File size: 20,067 Bytes
0db93dd
9b14558
2038164
 
0db93dd
d423f18
0db93dd
 
 
902ea7b
0db93dd
 
 
 
 
 
 
 
 
d423f18
cc0572c
d423f18
 
 
 
cc0572c
d423f18
 
 
 
 
 
 
 
 
0db93dd
d524551
d423f18
 
 
 
 
 
 
 
0db93dd
d423f18
 
 
 
 
 
 
 
 
 
 
 
 
 
902ea7b
 
 
11723f3
 
902ea7b
 
 
d423f18
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
11723f3
d423f18
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d524551
d423f18
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
11723f3
d423f18
 
 
 
 
0db93dd
d423f18
11723f3
d423f18
 
 
 
 
 
 
0db93dd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d423f18
 
 
 
11723f3
d423f18
 
 
 
 
 
 
 
 
 
 
 
 
0db93dd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
043ae31
0db93dd
 
 
 
 
 
043ae31
0db93dd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
cc0572c
0db93dd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
11723f3
0db93dd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
cc0572c
 
 
 
11723f3
 
 
 
 
 
 
 
 
 
0db93dd
 
 
cc0572c
11723f3
 
 
 
 
 
 
 
 
0db93dd
11723f3
 
 
 
0db93dd
 
 
 
11723f3
 
 
0db93dd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
043ae31
 
 
 
 
11723f3
043ae31
 
11723f3
 
 
 
 
 
0db93dd
043ae31
 
 
0db93dd
 
 
11723f3
cc0572c
 
 
 
11723f3
 
 
 
 
cc0572c
11723f3
 
cc0572c
 
 
11723f3
cc0572c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0db93dd
 
 
 
2038164
 
043ae31
 
 
 
 
 
902ea7b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2038164
 
902ea7b
2038164
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
902ea7b
 
 
 
2038164
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
902ea7b
2038164
 
 
 
 
 
 
 
 
902ea7b
 
 
 
2038164
 
 
 
 
902ea7b
 
 
 
 
2038164
 
 
 
 
 
 
 
 
 
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
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
import uuid
from abc import ABC, abstractmethod
from collections import Counter
from dataclasses import field
from typing import Any, Dict, Generator, List, Optional

import evaluate
import numpy

from .dataclass import InternalField
from .operator import (
    MultiStreamOperator,
    SingleStreamOperator,
    StreamingOperator,
    StreamInstanceOperator,
)
from .operators import CopyFields
from .stream import MultiStream, Stream


def abstract_factory():
    return {}


def abstract_field():
    return field(default_factory=abstract_factory)


class UpdateStream(StreamInstanceOperator):
    update: dict

    def process(self, instance: Dict[str, Any], stream_name: str = None) -> Dict[str, Any]:
        instance.update(self.update)
        return instance


# TODO: currently we have two classes with this name. metric.Metric and matrics.Metric...
class Metric(ABC):
    @property
    @abstractmethod
    def main_score(self):
        pass


class GlobalMetric(SingleStreamOperator, Metric):
    def process(self, stream: Stream, stream_name: str = None) -> Generator:
        references = []
        predictions = []
        global_score = {}

        instances = []

        for instance in stream:
            if "score" not in instance:
                instance["score"] = {"global": global_score, "instance": {}}
            else:
                global_score = instance["score"]["global"]

            refs, pred = instance["references"], instance["prediction"]

            try:
                instance_score = self._compute([refs], [pred])
            except:
                instance_score = {"score": None, "score_name": self.main_score}

                if isinstance(self.main_score, str) and self.main_score is not None:
                    instance_score[self.main_score] = None

            instance["score"]["instance"].update(instance_score)

            references.append(refs)
            predictions.append(pred)
            instances.append(instance)

        result = self._compute(references, predictions)

        global_score.update(result)

        for instance in instances:
            instance["score"]["global"] = global_score
            yield instance

    def _compute(self, references: List[List[str]], predictions: List[str]) -> dict:
        result = self.compute(references, predictions)
        result["score"] = result[self.main_score]
        result["score_name"] = self.main_score
        return result

    @abstractmethod
    def compute(self, references: List[List[str]], predictions: List[str]) -> dict:
        pass


class InstanceMetric(SingleStreamOperator, Metric):
    implemented_reductions: List[str] = field(default_factory=lambda: ["mean"])

    @property
    @abstractmethod
    def reduction_map(self) -> dict:
        pass

    def process(self, stream: Stream, stream_name: str = None) -> Generator:
        global_score = {}
        instances = []

        for instance in stream:
            refs, pred = instance["references"], instance["prediction"]

            instance_score = self._compute(refs, pred)

            if "score" not in instance:
                instance["score"] = {"global": global_score, "instance": {}}
            else:
                global_score = instance["score"]["global"]

            instance["score"]["instance"].update(instance_score)

            instances.append(instance)

        for reduction, fields in self.reduction_map.items():
            assert (
                reduction in self.implemented_reductions
            ), f"Reduction {reduction} is not implemented, use one of {self.implemented_reductions}"

            if reduction == "mean":
                from statistics import mean

                for field in fields:
                    global_score[field] = mean([instance["score"]["instance"][field] for instance in instances])
                    if field == self.main_score:
                        global_score["score"] = global_score[field]
                        global_score["score_name"] = self.main_score

        for instance in instances:
            yield instance

    def _compute(self, references: List[List[str]], predictions: List[str]) -> dict:
        result = self.compute(references=references, predictions=predictions)
        result["score"] = result[self.main_score]
        result["score_name"] = self.main_score
        return result

    @abstractmethod
    def compute(self, references: List[str], prediction: str) -> dict:
        pass


class Squad(GlobalMetric):
    _metric = None
    main_score = "f1"
    metric = "squad"

    def prepare(self):
        super(Squad, self).prepare()
        self._metric = evaluate.load(self.metric)

    def compute(self, references: List[List[str]], predictions: List[str]) -> dict:
        ids = [str(uuid.uuid4()).replace("-", "") for _ in range(len(predictions))]
        formatted_predictions = [
            {"prediction_text": prediction, "id": ids[i]} for i, prediction in enumerate(predictions)
        ]
        formatted_references = [
            {"answers": {"answer_start": [-1], "text": reference}, "id": ids[i]}
            for i, reference in enumerate(references)
        ]

        return self._metric.compute(predictions=formatted_predictions, references=formatted_references)


class SingleReferenceInstanceMetric(InstanceMetric):
    def _compute(self, references: List[str], prediction: str) -> dict:
        result = self.compute(references[0], prediction)
        result["score"] = result[self.main_score]
        result["score_name"] = self.main_score
        return result

    @abstractmethod
    def compute(self, reference, prediction: str) -> dict:
        pass


class Accuracy(SingleReferenceInstanceMetric):
    reduction_map = {"mean": ["accuracy"]}
    main_score = "accuracy"

    def compute(self, reference, prediction: str) -> dict:
        return {"accuracy": float(str(reference) == str(prediction))}


class MetricPipeline(MultiStreamOperator, Metric):
    main_score: str = None
    preprocess_steps: Optional[List[StreamingOperator]] = field(default_factory=list)
    postpreprocess_steps: Optional[List[StreamingOperator]] = field(default_factory=list)
    metric: Metric = None

    def verify(self):
        assert self.main_score is not None, "main_score is not set"

    def prepare(self):
        super().prepare()
        self.prepare_score = CopyFields(
            field_to_field=[
                [f"score/instance/{self.main_score}", "score/instance/score"],
                [f"score/global/{self.main_score}", "score/global/score"],
            ],
            use_query=True,
        )

    def process(self, multi_stream: MultiStream) -> MultiStream:
        for step in self.preprocess_steps:
            multi_stream = step(multi_stream)
        multi_stream = self.metric(multi_stream)
        for step in self.postpreprocess_steps:
            multi_stream = step(multi_stream)
        multi_stream = self.prepare_score(multi_stream)
        return multi_stream


class HuggingfaceMetric(GlobalMetric):
    metric_name: str = None
    main_score: str = None
    scale: float = 1.0
    hf_compute_args: dict = {}

    def prepare(self):
        super().prepare()
        self.metric = evaluate.load(self.metric_name)

    def compute(self, references: List[List[str]], predictions: List[str]) -> dict:
        result = self.metric.compute(predictions=predictions, references=references, **self.hf_compute_args)
        if self.scale != 1.0:
            for key in result:
                if isinstance(result[key], float):
                    result[key] /= self.scale
        return result


class F1(GlobalMetric):
    _metric = None
    main_score = "f1_macro"
    average = None  # Report per class then aggregate by mean
    metric = "f1"

    def prepare(self):
        super(F1, self).prepare()
        self._metric = evaluate.load(self.metric)

    def get_str_id(self, str):
        if str not in self.str_to_id:
            id = len(self.str_to_id)
            self.str_to_id[str] = id
            self.id_to_str[id] = str
        return self.str_to_id[str]

    def compute(self, references: List[List[str]], predictions: List[str]) -> dict:
        assert all(
            len(reference) == 1 for reference in references
        ), "Only a single reference per prediction is allowed in F1 metric"
        self.str_to_id = {}
        self.id_to_str = {}
        formatted_references = [self.get_str_id(reference[0]) for reference in references]
        unique_labels = self.str_to_id.keys()
        formatted_predictions = [self.get_str_id(prediction) for prediction in predictions]
        labels = list(set(formatted_references))
        result = self._metric.compute(
            predictions=formatted_predictions, references=formatted_references, labels=labels, average=self.average
        )
        if isinstance(result["f1"], numpy.ndarray):
            from statistics import mean

            final_result = {self.main_score: mean(result["f1"])}
            for i, label in enumerate(labels):
                final_result["f1_" + self.id_to_str[label]] = result["f1"][i]
        else:
            final_result = {self.main_score: result["f1"]}
        return final_result


class F1Micro(F1):
    main_score = "f1_micro"
    average = "micro"


class F1Macro(F1):
    main_score = "f1_macro"


class F1MultiLabel(GlobalMetric):
    _metric = None
    main_score = "f1_macro"
    average = None  # Report per class then aggregate by mean
    classes_to_ignore = ["none"]

    def prepare(self):
        super(F1MultiLabel, self).prepare()
        self._metric = evaluate.load("f1", "multilabel")

    def add_str_to_id(self, str):
        if not str in self.str_to_id:
            id = len(self.str_to_id)
            self.str_to_id[str] = id
            self.id_to_str[id] = str
        return

    def get_one_hot_vector(self, labels: List[str]):
        result = [0] * len(self.str_to_id)
        for label in labels:
            if label in self.str_to_id:
                result[self.str_to_id[label]] = 1
        return result

    def compute(self, references: List[List[str]], predictions: List[str]) -> dict:
        self.str_to_id = {}
        self.id_to_str = {}
        assert all(
            len(reference) == 1 for reference in references
        ), "Only a single reference per prediction is allowed in F1 metric"
        references = [reference[0] for reference in references]
        labels = [
            l
            for l in set([label for reference in references for label in reference])
            if l not in self.classes_to_ignore
        ]
        # if no classes are left then F1 is not defined
        # (e.g. only "none" in references)
        if len(labels) == 0:
            return {self.main_score: float("nan")}

        for label in labels:
            self.add_str_to_id(label)
        formatted_references = [self.get_one_hot_vector(reference) for reference in references]
        formatted_predictions = [self.get_one_hot_vector(prediction) for prediction in predictions]

        # There is odd behavior in scikit-learn that when passing a one-hot vector with a single
        # element, it is treated a class identifier. Therefore, we add labels=[1] to limit to only
        # to this class.
        if len(labels) == 1:
            labels_param = [1]
        else:
            labels_param = None

        result = self._metric.compute(
            predictions=formatted_predictions,
            references=formatted_references,
            average=self.average,
            labels=labels_param,
        )
        if isinstance(result["f1"], numpy.ndarray):
            from statistics import mean

            assert len(result["f1"]) == len(
                labels
            ), f'F1 result ({result["f1"]}) has more entries than labels ({labels})'
            final_result = {self.main_score: mean(result["f1"])}
            for i, label in enumerate(labels):
                final_result["f1_" + label] = result["f1"][i]
        else:
            final_result = {self.main_score: result["f1"]}
        return final_result


class F1MicroMultiLabel(F1MultiLabel):
    main_score = "f1_micro"
    average = "micro"


class F1MacroMultiLabel(F1MultiLabel):
    main_score = "f1_macro"
    average = None


class Rouge(HuggingfaceMetric):
    metric_name = "rouge"
    main_score = "rougeL"
    scale = 1.0

    use_aggregator: bool = True
    rouge_types: List[str] = ["rouge1", "rouge2", "rougeL", "rougeLsum"]

    sent_split_newline: bool = True

    def prepare(self):
        self.hf_compute_args = {"use_aggregator": self.use_aggregator, "rouge_types": self.rouge_types}

        super().prepare()
        import nltk

        nltk.download("punkt")
        self.sent_tokenize = nltk.sent_tokenize

    def compute(self, references, predictions):
        if self.sent_split_newline:
            predictions = ["\n".join(self.sent_tokenize(prediction.strip())) for prediction in predictions]
            references = [["\n".join(self.sent_tokenize(r.strip())) for r in reference] for reference in references]
        return super().compute(references, predictions)


# Computes chat edit distance, ignoring whitespace
class CharEditDistanceAccuracy(SingleReferenceInstanceMetric):
    reduction_map = {"mean": ["char_edit_dist_accuracy"]}
    main_score = "char_edit_dist_accuracy"

    def prepare(self):
        import editdistance

        self.eval = editdistance.eval

    def compute(self, reference, prediction: str) -> dict:
        formatted_prediction = "".join(prediction.split())
        formatted_reference = "".join(reference.split())
        max_length = max(len(formatted_reference), len(formatted_prediction))
        if max_length == 0:
            return 0
        edit_dist = self.eval(formatted_reference, formatted_prediction)
        return {"char_edit_dist_accuracy": (1 - edit_dist / max_length)}


class Wer(HuggingfaceMetric):
    metric_name = "wer"
    main_score = "wer"

    def prepare(self):
        super().prepare()
        self.metric = evaluate.load(self.metric_name)

    def compute(self, references: List[List[str]], predictions: List[str]) -> dict:
        assert all(
            len(reference) == 1 for reference in references
        ), "Only single reference per prediction is allowed in wer metric"
        formatted_references = [reference[0] for reference in references]
        result = self.metric.compute(predictions=predictions, references=formatted_references)
        return {self.main_score: result}


class Bleu(HuggingfaceMetric):
    metric_name = "bleu"
    main_score = "bleu"
    scale = 1.0


class SacreBleu(HuggingfaceMetric):
    metric_name = "sacrebleu"
    main_score = "score"
    scale = 1.0


class MatthewsCorrelation(HuggingfaceMetric):
    metric_name = "matthews_correlation"
    main_score = "matthews_correlation"
    str_to_id: dict = InternalField(default_factory=dict)

    def get_str_id(self, str):
        if str not in self.str_to_id:
            id = len(self.str_to_id)
            self.str_to_id[str] = id
        return self.str_to_id[str]

    def compute(self, references: List[List[str]], predictions: List[str]) -> dict:
        formatted_references = [self.get_str_id(reference[0]) for reference in references]
        formatted_predictions = [self.get_str_id(prediction) for prediction in predictions]
        result = self.metric.compute(predictions=formatted_predictions, references=formatted_references)
        return result


class CustomF1(GlobalMetric):
    main_score = "f1_micro"
    classes = None

    @abstractmethod
    def get_element_group(self, element):
        pass

    @abstractmethod
    def get_element_representation(self, element):
        pass

    def group_elements(self, l):
        return {
            k: Counter([self.get_element_representation(value) for value in l if self.get_element_group(value) == k])
            for k in set([self.get_element_group(e) for e in l])
        }

    def calculate_groups_ratio(self, actual_group, total_group):
        return sum([min(actual_group[k], total_group[k]) for k in actual_group.keys()]), sum(actual_group.values())

    def f1(self, pn, pd, rn, rd):
        precision = 1.0 if pn == 0 and pd == 0 else pn / pd
        recall = 1.0 if rn == 0 and rd == 0 else rn / rd
        try:
            return 2 * precision * recall / (precision + recall)
        except ZeroDivisionError:
            return 0.0

    def compute(self, references: List[Any], predictions: List[Any]) -> dict:
        # in case reference are List[List[List[Any]]] and predictions are List[List[Any]]:
        if isinstance(references[0], list) and isinstance(references[0][0], list):
            references = [element[0] for element in references]

        assert len(references) == len(predictions), (
            f"references size ({len(references)})" f" doesn't mach predictions sise ({len(references)})."
        )
        if self.classes is None:
            classes = set([self.get_element_group(e) for sublist in references for e in sublist])
        else:
            classes = self.classes
        groups_statistics = dict()
        for references_batch, predictions_batch in zip(references, predictions):
            grouped_references = self.group_elements(references_batch)
            grouped_predictions = self.group_elements(predictions_batch)
            all_groups = set(grouped_references.keys()).union(grouped_predictions.keys())
            for group in all_groups:
                if group not in groups_statistics:
                    groups_statistics[group] = {
                        "precision_numerator": 0,
                        "precision_denominator": 0,
                        "recall_numerator": 0,
                        "recall_denominator": 0,
                    }
                references_by_group = grouped_references.get(group, Counter([]))
                predictions_by_group = grouped_predictions.get(group, Counter([]))
                pn, pd = self.calculate_groups_ratio(
                    actual_group=predictions_by_group, total_group=references_by_group
                )
                rn, rd = self.calculate_groups_ratio(
                    actual_group=references_by_group, total_group=predictions_by_group
                )
                groups_statistics[group]["precision_numerator"] += pn
                groups_statistics[group]["precision_denominator"] += pd
                groups_statistics[group]["recall_numerator"] += rn
                groups_statistics[group]["recall_denominator"] += rd

        result = {}
        num_of_unknown_class_predictions = 0
        pn_total = pd_total = rn_total = rd_total = 0
        for group in groups_statistics.keys():
            pn, pd, rn, rd = (
                groups_statistics[group]["precision_numerator"],
                groups_statistics[group]["precision_denominator"],
                groups_statistics[group]["recall_numerator"],
                groups_statistics[group]["recall_denominator"],
            )
            pn_total, pd_total, rn_total, rd_total = pn_total + pn, pd_total + pd, rn_total + rn, rd_total + rd
            if group in classes:
                result[f"f1_{group}"] = self.f1(pn, pd, rn, rd)
            else:
                num_of_unknown_class_predictions += pd
        try:
            result["f1_macro"] = sum(result.values()) / len(result.keys())
        except ZeroDivisionError:
            result["f1_macro"] = 1.0

        amount_of_predictions = pd_total
        if amount_of_predictions == 0:
            result["in_classes_support"] = 1.0
        else:
            result["in_classes_support"] = 1.0 - num_of_unknown_class_predictions / amount_of_predictions
        result[f"f1_micro"] = self.f1(pn_total, pd_total, rn_total, rd_total)
        return result


class NER(CustomF1):
    def get_element_group(self, element):
        return element[1]

    def get_element_representation(self, element):
        return str(element)