zip(*items))
golds = unzipped_list[0]
preds = unzipped_list[1]
fscore = sklearn.metrics.f1_score(golds, preds)
return np.max(fscore)
@register_aggregation("matthews_corrcoef")
def matthews_corrcoef(items):
unzipped_list = list(zip(*items))
golds = unzipped_list[0]
preds = unzipped_list[1]
# print(preds)
return sklearn.metrics.matthews_corrcoef(golds, preds)
@register_aggregation("bleu")
def bleu(items):
"""The Bilingual Evaluation Understudy Score, or BLEU for short, is a metric
for evaluating a generated sentence to a reference sentence. It counts matching
n-grams in the candidate translation to n-grams in the reference text, where
1-gram or unigram would be each token and a bigram comparison would be each
word pair. The comparison is made regardless of word order
Source: https://machinelearningmastery.com/calculate-bleu-score-for-text-python/
Paper: https://www.aclweb.org/anthology/P02-1040/
Higher is better
"""
refs = list(zip(*items))[0]
preds = list(zip(*items))[1]
refs, preds = _sacreformat(refs, preds)
return sacrebleu.corpus_bleu(preds, refs).score
@register_aggregation("chrf")
def chrf(items):
"""chrF++ is a tool for automatic evaluation of machine translation output
based on character n-gram precision and recall enhanced with word n-grams.
Source: https://github.com/m-popovic/chrF
Paper: https://www.aclweb.org/anthology/W15-3049.pdf
Higher is better # TODO I think
"""
refs = list(zip(*items))[0]
preds = list(zip(*items))[1]
refs, preds = _sacreformat(refs, preds)
return sacrebleu.corpus_chrf(preds, refs).score
@register_aggregation("ter")
def ter(items):
"""Translation Error Rate is an error metric for machine translation that
measures the number of edits required to change a system output into one
of the references
Source: http://www.cs.umd.edu/~snover/tercom/
Paper: http://mt-archive.info/AMTA-2006-Snover.pdf
Lower is better
"""
refs = list(zip(*items))[0]
preds = list(zip(*items))[1]
refs, preds = _sacreformat(refs, preds)
return sacrebleu.corpus_ter(preds, refs).score
@register_aggregation("brier_score")
def brier_score(items): # This is a passthrough function
gold, predictions = list(zip(*items))
gold = list(gold)
gold_one_hot = np.eye(np.max(gold) + 1)[gold]
predictions = list(zip(*items))[1]
return np.mean(np.sum((predictions - gold_one_hot) ** 2, axis=1))
@register_metric(
metric="brier_score",
higher_is_better=False,
output_type=["multiple_choice"],
aggregation="brier_score",
)
def brier_score_fn(items): # This is a passthrough function
return items
@register_metric(
metric="acc",
higher_is_better=True,
output_type=["loglikelihood", "multiple_choice"],
aggregation="mean",
)
def acc_fn(items): # This is a passthrough function
return items
@register_metric(
metric="acc_norm",
higher_is_better=True,
output_type=["loglikelihood", "multiple_choice"],
aggregation="mean",
)
def acc_norm_fn(items): # This is a passthrough function
return items
@register_metric(
metric="acc_mutual_info",
higher_is_better=True,
output_type="multiple_choice",
aggregation="mean",
)
def acc_mutual_info_fn(items): # This is a passthrough function
return items
exact_match = hf_evaluate.load("exact_match")
@register_metric(
metric="exact_match",
higher_is_better=True,
output_type="generate_until",
aggregation="mean",
)
def exact_match_fn(**kwargs):
return exact_match.compute(**kwargs)
@register_metric(
metric="perplexity",
higher_is_better=False,
output_type="loglikelihood",
aggregation="perplexity",
)
def perplexity_fn(items): # This is a passthrough function
return items
@register_metric(
metric="word_perplexity",
higher_is_better=False,
output_type="loglikelihood_rolling",
aggregation="weighted_perplexity",
)
def word_perplexity_fn(items): # This is a passthrough function
return items
@register_metric(
metric="byte_perplexity",
higher_is_better=False,
output_type="loglikelihood_rolling",
aggregation="weighted_perplexity",
)
def byte_perplexity_fn(items): # This is a passthrough function
return items
@register_metric(
metric="bits_per_byte",
higher_is_better=False,
output_type="loglikelihood_rolling",
aggregation="bits_per_byte",
)
def bits_per_byte_fn(items): # This is a passthrough function
return items
def pop_stddev(arr):
mu = mean(arr)
return math.sqrt(sum([(x - mu) ** 2 for x in arr]) / len(arr))
def sample_stddev(arr):
mu = mean(arr)
return math.sqrt(sum([(x - mu) ** 2 for x in arr]) / (len(arr) - 1))
def mean_stderr(arr):
return sample_stddev(arr) / math.sqrt(len(arr))
@register_metric(
metric="bypass",
higher_is_better=True,
output_type=["loglikelihood", "multiple_choice", "generate_until"],
aggregation="bypass",
)
def bypass(items):
return None
@register_metric(
metric="mcc",
higher_is_better=True,
output_type="multiple_choice",
aggregation="matthews_corrcoef",
)
def mcc_fn(items): # This is a passthrough function
return items
@register_metric(
metric="f1",
higher_is_better=True,
output_type="multiple_choice",
aggregation="f1",
)
def f1_fn(items): # This is a passthrough function
return items
@register_metric(
metric="bleu",
higher_is_better=True,
output_type="generate_until",
aggregation="bleu",
)
def bleu_fn(items): # This is a passthrough function
return items
@register_metric(
metric="chrf",
higher_is_better=True,
output_type="generate_until",
aggregation="chrf",
)
def chrf_fn(items): # This is a passthrough function
return items
@register_metric(
metric="ter",
higher_is_better=True,
output_type="generate_until",
aggregation="ter",
)
def ter_fn(items): # This is a passthrough function
return items
@register_metric(
metric="acc_all",
higher_is_better=True,
output_type="loglikelihood",
aggregation="mean",
)
def acc_all(items):
# Only count as correct if all answers are labeled correctly for each question
question_scoring_dict = {}
preds = list(zip(*items))[0]
docs = list(zip(*items))[1]
for doc, pred in zip(docs, preds):
paragraph_id = doc["idx"]["paragraph"]
question_id = doc["idx"]["question"]
if (paragraph_id, question_id) not in question_scoring_dict:
question_scoring_dict[(paragraph_id, question_id)] = []
gold_label = doc["label"] == 1
question_scoring_dict[(paragraph_id, question_id)].append(gold_label == pred)
acc = np.mean([int(all(x)) for x in question_scoring_dict.values()])
return acc
def acc_all_stderr(items):
# Only count as correct if all answers are labeled correctly for each question
question_scoring_dict = {}
preds = list(zip(*items))[0]
docs = list(zip(*items))[1]
for doc, pred in zip(docs, preds):
question_id = doc["idx"]["question"]
if question_id not in question_scoring_dict:
question_scoring_dict[question_id] = []
gold_label = doc["label"] == 1
question_scoring_dict[question_id].append(gold_label == pred)
acc = mean_stderr([int(all(x)) for x in question_scoring_dict.values()])
return acc
def metric_max_over_ground_truths(metric_fn, prediction, ground_truths):
"""Compute max metric between prediction and each ground truth."""
scores_for_ground_truths = []
for ground_truth in ground_truths:
score = metric_fn(prediction, ground_truth)
scores_for_ground_truths.append(score)
return max(scores_for_ground_truths)
def weighted_mean(items):
a, b = zip(*items)
return sum(a) / sum(b)
def is_non_str_iterable(obj):
return isinstance(obj, Iterable) and not isinstance(obj, str)
def _sacreformat(refs, preds):
"""Format refs and preds for sacrebleu corpus calculation. It is very particular"""
# Sacrebleu expects (List[str], List[List[str])
# e.g. sacrebleu.corpus_bleu([pred_t], [[ref1_stream], [ref2_stream], ...])
# Note [ref1_stream] is the first reference for each pred.
# So lists are size N and (M, N) for N preds and M possible refs for each pred
# This is a different order of dimensions that I would expect
# We expect refs to be List[str] or List[List[str]], the outer list corresponding to preds
# Must become List[List[str]] with the inner list corresponding to preds
if not is_non_str_iterable(refs):
refs = list(refs)
if not is_non_str_iterable(refs[0]):
refs = [[ref] for ref in refs]
refs = list(zip(*refs))
# Note the number of refs in each ref list much match the number of preds
# We expect preds to be List[str] or List[List[str]]. Must become List[str]
if not is_non_str_iterable(preds):
preds = list(preds)
if is_non_str_iterable(preds[0]):
assert len(preds[0]) == 1, f"Pred must be a str, was {preds[0]}"
preds = [pred[0] for pred in preds]
return refs, preds
# stderr stuff
class _bootstrap_internal:
def __init__(self, f, n) -> None:
self.f = f
self.n = n
def __call__(self, v):
i, xs = v
rnd = random.Random()
rnd.seed(i)
res = []
for _ in range(self.n):
res.append(self.f(rnd.choices(xs, k=len(xs))))
return res
def bootstrap_stderr(f, xs, iters):
import multiprocessing as mp
pool = mp.Pool(mp.cpu_count())
# this gives a biased estimate of the stderr (i.e w/ the mean, it gives something
# equivalent to stderr calculated without Bessel's correction in the stddev.
# Unfortunately, I haven't been able to figure out what the right correction is
# to make the bootstrap unbiased - i considered multiplying by sqrt(n/(n-1)) but
# that would be ad-hoc and I can't prove that that would actually be an unbiased estimator)
# Thankfully, shouldn't matter because our samples are pretty big usually anyways
res = []
chunk_size = min(1000, iters)
from tqdm import tqdm
print("bootstrapping for stddev:", f.__name__)
for bootstrap in tqdm(
pool.imap(
_bootstrap_internal(f, chunk_size),
[(i, xs) for i in range(iters // chunk_size)],
),
total=iters // chunk_size,
):
# sample w replacement
res.extend(bootstrap)
pool.close()
return sample_stddev(res)
def stderr_for_metric(metric, bootstrap_iters):
bootstrappable = [
median,
matthews_corrcoef,
f1_score,
perplexity,
bleu,
chrf,
ter,
]
if metric in bootstrappable:
return lambda x: bootstrap_stderr(metric, x, iters=bootstrap_iters)
stderr = {mean: mean_stderr, acc_all: acc_all_stderr}
return stderr.get(metric, None)
def pooled_sample_stderr(stderrs: List[float], sizes: List[int]):
# Used to aggregate bootstrapped stderrs across subtasks in a group,
# when we are weighting by the size of each subtask.
#
assert len(stderrs) == len(sizes)
# formula source: https://en.wikipedia.org/wiki/Pooled_variance
# and: https://stats.stackexchange.com/a/4841331
# this empirically seems to match running `stderr_for_metric` on all instances
# from the subtasks concatenated with each other.
pooled_sample_var = (
sum([(size - 1) * stderr**2 * size for size, stderr in zip(sizes, stderrs)])
) / (sum(sizes) - len(sizes))
return np.sqrt(pooled_sample_var / sum(sizes))
def combined_sample_stderr(stderrs: List[float], sizes: List[int], metrics=None):
assert (
metrics is not None
), "Need to pass a list of each subtask's metric for this stderr aggregation"
assert len(stderrs) == len(sizes) and len(sizes) == len(metrics)
# See https://github.com/EleutherAI/lm-evaluation-harness/pull/1390 for more documentation.
# This formula depends on sample means.
# removed because it seems to give erroneously huge stderrs for groupings of tasks
# and does not seem to match up with bootstrap-calculated stderrs for groups.
### don't use this unless a statistician has told you it's the right thing to do ###
# accumulators: we'll aggregate pairwise N - 1 times
variance = stderrs[0] ** 2
curr_size = sizes[0]
curr_score = metrics[0]
for stderr, size, score in zip(stderrs[1:], sizes[1:], metrics[1:]):
curr_score = ((curr_score * curr_size) + (score * size)) / (
curr_size + size
) # NOTE: this assumes our aggregation fn is "mean"
variance = ((curr_size - 1) * variance + (size - 1) * (stderr**2)) / (
curr_size + size - 1
) + curr_size * size / ((curr_size + size) * (curr_size + size - 1)) * (
curr_score - score
) ** 2
return np.sqrt(variance)
def aggregate_subtask_metrics(metrics, sizes, weight_by_size=True):
# A helper function that is used to aggregate
# subtask scores cross-task.
# TODO: does not hold for non-mean aggregations
if not weight_by_size:
sizes = [1] * len(sizes)
assert len(metrics) == len(sizes)
return sum([metric * size for metric, size in zip(metrics, sizes)]) / sum(sizes)