Upload metrics.py with huggingface_hub
Browse files- metrics.py +594 -133
metrics.py
CHANGED
@@ -1,14 +1,18 @@
|
|
|
|
1 |
import re
|
2 |
import string
|
3 |
import uuid
|
4 |
-
from abc import
|
5 |
from collections import Counter
|
6 |
from dataclasses import field
|
7 |
from typing import Any, Dict, Generator, List, Optional, Tuple
|
8 |
|
9 |
import evaluate
|
10 |
import numpy
|
|
|
|
|
11 |
|
|
|
12 |
from .dataclass import InternalField, OptionalField
|
13 |
from .operator import (
|
14 |
MultiStreamOperator,
|
@@ -17,8 +21,14 @@ from .operator import (
|
|
17 |
StreamInstanceOperator,
|
18 |
)
|
19 |
from .operators import CopyFields
|
|
|
20 |
from .stream import MultiStream, Stream
|
21 |
|
|
|
|
|
|
|
|
|
|
|
22 |
|
23 |
def abstract_factory():
|
24 |
return {}
|
@@ -31,23 +41,166 @@ def abstract_field():
|
|
31 |
class UpdateStream(StreamInstanceOperator):
|
32 |
update: dict
|
33 |
|
34 |
-
def process(
|
|
|
|
|
35 |
instance.update(self.update)
|
36 |
return instance
|
37 |
|
38 |
|
39 |
# TODO: currently we have two classes with this name. metric.Metric and matrics.Metric...
|
40 |
-
class Metric(
|
41 |
@property
|
42 |
@abstractmethod
|
43 |
def main_score(self):
|
44 |
pass
|
45 |
|
46 |
|
47 |
-
class
|
48 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
49 |
references = []
|
50 |
predictions = []
|
|
|
51 |
global_score = {}
|
52 |
|
53 |
instances = []
|
@@ -58,58 +211,100 @@ class GlobalMetric(SingleStreamOperator, Metric):
|
|
58 |
else:
|
59 |
global_score = instance["score"]["global"]
|
60 |
|
61 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
62 |
|
|
|
|
|
|
|
|
|
63 |
try:
|
64 |
-
instance_score = self._compute(
|
|
|
|
|
|
|
|
|
65 |
except:
|
66 |
instance_score = {"score": None, "score_name": self.main_score}
|
67 |
|
68 |
-
if isinstance(self.main_score, str)
|
69 |
instance_score[self.main_score] = None
|
70 |
|
71 |
instance["score"]["instance"].update(instance_score)
|
72 |
|
73 |
-
|
74 |
-
predictions.append(pred)
|
75 |
-
instances.append(instance)
|
76 |
-
|
77 |
-
result = self._compute(references, predictions)
|
78 |
|
79 |
global_score.update(result)
|
80 |
|
|
|
|
|
|
|
|
|
|
|
|
|
81 |
for instance in instances:
|
82 |
instance["score"]["global"] = global_score
|
83 |
yield instance
|
84 |
|
85 |
-
def _compute(
|
86 |
-
|
|
|
|
|
|
|
|
|
|
|
87 |
result["score"] = result[self.main_score]
|
88 |
result["score_name"] = self.main_score
|
89 |
return result
|
90 |
|
91 |
@abstractmethod
|
92 |
-
def compute(
|
|
|
|
|
|
|
|
|
|
|
93 |
pass
|
94 |
|
95 |
|
96 |
-
class BulkInstanceMetric(SingleStreamOperator,
|
|
|
97 |
main_score: str
|
98 |
reduction_map: Dict[str, List[str]]
|
99 |
|
100 |
implemented_reductions: List[str] = field(default_factory=lambda: ["mean"])
|
101 |
|
102 |
-
def process(self, stream: Stream, stream_name: str = None) -> Generator:
|
103 |
global_score = {}
|
104 |
instances = []
|
105 |
|
106 |
# consume the stream
|
107 |
references, predictions = map(
|
108 |
-
list,
|
|
|
|
|
|
|
|
|
|
|
|
|
109 |
)
|
110 |
|
|
|
|
|
|
|
|
|
|
|
111 |
# compute the metric over all refs and preds
|
112 |
-
instance_scores = self.compute(
|
|
|
|
|
|
|
|
|
113 |
|
114 |
# add the score and score_name fields
|
115 |
for instance_score in instance_scores:
|
@@ -134,21 +329,38 @@ class BulkInstanceMetric(SingleStreamOperator, Metric):
|
|
134 |
if reduction == "mean":
|
135 |
from statistics import mean
|
136 |
|
137 |
-
for
|
138 |
-
global_score[
|
139 |
-
|
140 |
-
|
|
|
|
|
|
|
|
|
|
|
141 |
global_score["score_name"] = self.main_score
|
142 |
|
|
|
|
|
|
|
|
|
|
|
143 |
for instance in instances:
|
144 |
yield instance
|
145 |
|
146 |
@abstractmethod
|
147 |
-
def compute(
|
|
|
|
|
|
|
|
|
|
|
148 |
pass
|
149 |
|
150 |
|
151 |
-
class InstanceMetric(SingleStreamOperator,
|
|
|
|
|
152 |
implemented_reductions: List[str] = field(default_factory=lambda: ["mean"])
|
153 |
|
154 |
@property
|
@@ -156,15 +368,21 @@ class InstanceMetric(SingleStreamOperator, Metric):
|
|
156 |
def reduction_map(self) -> dict:
|
157 |
pass
|
158 |
|
159 |
-
def process(self, stream: Stream, stream_name: str = None) -> Generator:
|
160 |
global_score = {}
|
161 |
instances = []
|
162 |
|
163 |
for instance in stream:
|
164 |
refs, pred = instance["references"], instance["prediction"]
|
|
|
|
|
|
|
165 |
|
166 |
-
instance_score = self.
|
167 |
-
|
|
|
|
|
|
|
168 |
if "score" not in instance:
|
169 |
instance["score"] = {"global": global_score, "instance": {}}
|
170 |
else:
|
@@ -182,23 +400,28 @@ class InstanceMetric(SingleStreamOperator, Metric):
|
|
182 |
if reduction == "mean":
|
183 |
from statistics import mean
|
184 |
|
185 |
-
for
|
186 |
-
|
187 |
-
|
188 |
-
|
|
|
|
|
|
|
|
|
189 |
global_score["score_name"] = self.main_score
|
190 |
|
|
|
|
|
|
|
|
|
|
|
191 |
for instance in instances:
|
192 |
yield instance
|
193 |
|
194 |
-
def _compute(self, references: List[str], prediction: str) -> dict:
|
195 |
-
result = self.compute(references=references, prediction=prediction)
|
196 |
-
result["score"] = result[self.main_score]
|
197 |
-
result["score_name"] = self.main_score
|
198 |
-
return result
|
199 |
-
|
200 |
@abstractmethod
|
201 |
-
def compute(
|
|
|
|
|
202 |
pass
|
203 |
|
204 |
|
@@ -208,46 +431,54 @@ class Squad(GlobalMetric):
|
|
208 |
metric = "squad"
|
209 |
|
210 |
def prepare(self):
|
211 |
-
super(
|
212 |
self._metric = evaluate.load(self.metric)
|
213 |
|
214 |
-
def compute(
|
|
|
|
|
|
|
|
|
|
|
215 |
ids = [str(uuid.uuid4()).replace("-", "") for _ in range(len(predictions))]
|
216 |
formatted_predictions = [
|
217 |
-
{"prediction_text": prediction, "id": ids[i]}
|
|
|
218 |
]
|
219 |
formatted_references = [
|
220 |
{"answers": {"answer_start": [-1], "text": reference}, "id": ids[i]}
|
221 |
for i, reference in enumerate(references)
|
222 |
]
|
223 |
|
224 |
-
return self._metric.compute(
|
225 |
-
|
226 |
-
|
227 |
-
|
228 |
-
def _compute(self, references: List[str], prediction: str) -> dict:
|
229 |
-
result = self.compute(references[0], prediction)
|
230 |
-
result["score"] = result[self.main_score]
|
231 |
-
result["score_name"] = self.main_score
|
232 |
-
return result
|
233 |
-
|
234 |
-
@abstractmethod
|
235 |
-
def compute(self, reference, prediction: str) -> dict:
|
236 |
-
pass
|
237 |
|
238 |
|
239 |
-
class Accuracy(
|
240 |
reduction_map = {"mean": ["accuracy"]}
|
241 |
main_score = "accuracy"
|
242 |
|
243 |
-
def compute(
|
244 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
245 |
|
246 |
|
247 |
class MetricPipeline(MultiStreamOperator, Metric):
|
248 |
main_score: str = None
|
249 |
preprocess_steps: Optional[List[StreamingOperator]] = field(default_factory=list)
|
250 |
-
postpreprocess_steps: Optional[List[StreamingOperator]] = field(
|
|
|
|
|
251 |
metric: Metric = None
|
252 |
|
253 |
def verify(self):
|
@@ -269,14 +500,15 @@ class MetricPipeline(MultiStreamOperator, Metric):
|
|
269 |
multi_stream = self.metric(multi_stream)
|
270 |
for step in self.postpreprocess_steps:
|
271 |
multi_stream = step(multi_stream)
|
272 |
-
|
273 |
-
return multi_stream
|
274 |
|
275 |
|
276 |
class HuggingfaceMetric(GlobalMetric):
|
277 |
hf_metric_name: str = None
|
278 |
main_score: str = None # The main score returned from the metric
|
279 |
-
hf_main_score: str =
|
|
|
|
|
280 |
|
281 |
scale: float = 1.0 # optional scaling of main results
|
282 |
scaled_fields: list = None
|
@@ -285,24 +517,39 @@ class HuggingfaceMetric(GlobalMetric):
|
|
285 |
|
286 |
def prepare(self):
|
287 |
super().prepare()
|
288 |
-
self.metric = evaluate.load(
|
|
|
|
|
289 |
|
290 |
-
def compute(
|
291 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
292 |
if self.hf_main_score:
|
293 |
result[self.main_score] = result[self.hf_main_score]
|
294 |
del result[self.hf_main_score]
|
295 |
if self.scale != 1.0:
|
296 |
-
assert
|
|
|
|
|
297 |
for key in self.scaled_fields:
|
298 |
-
assert
|
|
|
|
|
299 |
if isinstance(result[key], list):
|
300 |
assert all(
|
301 |
isinstance(v, float) for v in result[key]
|
302 |
), "Not all scaled field '{key}' values are floats: {result[key]}"
|
303 |
result[key] = [v / self.scale for v in result[key]]
|
304 |
else:
|
305 |
-
assert isinstance(
|
|
|
|
|
306 |
result[key] /= self.scale
|
307 |
return result
|
308 |
|
@@ -317,8 +564,15 @@ class HuggingfaceBulkMetric(BulkInstanceMetric):
|
|
317 |
super().prepare()
|
318 |
self.metric = evaluate.load(self.hf_metric_name)
|
319 |
|
320 |
-
def compute(
|
321 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
322 |
|
323 |
# convert dict of lists to a list of dicts
|
324 |
results = [{} for _ in range(len(scores[self.hf_metric_fields[0]]))]
|
@@ -337,7 +591,7 @@ class F1(GlobalMetric):
|
|
337 |
metric = "f1"
|
338 |
|
339 |
def prepare(self):
|
340 |
-
super(
|
341 |
self._metric = evaluate.load(self.metric)
|
342 |
|
343 |
def get_str_id(self, str):
|
@@ -347,18 +601,30 @@ class F1(GlobalMetric):
|
|
347 |
self.id_to_str[id] = str
|
348 |
return self.str_to_id[str]
|
349 |
|
350 |
-
def compute(
|
|
|
|
|
|
|
|
|
|
|
351 |
assert all(
|
352 |
len(reference) == 1 for reference in references
|
353 |
), "Only a single reference per prediction is allowed in F1 metric"
|
354 |
self.str_to_id = {}
|
355 |
self.id_to_str = {}
|
356 |
-
formatted_references = [
|
357 |
-
|
358 |
-
|
|
|
|
|
|
|
|
|
359 |
labels = list(set(formatted_references))
|
360 |
result = self._metric.compute(
|
361 |
-
predictions=formatted_predictions,
|
|
|
|
|
|
|
362 |
)
|
363 |
if isinstance(result["f1"], numpy.ndarray):
|
364 |
from statistics import mean
|
@@ -380,6 +646,11 @@ class F1Macro(F1):
|
|
380 |
main_score = "f1_macro"
|
381 |
|
382 |
|
|
|
|
|
|
|
|
|
|
|
383 |
class F1MultiLabel(GlobalMetric):
|
384 |
_metric = None
|
385 |
main_score = "f1_macro"
|
@@ -387,11 +658,11 @@ class F1MultiLabel(GlobalMetric):
|
|
387 |
classes_to_ignore = ["none"]
|
388 |
|
389 |
def prepare(self):
|
390 |
-
super(
|
391 |
self._metric = evaluate.load("f1", "multilabel")
|
392 |
|
393 |
def add_str_to_id(self, str):
|
394 |
-
if not
|
395 |
id = len(self.str_to_id)
|
396 |
self.str_to_id[str] = id
|
397 |
self.id_to_str[id] = str
|
@@ -404,17 +675,34 @@ class F1MultiLabel(GlobalMetric):
|
|
404 |
result[self.str_to_id[label]] = 1
|
405 |
return result
|
406 |
|
407 |
-
def compute(
|
|
|
|
|
|
|
|
|
|
|
408 |
self.str_to_id = {}
|
409 |
self.id_to_str = {}
|
410 |
assert all(
|
411 |
len(reference) == 1 for reference in references
|
412 |
-
), "Only a single reference per prediction is allowed in F1 metric"
|
|
|
413 |
references = [reference[0] for reference in references]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
414 |
labels = [
|
415 |
-
|
416 |
-
for
|
417 |
-
if
|
418 |
]
|
419 |
# if no classes are left then F1 is not defined
|
420 |
# (e.g. only "none" in references)
|
@@ -423,8 +711,12 @@ class F1MultiLabel(GlobalMetric):
|
|
423 |
|
424 |
for label in labels:
|
425 |
self.add_str_to_id(label)
|
426 |
-
formatted_references = [
|
427 |
-
|
|
|
|
|
|
|
|
|
428 |
|
429 |
# There is odd behavior in scikit-learn that when passing a one-hot vector with a single
|
430 |
# element, it is treated a class identifier. Therefore, we add labels=[1] to limit to only
|
@@ -475,37 +767,53 @@ class Rouge(HuggingfaceMetric):
|
|
475 |
sent_split_newline: bool = True
|
476 |
|
477 |
def prepare(self):
|
478 |
-
self.hf_compute_args.update({"use_aggregator": self.use_aggregator, "rouge_types": self.rouge_types})
|
479 |
-
|
480 |
super().prepare()
|
|
|
|
|
|
|
|
|
|
|
481 |
import nltk
|
482 |
|
483 |
nltk.download("punkt")
|
484 |
self.sent_tokenize = nltk.sent_tokenize
|
485 |
|
486 |
-
def compute(self, references, predictions):
|
487 |
if self.sent_split_newline:
|
488 |
-
predictions = [
|
489 |
-
|
490 |
-
|
491 |
-
|
492 |
-
|
493 |
-
|
494 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
495 |
reduction_map = {"mean": ["char_edit_dist_accuracy"]}
|
496 |
main_score = "char_edit_dist_accuracy"
|
497 |
|
498 |
def prepare(self):
|
|
|
499 |
import editdistance
|
500 |
|
501 |
self.eval = editdistance.eval
|
502 |
|
503 |
-
def compute(
|
|
|
|
|
|
|
|
|
|
|
|
|
504 |
formatted_prediction = "".join(prediction.split())
|
505 |
-
formatted_reference = "".join(
|
506 |
max_length = max(len(formatted_reference), len(formatted_prediction))
|
507 |
if max_length == 0:
|
508 |
-
return 0
|
509 |
edit_dist = self.eval(formatted_reference, formatted_prediction)
|
510 |
return {"char_edit_dist_accuracy": (1 - edit_dist / max_length)}
|
511 |
|
@@ -514,12 +822,19 @@ class Wer(HuggingfaceMetric):
|
|
514 |
hf_metric_name = "wer"
|
515 |
main_score = "wer"
|
516 |
|
517 |
-
def compute(
|
|
|
|
|
|
|
|
|
|
|
518 |
assert all(
|
519 |
len(reference) == 1 for reference in references
|
520 |
), "Only single reference per prediction is allowed in wer metric"
|
521 |
formatted_references = [reference[0] for reference in references]
|
522 |
-
result = self.metric.compute(
|
|
|
|
|
523 |
return {self.main_score: result}
|
524 |
|
525 |
|
@@ -534,16 +849,27 @@ class MatthewsCorrelation(HuggingfaceMetric):
|
|
534 |
self.str_to_id[str] = id
|
535 |
return self.str_to_id[str]
|
536 |
|
537 |
-
def compute(
|
538 |
-
|
539 |
-
|
540 |
-
|
541 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
542 |
|
543 |
|
544 |
class CustomF1(GlobalMetric):
|
545 |
main_score = "f1_micro"
|
546 |
classes = None
|
|
|
547 |
|
548 |
@abstractmethod
|
549 |
def get_element_group(self, element):
|
@@ -553,40 +879,64 @@ class CustomF1(GlobalMetric):
|
|
553 |
def get_element_representation(self, element):
|
554 |
pass
|
555 |
|
556 |
-
def group_elements(self,
|
557 |
return {
|
558 |
-
k: Counter(
|
559 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
560 |
}
|
561 |
|
562 |
def calculate_groups_ratio(self, actual_group, total_group):
|
563 |
-
return sum(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
564 |
|
565 |
def f1(self, pn, pd, rn, rd):
|
566 |
-
precision =
|
567 |
-
recall =
|
568 |
try:
|
569 |
return 2 * precision * recall / (precision + recall)
|
570 |
except ZeroDivisionError:
|
571 |
-
return
|
572 |
-
|
573 |
-
def compute(
|
|
|
|
|
|
|
|
|
|
|
574 |
# in case reference are List[List[List[Any]]] and predictions are List[List[Any]]:
|
575 |
if isinstance(references[0], list) and isinstance(references[0][0], list):
|
576 |
references = [element[0] for element in references]
|
577 |
|
578 |
assert len(references) == len(predictions), (
|
579 |
-
f"references size ({len(references)})"
|
|
|
580 |
)
|
581 |
if self.classes is None:
|
582 |
-
classes =
|
|
|
|
|
583 |
else:
|
584 |
classes = self.classes
|
585 |
-
groups_statistics =
|
586 |
for references_batch, predictions_batch in zip(references, predictions):
|
587 |
grouped_references = self.group_elements(references_batch)
|
588 |
grouped_predictions = self.group_elements(predictions_batch)
|
589 |
-
all_groups = set(grouped_references.keys()).union(
|
|
|
|
|
590 |
for group in all_groups:
|
591 |
if group not in groups_statistics:
|
592 |
groups_statistics[group] = {
|
@@ -608,9 +958,11 @@ class CustomF1(GlobalMetric):
|
|
608 |
groups_statistics[group]["recall_numerator"] += rn
|
609 |
groups_statistics[group]["recall_denominator"] += rd
|
610 |
|
611 |
-
result = {}
|
612 |
num_of_unknown_class_predictions = 0
|
613 |
pn_total = pd_total = rn_total = rd_total = 0
|
|
|
|
|
|
|
614 |
for group in groups_statistics.keys():
|
615 |
pn, pd, rn, rd = (
|
616 |
groups_statistics[group]["precision_numerator"],
|
@@ -618,22 +970,45 @@ class CustomF1(GlobalMetric):
|
|
618 |
groups_statistics[group]["recall_numerator"],
|
619 |
groups_statistics[group]["recall_denominator"],
|
620 |
)
|
621 |
-
pn_total, pd_total, rn_total, rd_total =
|
|
|
|
|
|
|
|
|
|
|
622 |
if group in classes:
|
623 |
-
|
|
|
|
|
624 |
else:
|
625 |
num_of_unknown_class_predictions += pd
|
|
|
|
|
626 |
try:
|
627 |
-
result["f1_macro"] = sum(
|
|
|
|
|
|
|
|
|
|
|
|
|
628 |
except ZeroDivisionError:
|
629 |
-
result["f1_macro"] =
|
|
|
|
|
630 |
|
631 |
amount_of_predictions = pd_total
|
632 |
if amount_of_predictions == 0:
|
633 |
result["in_classes_support"] = 1.0
|
634 |
else:
|
635 |
-
result["in_classes_support"] =
|
636 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
637 |
return result
|
638 |
|
639 |
|
@@ -668,11 +1043,20 @@ class TokenOverlap(InstanceMetric):
|
|
668 |
reduction_map = {"mean": ["f1", "precision", "recall"]}
|
669 |
main_score = "f1"
|
670 |
|
671 |
-
def compute(
|
672 |
-
|
673 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
674 |
|
675 |
-
def _compute_single_ref(
|
|
|
|
|
676 |
prediction_tokens = normalize_answer(prediction).split()
|
677 |
reference_tokens = normalize_answer(reference).split()
|
678 |
common = Counter(prediction_tokens) & Counter(reference_tokens)
|
@@ -713,7 +1097,12 @@ class SentenceBert(BulkInstanceMetric):
|
|
713 |
self.model = SentenceTransformer(self.model_name)
|
714 |
self.util = sbert_util
|
715 |
|
716 |
-
def compute(
|
|
|
|
|
|
|
|
|
|
|
717 |
scores = []
|
718 |
|
719 |
# we are in a multi-reference case (each prediction may have multiple
|
@@ -728,7 +1117,9 @@ class SentenceBert(BulkInstanceMetric):
|
|
728 |
|
729 |
# compute s-bert embeddings
|
730 |
preds_emb = self.model.encode(predictions)
|
731 |
-
refs_emb = self.model.encode(
|
|
|
|
|
732 |
|
733 |
# for each candidate, pick the reference with the highest score
|
734 |
for pred_emb, ref_group_bounds in zip(preds_emb, ref_group_boundaries):
|
@@ -746,11 +1137,17 @@ class Reward(BulkInstanceMetric):
|
|
746 |
model_name: str
|
747 |
|
748 |
def prepare(self):
|
|
|
749 |
from transformers import pipeline
|
750 |
|
751 |
self.pipe = pipeline("text-classification", model=self.model_name)
|
752 |
|
753 |
-
def compute(
|
|
|
|
|
|
|
|
|
|
|
754 |
# treat the references as the questions and the predictions as answers
|
755 |
# assume a single reference
|
756 |
questions = [refs[0] for refs in references]
|
@@ -762,3 +1159,67 @@ class Reward(BulkInstanceMetric):
|
|
762 |
# compute the metric
|
763 |
# add function_to_apply="none" to disable sigmoid
|
764 |
return self.pipe(inputs, batch_size=self.batch_size)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import logging
|
2 |
import re
|
3 |
import string
|
4 |
import uuid
|
5 |
+
from abc import abstractmethod
|
6 |
from collections import Counter
|
7 |
from dataclasses import field
|
8 |
from typing import Any, Dict, Generator, List, Optional, Tuple
|
9 |
|
10 |
import evaluate
|
11 |
import numpy
|
12 |
+
import numpy as np
|
13 |
+
from scipy.stats import bootstrap
|
14 |
|
15 |
+
from .artifact import Artifact
|
16 |
from .dataclass import InternalField, OptionalField
|
17 |
from .operator import (
|
18 |
MultiStreamOperator,
|
|
|
21 |
StreamInstanceOperator,
|
22 |
)
|
23 |
from .operators import CopyFields
|
24 |
+
from .random_utils import get_seed
|
25 |
from .stream import MultiStream, Stream
|
26 |
|
27 |
+
# The default number of resamples used to estimate the confidence intervals
|
28 |
+
# global and instances metrics. Use None to disable confidence interval computation by default.
|
29 |
+
_N_RESAMPLES_DEFAULT_FOR_INSTANCE_METRICS = 1000
|
30 |
+
_N_RESAMPLES_DEFAULT_FOR_GLOBAL_METRICS = 100
|
31 |
+
|
32 |
|
33 |
def abstract_factory():
|
34 |
return {}
|
|
|
41 |
class UpdateStream(StreamInstanceOperator):
|
42 |
update: dict
|
43 |
|
44 |
+
def process(
|
45 |
+
self, instance: Dict[str, Any], stream_name: Optional[str] = None
|
46 |
+
) -> Dict[str, Any]:
|
47 |
instance.update(self.update)
|
48 |
return instance
|
49 |
|
50 |
|
51 |
# TODO: currently we have two classes with this name. metric.Metric and matrics.Metric...
|
52 |
+
class Metric(Artifact):
|
53 |
@property
|
54 |
@abstractmethod
|
55 |
def main_score(self):
|
56 |
pass
|
57 |
|
58 |
|
59 |
+
class MetricWithConfidenceInterval(Metric):
|
60 |
+
# The number of resamples used to estimate the confidence intervals of this metric.
|
61 |
+
# Use None to disable confidence interval computation.
|
62 |
+
n_resamples: int = None
|
63 |
+
confidence_level: float = 0.95
|
64 |
+
|
65 |
+
@staticmethod
|
66 |
+
def new_random_generator():
|
67 |
+
# The np.random.default_rng expects a 32-bit int, while hash(..) can return a 64-bit integer.
|
68 |
+
# So use '& MAX_32BIT' to get a 32-bit seed.
|
69 |
+
_max_32bit = 2**32 - 1
|
70 |
+
return np.random.default_rng(hash(get_seed()) & _max_32bit)
|
71 |
+
|
72 |
+
def disable_confidence_interval_calculation(self):
|
73 |
+
self.n_resamples = None
|
74 |
+
|
75 |
+
def _can_compute_confidence_intervals(self, num_predictions):
|
76 |
+
return (
|
77 |
+
self.n_resamples is not None
|
78 |
+
and self.n_resamples > 1
|
79 |
+
and num_predictions > 1
|
80 |
+
)
|
81 |
+
|
82 |
+
def score_based_confidence_interval(self, score_names: List[str], instances):
|
83 |
+
"""Compute confidence intervals based on existing scores, already computed on the input instances.
|
84 |
+
|
85 |
+
score_names: List[str]
|
86 |
+
Compute a confidence interval for each score_name from this list.
|
87 |
+
instances:
|
88 |
+
The instances for which the confidence intervals are computed.
|
89 |
+
"""
|
90 |
+
from statistics import mean
|
91 |
+
|
92 |
+
result = {}
|
93 |
+
|
94 |
+
if not self._can_compute_confidence_intervals(num_predictions=len(instances)):
|
95 |
+
return result
|
96 |
+
|
97 |
+
for score_name in score_names:
|
98 |
+
scores = [
|
99 |
+
instance["score"]["instance"][score_name] for instance in instances
|
100 |
+
]
|
101 |
+
ci = bootstrap(
|
102 |
+
(scores,),
|
103 |
+
statistic=mean,
|
104 |
+
n_resamples=self.n_resamples,
|
105 |
+
confidence_level=self.confidence_level,
|
106 |
+
random_state=self.new_random_generator(),
|
107 |
+
).confidence_interval
|
108 |
+
result[f"{score_name}_ci_low"] = ci.low
|
109 |
+
result[f"{score_name}_ci_high"] = ci.high
|
110 |
+
if score_name == self.main_score:
|
111 |
+
result["score_ci_low"] = ci.low
|
112 |
+
result["score_ci_high"] = ci.high
|
113 |
+
return result
|
114 |
+
|
115 |
+
def compute_global_confidence_intervals(
|
116 |
+
self, references, predictions, additional_inputs, score_name
|
117 |
+
):
|
118 |
+
"""Computed confidence intervals for a set of references and predictions."""
|
119 |
+
random_gen = self.new_random_generator()
|
120 |
+
|
121 |
+
def statistic(arr, axis):
|
122 |
+
# arr is a 2d array where each row is a resampling, so we
|
123 |
+
# iterate over the rows and compute the metric on each resampling
|
124 |
+
def metric(sample_refs, sample_preds, sample_additional_inputs):
|
125 |
+
try:
|
126 |
+
return self._compute(
|
127 |
+
references=sample_refs,
|
128 |
+
predictions=sample_preds,
|
129 |
+
additional_inputs=sample_additional_inputs,
|
130 |
+
)["score"]
|
131 |
+
except Exception as e:
|
132 |
+
# this happens in edge cases, for example, when the sampling creates a
|
133 |
+
# sample where all strings are empty and this fails bleu.
|
134 |
+
logging.info(f"Warning in {self.__class__.__name__}", e)
|
135 |
+
return np.nan
|
136 |
+
|
137 |
+
scores = numpy.apply_along_axis(
|
138 |
+
lambda x: metric(
|
139 |
+
sample_refs=[references[i] for i in x],
|
140 |
+
sample_preds=[predictions[i] for i in x],
|
141 |
+
sample_additional_inputs=[additional_inputs[i] for i in x],
|
142 |
+
),
|
143 |
+
axis=axis,
|
144 |
+
arr=arr,
|
145 |
+
)
|
146 |
+
|
147 |
+
# when running with bca interval (default), the statistic is called twice: with the
|
148 |
+
# original data and with the resamples. here we want to focus only on the latter.
|
149 |
+
if scores.size > 1:
|
150 |
+
# here we deal with samples on which the metric could not be computed. These are
|
151 |
+
# edge cases - for example, when the sample contains only empty strings.
|
152 |
+
# CI is about the distribution around the statistic (e.g. mean), it doesn't deal with
|
153 |
+
# cases in which the metric is not computable. Therefore, we ignore these edge cases
|
154 |
+
# as part of the computation of CI. The question is how to implement this policy.
|
155 |
+
# Options:
|
156 |
+
# 1. skip the errors and return a shorter array => this fails because Scipy demans
|
157 |
+
# this callback (i.e. the statistic() callback) to return an array of the same size
|
158 |
+
# as the number of resamples
|
159 |
+
# 2. Put np.nan for the errors => this fails because in such case the ci itself
|
160 |
+
# becomes np.nan. So one edge case can fail the whole CI computation.
|
161 |
+
# 3. Replace the errors with a sampling from the successful cases => this is what
|
162 |
+
# is implemented.
|
163 |
+
error_indices = numpy.isnan(scores)
|
164 |
+
n_errors = sum(error_indices)
|
165 |
+
if n_errors > 0:
|
166 |
+
new_scores = random_gen.choice(scores, n_errors, replace=True)
|
167 |
+
scores = scores[~error_indices]
|
168 |
+
scores = np.concatenate([scores, new_scores])
|
169 |
+
|
170 |
+
return scores
|
171 |
+
|
172 |
+
result = {}
|
173 |
+
num_predictions = len(predictions)
|
174 |
+
if self._can_compute_confidence_intervals(num_predictions=num_predictions):
|
175 |
+
identifiers = list(range(num_predictions))
|
176 |
+
ci = bootstrap(
|
177 |
+
(identifiers,),
|
178 |
+
statistic=statistic,
|
179 |
+
n_resamples=self.n_resamples,
|
180 |
+
confidence_level=self.confidence_level,
|
181 |
+
random_state=random_gen,
|
182 |
+
).confidence_interval
|
183 |
+
result["score_ci_low"] = ci.low
|
184 |
+
result["score_ci_high"] = ci.high
|
185 |
+
result[f"{score_name}_ci_low"] = ci.low
|
186 |
+
result[f"{score_name}_ci_high"] = ci.high
|
187 |
+
return result
|
188 |
+
|
189 |
+
|
190 |
+
class GlobalMetric(SingleStreamOperator, MetricWithConfidenceInterval):
|
191 |
+
"""A class for computing metrics that require joint calculations over all instances and are not just aggregation of scores of individuals instances.
|
192 |
+
|
193 |
+
For example, macro_F1 requires
|
194 |
+
calculation requires calculation of recall and precision per class, so all instances of the class
|
195 |
+
need to be considered. Accuracy, on the other hand, is just an average of the accuracy of all the instances.
|
196 |
+
"""
|
197 |
+
|
198 |
+
n_resamples = _N_RESAMPLES_DEFAULT_FOR_GLOBAL_METRICS
|
199 |
+
|
200 |
+
def process(self, stream: Stream, stream_name: Optional[str] = None) -> Generator:
|
201 |
references = []
|
202 |
predictions = []
|
203 |
+
additional_inputs = []
|
204 |
global_score = {}
|
205 |
|
206 |
instances = []
|
|
|
211 |
else:
|
212 |
global_score = instance["score"]["global"]
|
213 |
|
214 |
+
instance_references, instance_prediction = (
|
215 |
+
instance["references"],
|
216 |
+
instance["prediction"],
|
217 |
+
)
|
218 |
+
references.append(instance_references)
|
219 |
+
predictions.append(instance_prediction)
|
220 |
+
instances.append(instance)
|
221 |
|
222 |
+
instance_additional_inputs = (
|
223 |
+
instance["additional_inputs"] if "additional_inputs" in instance else {}
|
224 |
+
)
|
225 |
+
additional_inputs.append(instance_additional_inputs)
|
226 |
try:
|
227 |
+
instance_score = self._compute(
|
228 |
+
[instance_references],
|
229 |
+
[instance_prediction],
|
230 |
+
[instance_additional_inputs],
|
231 |
+
)
|
232 |
except:
|
233 |
instance_score = {"score": None, "score_name": self.main_score}
|
234 |
|
235 |
+
if isinstance(self.main_score, str):
|
236 |
instance_score[self.main_score] = None
|
237 |
|
238 |
instance["score"]["instance"].update(instance_score)
|
239 |
|
240 |
+
result = self._compute(references, predictions, additional_inputs)
|
|
|
|
|
|
|
|
|
241 |
|
242 |
global_score.update(result)
|
243 |
|
244 |
+
score_name = global_score["score_name"]
|
245 |
+
confidence_interval = self.compute_global_confidence_intervals(
|
246 |
+
references, predictions, additional_inputs, score_name
|
247 |
+
)
|
248 |
+
global_score.update(confidence_interval)
|
249 |
+
|
250 |
for instance in instances:
|
251 |
instance["score"]["global"] = global_score
|
252 |
yield instance
|
253 |
|
254 |
+
def _compute(
|
255 |
+
self,
|
256 |
+
references: List[List[str]],
|
257 |
+
predictions: List[str],
|
258 |
+
additional_inputs: List[Any],
|
259 |
+
) -> dict:
|
260 |
+
result = self.compute(references, predictions, additional_inputs)
|
261 |
result["score"] = result[self.main_score]
|
262 |
result["score_name"] = self.main_score
|
263 |
return result
|
264 |
|
265 |
@abstractmethod
|
266 |
+
def compute(
|
267 |
+
self,
|
268 |
+
references: List[List[Any]],
|
269 |
+
predictions: List[Any],
|
270 |
+
additional_inputs: List[Any],
|
271 |
+
) -> dict:
|
272 |
pass
|
273 |
|
274 |
|
275 |
+
class BulkInstanceMetric(SingleStreamOperator, MetricWithConfidenceInterval):
|
276 |
+
n_resamples = _N_RESAMPLES_DEFAULT_FOR_INSTANCE_METRICS
|
277 |
main_score: str
|
278 |
reduction_map: Dict[str, List[str]]
|
279 |
|
280 |
implemented_reductions: List[str] = field(default_factory=lambda: ["mean"])
|
281 |
|
282 |
+
def process(self, stream: Stream, stream_name: Optional[str] = None) -> Generator:
|
283 |
global_score = {}
|
284 |
instances = []
|
285 |
|
286 |
# consume the stream
|
287 |
references, predictions = map(
|
288 |
+
list,
|
289 |
+
zip(
|
290 |
+
*[
|
291 |
+
(instance["references"], instance["prediction"])
|
292 |
+
for instance in stream
|
293 |
+
]
|
294 |
+
),
|
295 |
)
|
296 |
|
297 |
+
additional_inputs = [
|
298 |
+
instance["additional_inputs"] if "additional_inputs" in instance else {}
|
299 |
+
for instance in stream
|
300 |
+
]
|
301 |
+
|
302 |
# compute the metric over all refs and preds
|
303 |
+
instance_scores = self.compute(
|
304 |
+
references=references,
|
305 |
+
predictions=predictions,
|
306 |
+
additional_inputs=additional_inputs,
|
307 |
+
)
|
308 |
|
309 |
# add the score and score_name fields
|
310 |
for instance_score in instance_scores:
|
|
|
329 |
if reduction == "mean":
|
330 |
from statistics import mean
|
331 |
|
332 |
+
for field_name in fields:
|
333 |
+
global_score[field_name] = mean(
|
334 |
+
[
|
335 |
+
instance["score"]["instance"][field_name]
|
336 |
+
for instance in instances
|
337 |
+
]
|
338 |
+
)
|
339 |
+
if field_name == self.main_score:
|
340 |
+
global_score["score"] = global_score[field_name]
|
341 |
global_score["score_name"] = self.main_score
|
342 |
|
343 |
+
confidence_interval = self.score_based_confidence_interval(
|
344 |
+
score_names=[self.main_score], instances=instances
|
345 |
+
)
|
346 |
+
global_score.update(confidence_interval)
|
347 |
+
|
348 |
for instance in instances:
|
349 |
yield instance
|
350 |
|
351 |
@abstractmethod
|
352 |
+
def compute(
|
353 |
+
self,
|
354 |
+
references: List[List[Any]],
|
355 |
+
predictions: List[Any],
|
356 |
+
additional_inputs: List[Dict],
|
357 |
+
) -> Dict[str, Any]:
|
358 |
pass
|
359 |
|
360 |
|
361 |
+
class InstanceMetric(SingleStreamOperator, MetricWithConfidenceInterval):
|
362 |
+
n_resamples = _N_RESAMPLES_DEFAULT_FOR_INSTANCE_METRICS
|
363 |
+
|
364 |
implemented_reductions: List[str] = field(default_factory=lambda: ["mean"])
|
365 |
|
366 |
@property
|
|
|
368 |
def reduction_map(self) -> dict:
|
369 |
pass
|
370 |
|
371 |
+
def process(self, stream: Stream, stream_name: Optional[str] = None) -> Generator:
|
372 |
global_score = {}
|
373 |
instances = []
|
374 |
|
375 |
for instance in stream:
|
376 |
refs, pred = instance["references"], instance["prediction"]
|
377 |
+
additional_inputs = (
|
378 |
+
instance["additional_inputs"] if "additional_inputs" in instance else {}
|
379 |
+
)
|
380 |
|
381 |
+
instance_score = self.compute(
|
382 |
+
references=refs, prediction=pred, additional_inputs=additional_inputs
|
383 |
+
)
|
384 |
+
instance_score["score"] = instance_score[self.main_score]
|
385 |
+
instance_score["score_name"] = self.main_score
|
386 |
if "score" not in instance:
|
387 |
instance["score"] = {"global": global_score, "instance": {}}
|
388 |
else:
|
|
|
400 |
if reduction == "mean":
|
401 |
from statistics import mean
|
402 |
|
403 |
+
for field_name in fields:
|
404 |
+
scores = [
|
405 |
+
instance["score"]["instance"][field_name]
|
406 |
+
for instance in instances
|
407 |
+
]
|
408 |
+
global_score[field_name] = mean(scores)
|
409 |
+
if field_name == self.main_score:
|
410 |
+
global_score["score"] = global_score[field_name]
|
411 |
global_score["score_name"] = self.main_score
|
412 |
|
413 |
+
confidence_interval = self.score_based_confidence_interval(
|
414 |
+
score_names=[self.main_score], instances=instances
|
415 |
+
)
|
416 |
+
global_score.update(confidence_interval)
|
417 |
+
|
418 |
for instance in instances:
|
419 |
yield instance
|
420 |
|
|
|
|
|
|
|
|
|
|
|
|
|
421 |
@abstractmethod
|
422 |
+
def compute(
|
423 |
+
self, references: List[Any], prediction: Any, additional_inputs: Dict
|
424 |
+
) -> dict:
|
425 |
pass
|
426 |
|
427 |
|
|
|
431 |
metric = "squad"
|
432 |
|
433 |
def prepare(self):
|
434 |
+
super().prepare()
|
435 |
self._metric = evaluate.load(self.metric)
|
436 |
|
437 |
+
def compute(
|
438 |
+
self,
|
439 |
+
references: List[List[str]],
|
440 |
+
predictions: List[str],
|
441 |
+
additional_inputs: List[Dict],
|
442 |
+
) -> dict:
|
443 |
ids = [str(uuid.uuid4()).replace("-", "") for _ in range(len(predictions))]
|
444 |
formatted_predictions = [
|
445 |
+
{"prediction_text": prediction, "id": ids[i]}
|
446 |
+
for i, prediction in enumerate(predictions)
|
447 |
]
|
448 |
formatted_references = [
|
449 |
{"answers": {"answer_start": [-1], "text": reference}, "id": ids[i]}
|
450 |
for i, reference in enumerate(references)
|
451 |
]
|
452 |
|
453 |
+
return self._metric.compute(
|
454 |
+
predictions=formatted_predictions,
|
455 |
+
references=formatted_references,
|
456 |
+
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
457 |
|
458 |
|
459 |
+
class Accuracy(InstanceMetric):
|
460 |
reduction_map = {"mean": ["accuracy"]}
|
461 |
main_score = "accuracy"
|
462 |
|
463 |
+
def compute(
|
464 |
+
self, references: List[Any], prediction: Any, additional_inputs: List[Dict]
|
465 |
+
) -> dict:
|
466 |
+
result = {
|
467 |
+
self.main_score: float(
|
468 |
+
str(prediction) in [str(reference) for reference in references]
|
469 |
+
)
|
470 |
+
}
|
471 |
+
result["score"] = result[self.main_score]
|
472 |
+
result["score_name"] = self.main_score
|
473 |
+
return result
|
474 |
|
475 |
|
476 |
class MetricPipeline(MultiStreamOperator, Metric):
|
477 |
main_score: str = None
|
478 |
preprocess_steps: Optional[List[StreamingOperator]] = field(default_factory=list)
|
479 |
+
postpreprocess_steps: Optional[List[StreamingOperator]] = field(
|
480 |
+
default_factory=list
|
481 |
+
)
|
482 |
metric: Metric = None
|
483 |
|
484 |
def verify(self):
|
|
|
500 |
multi_stream = self.metric(multi_stream)
|
501 |
for step in self.postpreprocess_steps:
|
502 |
multi_stream = step(multi_stream)
|
503 |
+
return self.prepare_score(multi_stream)
|
|
|
504 |
|
505 |
|
506 |
class HuggingfaceMetric(GlobalMetric):
|
507 |
hf_metric_name: str = None
|
508 |
main_score: str = None # The main score returned from the metric
|
509 |
+
hf_main_score: str = (
|
510 |
+
None # USed if HF returns uses a different score name for the main metric
|
511 |
+
)
|
512 |
|
513 |
scale: float = 1.0 # optional scaling of main results
|
514 |
scaled_fields: list = None
|
|
|
517 |
|
518 |
def prepare(self):
|
519 |
super().prepare()
|
520 |
+
self.metric = evaluate.load(
|
521 |
+
self.hf_metric_name, experiment_id=self.experiment_id
|
522 |
+
)
|
523 |
|
524 |
+
def compute(
|
525 |
+
self,
|
526 |
+
references: List[List[Any]],
|
527 |
+
predictions: List[Any],
|
528 |
+
additional_inputs: List[Dict],
|
529 |
+
) -> dict:
|
530 |
+
result = self.metric.compute(
|
531 |
+
predictions=predictions, references=references, **self.hf_compute_args
|
532 |
+
)
|
533 |
if self.hf_main_score:
|
534 |
result[self.main_score] = result[self.hf_main_score]
|
535 |
del result[self.hf_main_score]
|
536 |
if self.scale != 1.0:
|
537 |
+
assert (
|
538 |
+
self.scaled_fields is not None
|
539 |
+
), f"Scaling factor was set to {self.scale}, but no fields specified"
|
540 |
for key in self.scaled_fields:
|
541 |
+
assert (
|
542 |
+
key in result
|
543 |
+
), f"Trying to scale field '{key}' which is not in results of metrics: {result}"
|
544 |
if isinstance(result[key], list):
|
545 |
assert all(
|
546 |
isinstance(v, float) for v in result[key]
|
547 |
), "Not all scaled field '{key}' values are floats: {result[key]}"
|
548 |
result[key] = [v / self.scale for v in result[key]]
|
549 |
else:
|
550 |
+
assert isinstance(
|
551 |
+
result[key], float
|
552 |
+
), "Scaled field '{key}' is not float: {result[key]}"
|
553 |
result[key] /= self.scale
|
554 |
return result
|
555 |
|
|
|
564 |
super().prepare()
|
565 |
self.metric = evaluate.load(self.hf_metric_name)
|
566 |
|
567 |
+
def compute(
|
568 |
+
self,
|
569 |
+
references: List[List[str]],
|
570 |
+
predictions: List[str],
|
571 |
+
additional_inputs: List[Any],
|
572 |
+
) -> List[Dict[str, Any]]:
|
573 |
+
scores = self.metric.compute(
|
574 |
+
predictions=predictions, references=references, **self.hf_compute_args
|
575 |
+
)
|
576 |
|
577 |
# convert dict of lists to a list of dicts
|
578 |
results = [{} for _ in range(len(scores[self.hf_metric_fields[0]]))]
|
|
|
591 |
metric = "f1"
|
592 |
|
593 |
def prepare(self):
|
594 |
+
super().prepare()
|
595 |
self._metric = evaluate.load(self.metric)
|
596 |
|
597 |
def get_str_id(self, str):
|
|
|
601 |
self.id_to_str[id] = str
|
602 |
return self.str_to_id[str]
|
603 |
|
604 |
+
def compute(
|
605 |
+
self,
|
606 |
+
references: List[List[str]],
|
607 |
+
predictions: List[str],
|
608 |
+
additional_inputs: List[Dict],
|
609 |
+
) -> dict:
|
610 |
assert all(
|
611 |
len(reference) == 1 for reference in references
|
612 |
), "Only a single reference per prediction is allowed in F1 metric"
|
613 |
self.str_to_id = {}
|
614 |
self.id_to_str = {}
|
615 |
+
formatted_references = [
|
616 |
+
self.get_str_id(reference[0]) for reference in references
|
617 |
+
]
|
618 |
+
self.str_to_id.keys()
|
619 |
+
formatted_predictions = [
|
620 |
+
self.get_str_id(prediction) for prediction in predictions
|
621 |
+
]
|
622 |
labels = list(set(formatted_references))
|
623 |
result = self._metric.compute(
|
624 |
+
predictions=formatted_predictions,
|
625 |
+
references=formatted_references,
|
626 |
+
labels=labels,
|
627 |
+
average=self.average,
|
628 |
)
|
629 |
if isinstance(result["f1"], numpy.ndarray):
|
630 |
from statistics import mean
|
|
|
646 |
main_score = "f1_macro"
|
647 |
|
648 |
|
649 |
+
class F1Weighted(F1):
|
650 |
+
main_score = "f1_weighted"
|
651 |
+
average = "weighted"
|
652 |
+
|
653 |
+
|
654 |
class F1MultiLabel(GlobalMetric):
|
655 |
_metric = None
|
656 |
main_score = "f1_macro"
|
|
|
658 |
classes_to_ignore = ["none"]
|
659 |
|
660 |
def prepare(self):
|
661 |
+
super().prepare()
|
662 |
self._metric = evaluate.load("f1", "multilabel")
|
663 |
|
664 |
def add_str_to_id(self, str):
|
665 |
+
if str not in self.str_to_id:
|
666 |
id = len(self.str_to_id)
|
667 |
self.str_to_id[str] = id
|
668 |
self.id_to_str[id] = str
|
|
|
675 |
result[self.str_to_id[label]] = 1
|
676 |
return result
|
677 |
|
678 |
+
def compute(
|
679 |
+
self,
|
680 |
+
references: List[List[str]],
|
681 |
+
predictions: List[List[str]],
|
682 |
+
additional_inputs: List[Dict],
|
683 |
+
) -> dict:
|
684 |
self.str_to_id = {}
|
685 |
self.id_to_str = {}
|
686 |
assert all(
|
687 |
len(reference) == 1 for reference in references
|
688 |
+
), "Only a single reference per prediction is allowed in F1 multi label metric"
|
689 |
+
|
690 |
references = [reference[0] for reference in references]
|
691 |
+
|
692 |
+
for reference in references:
|
693 |
+
assert isinstance(
|
694 |
+
references, list
|
695 |
+
), f"Each reference is expected to list of strings in F1 multi label metric. Received reference: {reference}"
|
696 |
+
|
697 |
+
for prediction in predictions:
|
698 |
+
assert isinstance(
|
699 |
+
prediction, list
|
700 |
+
), f"Each prediction is expected to list of strings in F1 multi label metric. Received prediction: {prediction}"
|
701 |
+
|
702 |
labels = [
|
703 |
+
lbl
|
704 |
+
for lbl in {label for reference in references for label in reference}
|
705 |
+
if lbl not in self.classes_to_ignore
|
706 |
]
|
707 |
# if no classes are left then F1 is not defined
|
708 |
# (e.g. only "none" in references)
|
|
|
711 |
|
712 |
for label in labels:
|
713 |
self.add_str_to_id(label)
|
714 |
+
formatted_references = [
|
715 |
+
self.get_one_hot_vector(reference) for reference in references
|
716 |
+
]
|
717 |
+
formatted_predictions = [
|
718 |
+
self.get_one_hot_vector(prediction) for prediction in predictions
|
719 |
+
]
|
720 |
|
721 |
# There is odd behavior in scikit-learn that when passing a one-hot vector with a single
|
722 |
# element, it is treated a class identifier. Therefore, we add labels=[1] to limit to only
|
|
|
767 |
sent_split_newline: bool = True
|
768 |
|
769 |
def prepare(self):
|
|
|
|
|
770 |
super().prepare()
|
771 |
+
|
772 |
+
self.hf_compute_args.update(
|
773 |
+
{"use_aggregator": self.use_aggregator, "rouge_types": self.rouge_types}
|
774 |
+
)
|
775 |
+
|
776 |
import nltk
|
777 |
|
778 |
nltk.download("punkt")
|
779 |
self.sent_tokenize = nltk.sent_tokenize
|
780 |
|
781 |
+
def compute(self, references, predictions, additional_inputs: List[Dict]):
|
782 |
if self.sent_split_newline:
|
783 |
+
predictions = [
|
784 |
+
"\n".join(self.sent_tokenize(prediction.strip()))
|
785 |
+
for prediction in predictions
|
786 |
+
]
|
787 |
+
references = [
|
788 |
+
["\n".join(self.sent_tokenize(r.strip())) for r in reference]
|
789 |
+
for reference in references
|
790 |
+
]
|
791 |
+
return super().compute(references, predictions, additional_inputs)
|
792 |
+
|
793 |
+
|
794 |
+
# Computes char edit distance, ignoring whitespace
|
795 |
+
class CharEditDistanceAccuracy(InstanceMetric):
|
796 |
reduction_map = {"mean": ["char_edit_dist_accuracy"]}
|
797 |
main_score = "char_edit_dist_accuracy"
|
798 |
|
799 |
def prepare(self):
|
800 |
+
super().prepare()
|
801 |
import editdistance
|
802 |
|
803 |
self.eval = editdistance.eval
|
804 |
|
805 |
+
def compute(
|
806 |
+
self, references, prediction: str, additional_inputs: List[Dict]
|
807 |
+
) -> dict:
|
808 |
+
assert (
|
809 |
+
len(references) == 1
|
810 |
+
), f"Expected only one reference , but received: {references}"
|
811 |
+
|
812 |
formatted_prediction = "".join(prediction.split())
|
813 |
+
formatted_reference = "".join(references[0].split())
|
814 |
max_length = max(len(formatted_reference), len(formatted_prediction))
|
815 |
if max_length == 0:
|
816 |
+
return {"char_edit_dist_accuracy": 0.0}
|
817 |
edit_dist = self.eval(formatted_reference, formatted_prediction)
|
818 |
return {"char_edit_dist_accuracy": (1 - edit_dist / max_length)}
|
819 |
|
|
|
822 |
hf_metric_name = "wer"
|
823 |
main_score = "wer"
|
824 |
|
825 |
+
def compute(
|
826 |
+
self,
|
827 |
+
references: List[List[str]],
|
828 |
+
predictions: List[str],
|
829 |
+
additional_inputs: List[Dict],
|
830 |
+
) -> dict:
|
831 |
assert all(
|
832 |
len(reference) == 1 for reference in references
|
833 |
), "Only single reference per prediction is allowed in wer metric"
|
834 |
formatted_references = [reference[0] for reference in references]
|
835 |
+
result = self.metric.compute(
|
836 |
+
predictions=predictions, references=formatted_references
|
837 |
+
)
|
838 |
return {self.main_score: result}
|
839 |
|
840 |
|
|
|
849 |
self.str_to_id[str] = id
|
850 |
return self.str_to_id[str]
|
851 |
|
852 |
+
def compute(
|
853 |
+
self,
|
854 |
+
references: List[List[str]],
|
855 |
+
predictions: List[str],
|
856 |
+
additional_inputs: List[Dict],
|
857 |
+
) -> dict:
|
858 |
+
formatted_references = [
|
859 |
+
self.get_str_id(reference[0]) for reference in references
|
860 |
+
]
|
861 |
+
formatted_predictions = [
|
862 |
+
self.get_str_id(prediction) for prediction in predictions
|
863 |
+
]
|
864 |
+
return self.metric.compute(
|
865 |
+
predictions=formatted_predictions, references=formatted_references
|
866 |
+
)
|
867 |
|
868 |
|
869 |
class CustomF1(GlobalMetric):
|
870 |
main_score = "f1_micro"
|
871 |
classes = None
|
872 |
+
zero_division = 0.0
|
873 |
|
874 |
@abstractmethod
|
875 |
def get_element_group(self, element):
|
|
|
879 |
def get_element_representation(self, element):
|
880 |
pass
|
881 |
|
882 |
+
def group_elements(self, elements_list):
|
883 |
return {
|
884 |
+
k: Counter(
|
885 |
+
[
|
886 |
+
self.get_element_representation(value)
|
887 |
+
for value in elements_list
|
888 |
+
if self.get_element_group(value) == k
|
889 |
+
]
|
890 |
+
)
|
891 |
+
for k in {self.get_element_group(e) for e in elements_list}
|
892 |
}
|
893 |
|
894 |
def calculate_groups_ratio(self, actual_group, total_group):
|
895 |
+
return sum(
|
896 |
+
[min(actual_group[k], total_group[k]) for k in actual_group.keys()]
|
897 |
+
), sum(actual_group.values())
|
898 |
+
|
899 |
+
def precision(self, pn, pd, rn, rd):
|
900 |
+
return self.zero_division if pn == 0 and pd == 0 else pn / pd
|
901 |
+
|
902 |
+
def recall(self, pn, pd, rn, rd):
|
903 |
+
return self.zero_division if rn == 0 and rd == 0 else rn / rd
|
904 |
|
905 |
def f1(self, pn, pd, rn, rd):
|
906 |
+
precision = self.precision(pn, pd, rn, rd)
|
907 |
+
recall = self.recall(pn, pd, rn, rd)
|
908 |
try:
|
909 |
return 2 * precision * recall / (precision + recall)
|
910 |
except ZeroDivisionError:
|
911 |
+
return self.zero_division
|
912 |
+
|
913 |
+
def compute(
|
914 |
+
self,
|
915 |
+
references: List[Any],
|
916 |
+
predictions: List[Any],
|
917 |
+
additional_inputs: List[Dict],
|
918 |
+
) -> dict:
|
919 |
# in case reference are List[List[List[Any]]] and predictions are List[List[Any]]:
|
920 |
if isinstance(references[0], list) and isinstance(references[0][0], list):
|
921 |
references = [element[0] for element in references]
|
922 |
|
923 |
assert len(references) == len(predictions), (
|
924 |
+
f"references size ({len(references)})"
|
925 |
+
f" doesn't mach predictions sise ({len(references)})."
|
926 |
)
|
927 |
if self.classes is None:
|
928 |
+
classes = {
|
929 |
+
self.get_element_group(e) for sublist in references for e in sublist
|
930 |
+
}
|
931 |
else:
|
932 |
classes = self.classes
|
933 |
+
groups_statistics = {}
|
934 |
for references_batch, predictions_batch in zip(references, predictions):
|
935 |
grouped_references = self.group_elements(references_batch)
|
936 |
grouped_predictions = self.group_elements(predictions_batch)
|
937 |
+
all_groups = set(grouped_references.keys()).union(
|
938 |
+
grouped_predictions.keys()
|
939 |
+
)
|
940 |
for group in all_groups:
|
941 |
if group not in groups_statistics:
|
942 |
groups_statistics[group] = {
|
|
|
958 |
groups_statistics[group]["recall_numerator"] += rn
|
959 |
groups_statistics[group]["recall_denominator"] += rd
|
960 |
|
|
|
961 |
num_of_unknown_class_predictions = 0
|
962 |
pn_total = pd_total = rn_total = rd_total = 0
|
963 |
+
f1_result = {}
|
964 |
+
recall_result = {}
|
965 |
+
precision_result = {}
|
966 |
for group in groups_statistics.keys():
|
967 |
pn, pd, rn, rd = (
|
968 |
groups_statistics[group]["precision_numerator"],
|
|
|
970 |
groups_statistics[group]["recall_numerator"],
|
971 |
groups_statistics[group]["recall_denominator"],
|
972 |
)
|
973 |
+
pn_total, pd_total, rn_total, rd_total = (
|
974 |
+
pn_total + pn,
|
975 |
+
pd_total + pd,
|
976 |
+
rn_total + rn,
|
977 |
+
rd_total + rd,
|
978 |
+
)
|
979 |
if group in classes:
|
980 |
+
f1_result[f"f1_{group}"] = self.f1(pn, pd, rn, rd)
|
981 |
+
recall_result[f"recall_{group}"] = self.recall(pn, pd, rn, rd)
|
982 |
+
precision_result[f"precision_{group}"] = self.precision(pn, pd, rn, rd)
|
983 |
else:
|
984 |
num_of_unknown_class_predictions += pd
|
985 |
+
|
986 |
+
result = f1_result
|
987 |
try:
|
988 |
+
result["f1_macro"] = sum(f1_result.values()) / len(result.keys())
|
989 |
+
result["recall_macro"] = sum(recall_result.values()) / len(
|
990 |
+
recall_result.keys()
|
991 |
+
)
|
992 |
+
result["precision_macro"] = sum(precision_result.values()) / len(
|
993 |
+
precision_result.keys()
|
994 |
+
)
|
995 |
except ZeroDivisionError:
|
996 |
+
result["f1_macro"] = self.zero_division
|
997 |
+
result["recall_macro"] = self.zero_division
|
998 |
+
result["micro_macro"] = self.zero_division
|
999 |
|
1000 |
amount_of_predictions = pd_total
|
1001 |
if amount_of_predictions == 0:
|
1002 |
result["in_classes_support"] = 1.0
|
1003 |
else:
|
1004 |
+
result["in_classes_support"] = (
|
1005 |
+
1.0 - num_of_unknown_class_predictions / amount_of_predictions
|
1006 |
+
)
|
1007 |
+
result["f1_micro"] = self.f1(pn_total, pd_total, rn_total, rd_total)
|
1008 |
+
result["recall_micro"] = self.recall(pn_total, pd_total, rn_total, rd_total)
|
1009 |
+
result["precision_micro"] = self.precision(
|
1010 |
+
pn_total, pd_total, rn_total, rd_total
|
1011 |
+
)
|
1012 |
return result
|
1013 |
|
1014 |
|
|
|
1043 |
reduction_map = {"mean": ["f1", "precision", "recall"]}
|
1044 |
main_score = "f1"
|
1045 |
|
1046 |
+
def compute(
|
1047 |
+
self, references: List[Any], prediction: Any, additional_inputs: List[Dict]
|
1048 |
+
) -> dict:
|
1049 |
+
results = [
|
1050 |
+
self._compute_single_ref(reference, prediction) for reference in references
|
1051 |
+
]
|
1052 |
+
return {
|
1053 |
+
measure: max(r[i] for r in results)
|
1054 |
+
for i, measure in enumerate(["precision", "recall", "f1"])
|
1055 |
+
}
|
1056 |
|
1057 |
+
def _compute_single_ref(
|
1058 |
+
self, reference: Any, prediction: Any
|
1059 |
+
) -> Tuple[float, float, float]:
|
1060 |
prediction_tokens = normalize_answer(prediction).split()
|
1061 |
reference_tokens = normalize_answer(reference).split()
|
1062 |
common = Counter(prediction_tokens) & Counter(reference_tokens)
|
|
|
1097 |
self.model = SentenceTransformer(self.model_name)
|
1098 |
self.util = sbert_util
|
1099 |
|
1100 |
+
def compute(
|
1101 |
+
self,
|
1102 |
+
references: List[List[Any]],
|
1103 |
+
predictions: List[Any],
|
1104 |
+
additional_inputs: List[Dict],
|
1105 |
+
) -> List[Any]:
|
1106 |
scores = []
|
1107 |
|
1108 |
# we are in a multi-reference case (each prediction may have multiple
|
|
|
1117 |
|
1118 |
# compute s-bert embeddings
|
1119 |
preds_emb = self.model.encode(predictions)
|
1120 |
+
refs_emb = self.model.encode(
|
1121 |
+
[ref for ref_group in references for ref in ref_group]
|
1122 |
+
)
|
1123 |
|
1124 |
# for each candidate, pick the reference with the highest score
|
1125 |
for pred_emb, ref_group_bounds in zip(preds_emb, ref_group_boundaries):
|
|
|
1137 |
model_name: str
|
1138 |
|
1139 |
def prepare(self):
|
1140 |
+
super().prepare()
|
1141 |
from transformers import pipeline
|
1142 |
|
1143 |
self.pipe = pipeline("text-classification", model=self.model_name)
|
1144 |
|
1145 |
+
def compute(
|
1146 |
+
self,
|
1147 |
+
references: List[List[Any]],
|
1148 |
+
predictions: List[Any],
|
1149 |
+
additional_inputs: List[Dict],
|
1150 |
+
) -> List[Any]:
|
1151 |
# treat the references as the questions and the predictions as answers
|
1152 |
# assume a single reference
|
1153 |
questions = [refs[0] for refs in references]
|
|
|
1159 |
# compute the metric
|
1160 |
# add function_to_apply="none" to disable sigmoid
|
1161 |
return self.pipe(inputs, batch_size=self.batch_size)
|
1162 |
+
|
1163 |
+
|
1164 |
+
class NDCG(GlobalMetric):
|
1165 |
+
"""Normalized Discounted Cumulative Gain: measures the quality of ranking with respect to ground truth ranking scores.
|
1166 |
+
|
1167 |
+
As this measures ranking, it is a global metric that can only be calculated over groups of instances. In the
|
1168 |
+
common use case where the instances are grouped by different queries, i.e., where the task is to provide a
|
1169 |
+
relevance score for a search result w.r.t. a query, an nDCG score is calculated per each query (specified in the
|
1170 |
+
"query" input field of an instance) and the final score is the average across all queries.
|
1171 |
+
Note that the expected scores are relevance scores (i.e., higher is better) and not rank indices. The absolute
|
1172 |
+
value of the scores is only meaningful for the reference scores; for the predictions, only the ordering of the
|
1173 |
+
scores affects the outcome - for example, predicted scores of [80, 1, 2] and [0.8, 0.5, 0.6] will receive
|
1174 |
+
the same nDCG score w.r.t. a given set of reference scores.
|
1175 |
+
|
1176 |
+
See also https://en.wikipedia.org/wiki/Discounted_cumulative_gain
|
1177 |
+
"""
|
1178 |
+
|
1179 |
+
main_score = "nDCG"
|
1180 |
+
|
1181 |
+
def prepare(self):
|
1182 |
+
from sklearn.metrics import ndcg_score
|
1183 |
+
|
1184 |
+
super().prepare()
|
1185 |
+
self.eval = ndcg_score
|
1186 |
+
|
1187 |
+
def compute(
|
1188 |
+
self,
|
1189 |
+
references: List[List[Any]],
|
1190 |
+
predictions: List[Any],
|
1191 |
+
additional_inputs: List[Any],
|
1192 |
+
) -> dict:
|
1193 |
+
from collections import defaultdict
|
1194 |
+
from statistics import mean
|
1195 |
+
|
1196 |
+
query_to_predictions_and_references = defaultdict(lambda: [[], []])
|
1197 |
+
for reference, pred, inputs_dict in zip(
|
1198 |
+
references, predictions, additional_inputs
|
1199 |
+
):
|
1200 |
+
query = inputs_dict.get("query")
|
1201 |
+
query_to_predictions_and_references[query][0].append(pred)
|
1202 |
+
query_to_predictions_and_references[query][1].append(reference)
|
1203 |
+
|
1204 |
+
scores = []
|
1205 |
+
for q_predictions, q_references in query_to_predictions_and_references.values():
|
1206 |
+
if len(q_references) == 1:
|
1207 |
+
continue
|
1208 |
+
|
1209 |
+
if (
|
1210 |
+
None in q_predictions
|
1211 |
+
): # model failed to predict numeric scores for some instances
|
1212 |
+
numeric_predictions = [
|
1213 |
+
pred for pred in q_predictions if pred is not None
|
1214 |
+
]
|
1215 |
+
if len(numeric_predictions) <= 1: # no meaningful ranking
|
1216 |
+
scores.append(0)
|
1217 |
+
continue
|
1218 |
+
# consider non-numeric model predictions as ranked last
|
1219 |
+
min_value = min(numeric_predictions)
|
1220 |
+
q_predictions = [
|
1221 |
+
1 + (pred - min_value) if pred is not None else 0
|
1222 |
+
for pred in q_predictions
|
1223 |
+
]
|
1224 |
+
scores.append(self.eval([q_references], [q_predictions]))
|
1225 |
+
return {self.main_score: mean(scores) if len(scores) > 0 else np.nan}
|