metrics.py 11.9 KB
Newer Older
&'s avatar
& committed
1
import math
2
3
4
5
6
from collections.abc import Iterable

import numpy as np
import sacrebleu
import sklearn.metrics
Leo Gao's avatar
Leo Gao committed
7
import random
lintangsutawika's avatar
lintangsutawika committed
8
import evaluate
&'s avatar
& committed
9

lintangsutawika's avatar
lintangsutawika committed
10
from Levenshtein import distance
11
12
13
14
15
16
17
18
19
20
21
22
23
24
from lm_eval.api.registry import register_metric, register_aggregation


# 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]


25
# Certain metrics must be calculated across all documents in a benchmark.
haileyschoelkopf's avatar
haileyschoelkopf committed
26
# We use them as aggregation metrics, paired with no-op passthrough metric fns.
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
@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
42
43
44
45
46
47
48
49
50
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)


@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)


61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
@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


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
@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


lintangsutawika's avatar
lintangsutawika committed
110
111
@register_aggregation("brier_score")
def brier_score(items):  # This is a passthrough function
112
113
114
    gold, predictions = list(zip(*items))
    gold = list(gold)
    gold_one_hot = np.eye(np.max(gold)+1)[gold]
lintangsutawika's avatar
lintangsutawika committed
115
116
117
118
119
120
121
122
123
124
125
126
127
128
    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


129
130
131
132
133
134
135
136
137
138
@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


139
140
141
142
143
144
145
146
147
148
@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


149
150
151
152
153
154
155
156
157
158
@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


159
160
161
exact_match = evaluate.load("exact_match")


lintangsutawika's avatar
lintangsutawika committed
162
163
164
165
166
167
168
169
170
171
172
173
# @register_metric(
#     metric="token_edit_distance",
#     higher_is_better=False,
#     output_type=["generate_until"],
#     aggregation="mean",
# )
# def ted_fn(items):  # This is a passthrough function
    
#     references, predictions = items
#     return distance(references, predictions)


174
175
176
177
178
179
@register_metric(
    metric="exact_match",
    higher_is_better=True,
    output_type="generate_until",
    aggregation="mean",
)
180
181
def exact_match_fn(**kwargs):
    return exact_match.compute(**kwargs)
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
@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
223

Leo Gao's avatar
Leo Gao committed
224
def pop_stddev(arr):
225
    mu = mean(arr)
Leo Gao's avatar
Leo Gao committed
226
227
228
    return math.sqrt(sum([(x - mu) ** 2 for x in arr]) / len(arr))


Leo Gao's avatar
Leo Gao committed
229
def sample_stddev(arr):
230
    mu = mean(arr)
Leo Gao's avatar
Leo Gao committed
231
232
233
    return math.sqrt(sum([(x - mu) ** 2 for x in arr]) / (len(arr) - 1))


Leo Gao's avatar
Leo Gao committed
234
def mean_stderr(arr):
Leo Gao's avatar
Leo Gao committed
235
    return sample_stddev(arr) / math.sqrt(len(arr))
Leo Gao's avatar
Leo Gao committed
236
237


haileyschoelkopf's avatar
haileyschoelkopf committed
238
239
240
241
242
243
244
245
@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
246
247
248


@register_metric(
249
    metric="f1",
250
251
    higher_is_better=True,
    output_type="multiple_choice",
haileyschoelkopf's avatar
haileyschoelkopf committed
252
    aggregation="f1",
253
)
254
def f1_fn(items):  # This is a passthrough function
haileyschoelkopf's avatar
haileyschoelkopf committed
255
    return items
256
257


258
259
260
@register_metric(
    metric="bleu",
    higher_is_better=True,
261
    output_type="generate_until",
262
263
264
265
266
267
    aggregation="bleu",
)
def bleu_fn(items):  # This is a passthrough function
    return items


268
269
270
@register_metric(
    metric="chrf",
    higher_is_better=True,
271
    output_type="generate_until",
272
273
274
275
276
277
278
279
280
    aggregation="chrf",
)
def chrf_fn(items):  # This is a passthrough function
    return items


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


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
@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
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
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
330
331
332
333
334
335
336
337
338
339

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)


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
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
380
class _bootstrap_internal:
Ethan Smith's avatar
Ethan Smith committed
381
    def __init__(self, f, n) -> None:
Leo Gao's avatar
Leo Gao committed
382
383
        self.f = f
        self.n = n
384

Leo Gao's avatar
Leo Gao committed
385
386
387
388
389
390
391
392
393
    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
394

395
def bootstrap_stderr(f, xs, iters):
Leo Gao's avatar
Leo Gao committed
396
    import multiprocessing as mp
Fabrizio Milo's avatar
Fabrizio Milo committed
397

Leo Gao's avatar
Leo Gao committed
398
    pool = mp.Pool(mp.cpu_count())
Leo Gao's avatar
Leo Gao committed
399
    # this gives a biased estimate of the stderr (i.e w/ the mean, it gives something
Fabrizio Milo's avatar
Fabrizio Milo committed
400
    # equivalent to stderr calculated without Bessel's correction in the stddev.
Leo Gao's avatar
Leo Gao committed
401
402
403
404
    # 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
405
    res = []
406
    chunk_size = min(1000, iters)
Leo Gao's avatar
Leo Gao committed
407
    from tqdm import tqdm
Fabrizio Milo's avatar
Fabrizio Milo committed
408

Leo Gao's avatar
Leo Gao committed
409
    print("bootstrapping for stddev:", f.__name__)
Fabrizio Milo's avatar
Fabrizio Milo committed
410
411
    for bootstrap in tqdm(
        pool.imap(
412
            _bootstrap_internal(f, chunk_size),
Fabrizio Milo's avatar
Fabrizio Milo committed
413
414
415
416
            [(i, xs) for i in range(iters // chunk_size)],
        ),
        total=iters // chunk_size,
    ):
Leo Gao's avatar
Leo Gao committed
417
        # sample w replacement
Leo Gao's avatar
Leo Gao committed
418
        res.extend(bootstrap)
Leo Gao's avatar
Leo Gao committed
419

Leo Gao's avatar
Leo Gao committed
420
    pool.close()
Leo Gao's avatar
Leo Gao committed
421
    return sample_stddev(res)
Leo Gao's avatar
Leo Gao committed
422
423


424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
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)