metrics.py 12.8 KB
Newer Older
1
import logging
&'s avatar
& committed
2
import math
3
import random
4
5
from collections.abc import Iterable

6
import evaluate
7
8
9
import numpy as np
import sacrebleu
import sklearn.metrics
&'s avatar
& committed
10

11
from lm_eval.api.registry import register_aggregation, register_metric
12

13
eval_logger = logging.getLogger("lm-eval")
14
15
16
17
18
19
20
21
22
23
24
25
26


# Register Aggregations First
@register_aggregation("mean")
def mean(arr):
    return sum(arr) / len(arr)


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


27
# Certain metrics must be calculated across all documents in a benchmark.
haileyschoelkopf's avatar
haileyschoelkopf committed
28
# We use them as aggregation metrics, paired with no-op passthrough metric fns.
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
@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
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
@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)


63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
@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


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
@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


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
@register_aggregation("ece")
def ece(items: list) -> float:
    probs: list[float] = []
    scores: list[float] = []
    for i in range(len(items)):
        # Get only largest probability from each example
        largest_idx = np.argmax(items[i]["probs"])
        probs.append(items[i]["probs"][largest_idx])
        scores.append(items[i]["scores"][largest_idx])

    sorted_indices = np.argsort(probs)
    sorted_probs = np.asarray(probs)[sorted_indices]
    sorted_scores = np.asarray(scores)[sorted_indices]

    def bin_to_subsets(array: np.ndarray, num_subsets: int = 10) -> np.ndarray:
        subset_size: int = len(array) // num_subsets
        remainder: int = len(array) % num_subsets
        subsets: list[np.ndarray] = []
        start: int = 0
        for _ in range(num_subsets):
            subset_end: int = start + subset_size + (1 if remainder > 0 else 0)
            subsets.append(array[start:subset_end])
            start = subset_end
            remainder -= 1
        return subsets

    probs = np.asarray([np.mean(x) for x in bin_to_subsets(sorted_probs, 10)])
    freqs = np.asarray([np.mean(x) for x in bin_to_subsets(sorted_scores, 10)])
    return np.sum(np.abs(freqs - probs)) / len(freqs)


143
144
145
146
147
148
149
150
151
152
@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


153
154
155
156
157
158
159
160
161
162
@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


163
164
165
166
167
168
169
170
171
172
@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


173
174
175
exact_match = evaluate.load("exact_match")


176
177
178
179
180
181
@register_metric(
    metric="exact_match",
    higher_is_better=True,
    output_type="generate_until",
    aggregation="mean",
)
182
183
def exact_match_fn(**kwargs):
    return exact_match.compute(**kwargs)
184
185


186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
@register_metric(
    metric="ece",
    higher_is_better=False,
    output_type="multiple_choice",
    aggregation="ece",
)
def ece_fn(items):  # This is a passthrough function
    """
    Expected Calibration Error (ECE).

    This consists of the average absolute difference between the fraction of
    model predictions which are correct and the mean of the model's normalized
    probability for those predictions (after binning), for multiple choice questions.

    Paper: https://arxiv.org/abs/2207.05221
    """
    return items


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
@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
244

Leo Gao's avatar
Leo Gao committed
245
def pop_stddev(arr):
246
    mu = mean(arr)
Leo Gao's avatar
Leo Gao committed
247
248
249
    return math.sqrt(sum([(x - mu) ** 2 for x in arr]) / len(arr))


Leo Gao's avatar
Leo Gao committed
250
def sample_stddev(arr):
251
    mu = mean(arr)
Leo Gao's avatar
Leo Gao committed
252
253
254
    return math.sqrt(sum([(x - mu) ** 2 for x in arr]) / (len(arr) - 1))


Leo Gao's avatar
Leo Gao committed
255
def mean_stderr(arr):
Leo Gao's avatar
Leo Gao committed
256
    return sample_stddev(arr) / math.sqrt(len(arr))
Leo Gao's avatar
Leo Gao committed
257
258


haileyschoelkopf's avatar
haileyschoelkopf committed
259
260
261
262
263
264
265
266
@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
267
268
269


@register_metric(
270
    metric="f1",
271
272
    higher_is_better=True,
    output_type="multiple_choice",
haileyschoelkopf's avatar
haileyschoelkopf committed
273
    aggregation="f1",
274
)
275
def f1_fn(items):  # This is a passthrough function
haileyschoelkopf's avatar
haileyschoelkopf committed
276
    return items
277
278


279
280
281
@register_metric(
    metric="bleu",
    higher_is_better=True,
282
    output_type="generate_until",
283
284
285
286
287
288
    aggregation="bleu",
)
def bleu_fn(items):  # This is a passthrough function
    return items


289
290
291
@register_metric(
    metric="chrf",
    higher_is_better=True,
292
    output_type="generate_until",
293
294
295
296
297
298
299
300
301
    aggregation="chrf",
)
def chrf_fn(items):  # This is a passthrough function
    return items


@register_metric(
    metric="ter",
    higher_is_better=True,
302
    output_type="generate_until",
303
304
305
306
307
308
    aggregation="ter",
)
def ter_fn(items):  # This is a passthrough function
    return items


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
@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
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
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
351
352
353
354
355
356
357
358
359
360

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)


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
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
401
class _bootstrap_internal:
Ethan Smith's avatar
Ethan Smith committed
402
    def __init__(self, f, n) -> None:
Leo Gao's avatar
Leo Gao committed
403
404
        self.f = f
        self.n = n
405

Leo Gao's avatar
Leo Gao committed
406
407
408
409
410
411
412
413
414
    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
415

416
def bootstrap_stderr(f, xs, iters):
Leo Gao's avatar
Leo Gao committed
417
    import multiprocessing as mp
Fabrizio Milo's avatar
Fabrizio Milo committed
418

Leo Gao's avatar
Leo Gao committed
419
    pool = mp.Pool(mp.cpu_count())
Leo Gao's avatar
Leo Gao committed
420
    # this gives a biased estimate of the stderr (i.e w/ the mean, it gives something
Fabrizio Milo's avatar
Fabrizio Milo committed
421
    # equivalent to stderr calculated without Bessel's correction in the stddev.
Leo Gao's avatar
Leo Gao committed
422
423
424
425
    # 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
426
    res = []
427
    chunk_size = min(1000, iters)
Leo Gao's avatar
Leo Gao committed
428
    from tqdm import tqdm
Fabrizio Milo's avatar
Fabrizio Milo committed
429

Leo Gao's avatar
Leo Gao committed
430
    print("bootstrapping for stddev:", f.__name__)
Fabrizio Milo's avatar
Fabrizio Milo committed
431
432
    for bootstrap in tqdm(
        pool.imap(
433
            _bootstrap_internal(f, chunk_size),
Fabrizio Milo's avatar
Fabrizio Milo committed
434
435
436
437
            [(i, xs) for i in range(iters // chunk_size)],
        ),
        total=iters // chunk_size,
    ):
Leo Gao's avatar
Leo Gao committed
438
        # sample w replacement
Leo Gao's avatar
Leo Gao committed
439
        res.extend(bootstrap)
Leo Gao's avatar
Leo Gao committed
440

Leo Gao's avatar
Leo Gao committed
441
    pool.close()
Leo Gao's avatar
Leo Gao committed
442
    return sample_stddev(res)
Leo Gao's avatar
Leo Gao committed
443
444


445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
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)