metrics.py 15.1 KB
Newer Older
1
import logging
&'s avatar
& committed
2
import math
3
import random
4
from collections.abc import Iterable
5
from typing import List
6

Baber Abbasi's avatar
Baber Abbasi committed
7
import evaluate as hf_evaluate
8
9
10
import numpy as np
import sacrebleu
import sklearn.metrics
&'s avatar
& committed
11

12
from lm_eval.api.registry import register_aggregation, register_metric
13

lintangsutawika's avatar
lintangsutawika committed
14

15
eval_logger = logging.getLogger("lm-eval")
16

17

18
# Register Aggregations First
Baber Abbasi's avatar
Baber Abbasi committed
19
20
21
22
23
@register_aggregation("bypass")
def bypass_agg(arr):
    return 999


24
25
@register_aggregation("mean")
def mean(arr):
26
    return sum(arr) / len(arr)
27
28
29
30
31
32
33


@register_aggregation("median")
def median(arr):
    return arr[len(arr) // 2]


34
# Certain metrics must be calculated across all documents in a benchmark.
haileyschoelkopf's avatar
haileyschoelkopf committed
35
# We use them as aggregation metrics, paired with no-op passthrough metric fns.
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
@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)


haileyschoelkopf's avatar
haileyschoelkopf committed
51
52
53
54
55
56
57
58
59
60
@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)


61
62
63
64
65
66
67
68
69
70
71
72
73
@register_aggregation("squad_f1")
def squad_f1_score(items):
    gold_squad, pred_squad = [], []
    for index, (ref, pred) in enumerate(items):
        pred_dict = {'prediction_text': str(pred), 'id': str(index)}
        ref_dict = {'answers': {'answer_start': [0], 'text': [str(ref)]}, 'id': str(index)}
        gold_squad.append(ref_dict)
        pred_squad.append(pred_dict)
    squad_metric = hf_evaluate.load("squad")
    results_squad = squad_metric.compute(predictions=pred_squad, references=gold_squad)
    return results_squad['f1']/100


haileyschoelkopf's avatar
haileyschoelkopf committed
74
75
76
77
78
79
80
81
82
@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)


83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
@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


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
@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


Lintang Sutawika's avatar
Lintang Sutawika committed
132
133
134
@register_aggregation("brier_score")
def brier_score(items):  # This is a passthrough function
    gold, predictions = list(zip(*items))
Lintang Sutawika's avatar
Lintang Sutawika committed
135
136
    bs, num_class = np.array(predictions).shape

Lintang Sutawika's avatar
Lintang Sutawika committed
137
    gold = list(gold)
Lintang Sutawika's avatar
Lintang Sutawika committed
138
    gold_one_hot = np.eye(num_class)[gold]
Lintang Sutawika's avatar
Lintang Sutawika committed
139
140
141
142
143
144
145
146
147
148
149
150
151
    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


152
153
154
@register_metric(
    metric="acc",
    higher_is_better=True,
155
    output_type=["loglikelihood", "multiple_choice"],
156
157
158
159
160
161
    aggregation="mean",
)
def acc_fn(items):  # This is a passthrough function
    return items


162
163
164
165
166
167
168
169
170
171
@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


172
173
174
175
176
177
178
179
180
181
@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


Baber Abbasi's avatar
Baber Abbasi committed
182
exact_match = hf_evaluate.load("exact_match")
183
184


185
186
187
188
189
190
@register_metric(
    metric="exact_match",
    higher_is_better=True,
    output_type="generate_until",
    aggregation="mean",
)
191
192
def exact_match_fn(**kwargs):
    return exact_match.compute(**kwargs)
193
194


195
196
197
198
199
200
201
202
203
@register_metric(
    metric="squad",
    higher_is_better=True,
    output_type="generate_until",
    aggregation="squad_f1"
)
def squad_fn(items):
    return items

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
@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

&'s avatar
& committed
243

Leo Gao's avatar
Leo Gao committed
244
def pop_stddev(arr):
245
    mu = mean(arr)
Leo Gao's avatar
Leo Gao committed
246
247
248
    return math.sqrt(sum([(x - mu) ** 2 for x in arr]) / len(arr))


Leo Gao's avatar
Leo Gao committed
249
def sample_stddev(arr):
250
    mu = mean(arr)
Leo Gao's avatar
Leo Gao committed
251
252
253
    return math.sqrt(sum([(x - mu) ** 2 for x in arr]) / (len(arr) - 1))


Leo Gao's avatar
Leo Gao committed
254
def mean_stderr(arr):
Leo Gao's avatar
Leo Gao committed
255
    return sample_stddev(arr) / math.sqrt(len(arr))
Leo Gao's avatar
Leo Gao committed
256
257


Baber Abbasi's avatar
Baber Abbasi committed
258
259
260
261
262
263
264
265
266
267
@register_metric(
    metric="bypass",
    higher_is_better=True,
    output_type=["loglikelihood", "multiple_choice", "generate_until"],
    aggregation="bypass",
)
def bypass(items):
    return None


haileyschoelkopf's avatar
haileyschoelkopf committed
268
269
270
271
272
273
274
275
@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
276
277
278


@register_metric(
279
    metric="f1",
280
    higher_is_better=True,
281
    output_type=["multiple_choice"],
haileyschoelkopf's avatar
haileyschoelkopf committed
282
    aggregation="f1",
283
)
284
def f1_fn(items):  # This is a passthrough function
haileyschoelkopf's avatar
haileyschoelkopf committed
285
    return items
286
287


288
289
290
@register_metric(
    metric="bleu",
    higher_is_better=True,
291
    output_type="generate_until",
292
293
294
295
296
297
    aggregation="bleu",
)
def bleu_fn(items):  # This is a passthrough function
    return items


298
299
300
@register_metric(
    metric="chrf",
    higher_is_better=True,
301
    output_type="generate_until",
302
303
304
305
306
307
308
309
310
    aggregation="chrf",
)
def chrf_fn(items):  # This is a passthrough function
    return items


@register_metric(
    metric="ter",
    higher_is_better=True,
311
    output_type="generate_until",
312
313
314
315
316
317
    aggregation="ter",
)
def ter_fn(items):  # This is a passthrough function
    return items


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
@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


Leo Gao's avatar
Leo Gao committed
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
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

&'s avatar
& committed
360
361
362
363
364
365
366
367
368
369

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)


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
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


Leo Gao's avatar
Leo Gao committed
410
class _bootstrap_internal:
Ethan Smith's avatar
Ethan Smith committed
411
    def __init__(self, f, n) -> None:
Leo Gao's avatar
Leo Gao committed
412
413
        self.f = f
        self.n = n
414

Leo Gao's avatar
Leo Gao committed
415
416
417
418
419
420
421
422
423
    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

Leo Gao's avatar
Leo Gao committed
424

425
def bootstrap_stderr(f, xs, iters):
Leo Gao's avatar
Leo Gao committed
426
    import multiprocessing as mp
Fabrizio Milo's avatar
Fabrizio Milo committed
427

Leo Gao's avatar
Leo Gao committed
428
    pool = mp.Pool(mp.cpu_count())
Leo Gao's avatar
Leo Gao committed
429
    # this gives a biased estimate of the stderr (i.e w/ the mean, it gives something
Fabrizio Milo's avatar
Fabrizio Milo committed
430
    # equivalent to stderr calculated without Bessel's correction in the stddev.
Leo Gao's avatar
Leo Gao committed
431
432
433
434
    # 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
Leo Gao's avatar
Leo Gao committed
435
    res = []
436
    chunk_size = min(1000, iters)
Leo Gao's avatar
Leo Gao committed
437
    from tqdm import tqdm
Fabrizio Milo's avatar
Fabrizio Milo committed
438

Leo Gao's avatar
Leo Gao committed
439
    print("bootstrapping for stddev:", f.__name__)
Fabrizio Milo's avatar
Fabrizio Milo committed
440
441
    for bootstrap in tqdm(
        pool.imap(
442
            _bootstrap_internal(f, chunk_size),
Fabrizio Milo's avatar
Fabrizio Milo committed
443
444
445
446
            [(i, xs) for i in range(iters // chunk_size)],
        ),
        total=iters // chunk_size,
    ):
Leo Gao's avatar
Leo Gao committed
447
        # sample w replacement
Leo Gao's avatar
Leo Gao committed
448
        res.extend(bootstrap)
Leo Gao's avatar
Leo Gao committed
449

Leo Gao's avatar
Leo Gao committed
450
    pool.close()
Leo Gao's avatar
Leo Gao committed
451
    return sample_stddev(res)
Leo Gao's avatar
Leo Gao committed
452
453


454
455
456
457
458
def stderr_for_metric(metric, bootstrap_iters: int):
    if bootstrap_iters <= 0:
        # return no function (don't compute stderr) if bootstrap iters = 0
        return None

459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
    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)
475
476
477
478
479
480
481
482
483
484


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
485
486
    # and: https://stats.stackexchange.com/a/4841331
    # this empirically seems to match running `stderr_for_metric` on all instances
487
488
    # from the subtasks concatenated with each other.
    pooled_sample_var = (
489
        sum([(size - 1) * stderr**2 * size for size, stderr in zip(sizes, stderrs)])
490
491
    ) / (sum(sizes) - len(sizes))

492
    return np.sqrt(pooled_sample_var / sum(sizes))
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530


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
531
    if not weight_by_size:
532
533
534
535
536
        sizes = [1] * len(sizes)

    assert len(metrics) == len(sizes)

    return sum([metric * size for metric, size in zip(metrics, sizes)]) / sum(sizes)