Upload metrics.py with huggingface_hub
Browse files- metrics.py +356 -49
metrics.py
CHANGED
@@ -1,4 +1,3 @@
|
|
1 |
-
import logging
|
2 |
import re
|
3 |
import string
|
4 |
import uuid
|
@@ -14,6 +13,7 @@ from scipy.stats import bootstrap
|
|
14 |
|
15 |
from .artifact import Artifact
|
16 |
from .dataclass import InternalField, OptionalField
|
|
|
17 |
from .operator import (
|
18 |
MultiStreamOperator,
|
19 |
SingleStreamOperator,
|
@@ -23,7 +23,9 @@ from .operator import (
|
|
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
|
@@ -61,6 +63,7 @@ class MetricWithConfidenceInterval(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():
|
@@ -79,7 +82,7 @@ class MetricWithConfidenceInterval(Metric):
|
|
79 |
and num_predictions > 1
|
80 |
)
|
81 |
|
82 |
-
def score_based_confidence_interval(self,
|
83 |
"""Compute confidence intervals based on existing scores, already computed on the input instances.
|
84 |
|
85 |
score_names: List[str]
|
@@ -94,6 +97,10 @@ class MetricWithConfidenceInterval(Metric):
|
|
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
|
@@ -131,7 +138,7 @@ class MetricWithConfidenceInterval(Metric):
|
|
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 |
-
|
135 |
return np.nan
|
136 |
|
137 |
scores = numpy.apply_along_axis(
|
@@ -341,7 +348,7 @@ class BulkInstanceMetric(SingleStreamOperator, MetricWithConfidenceInterval):
|
|
341 |
global_score["score_name"] = self.main_score
|
342 |
|
343 |
confidence_interval = self.score_based_confidence_interval(
|
344 |
-
|
345 |
)
|
346 |
global_score.update(confidence_interval)
|
347 |
|
@@ -411,7 +418,7 @@ class InstanceMetric(SingleStreamOperator, MetricWithConfidenceInterval):
|
|
411 |
global_score["score_name"] = self.main_score
|
412 |
|
413 |
confidence_interval = self.score_based_confidence_interval(
|
414 |
-
|
415 |
)
|
416 |
global_score.update(confidence_interval)
|
417 |
|
@@ -473,6 +480,23 @@ class Accuracy(InstanceMetric):
|
|
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)
|
@@ -512,9 +536,29 @@ class HuggingfaceMetric(GlobalMetric):
|
|
512 |
|
513 |
scale: float = 1.0 # optional scaling of main results
|
514 |
scaled_fields: list = None
|
|
|
515 |
hf_compute_args: Dict[str, Any] = OptionalField(default_factory=dict)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
516 |
experiment_id: str = OptionalField(default_factory=lambda: str(uuid.uuid4()))
|
517 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
518 |
def prepare(self):
|
519 |
super().prepare()
|
520 |
self.metric = evaluate.load(
|
@@ -527,8 +571,36 @@ class HuggingfaceMetric(GlobalMetric):
|
|
527 |
predictions: List[Any],
|
528 |
additional_inputs: List[Dict],
|
529 |
) -> dict:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
530 |
result = self.metric.compute(
|
531 |
-
predictions=predictions,
|
|
|
|
|
|
|
532 |
)
|
533 |
if self.hf_main_score:
|
534 |
result[self.main_score] = result[self.hf_main_score]
|
@@ -559,6 +631,7 @@ class HuggingfaceBulkMetric(BulkInstanceMetric):
|
|
559 |
|
560 |
hf_metric_fields: List[str]
|
561 |
hf_compute_args: dict = {}
|
|
|
562 |
|
563 |
def prepare(self):
|
564 |
super().prepare()
|
@@ -570,8 +643,23 @@ class HuggingfaceBulkMetric(BulkInstanceMetric):
|
|
570 |
predictions: List[str],
|
571 |
additional_inputs: List[Any],
|
572 |
) -> List[Dict[str, Any]]:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
573 |
scores = self.metric.compute(
|
574 |
-
predictions=predictions,
|
|
|
|
|
|
|
575 |
)
|
576 |
|
577 |
# convert dict of lists to a list of dicts
|
@@ -656,10 +744,11 @@ class F1MultiLabel(GlobalMetric):
|
|
656 |
main_score = "f1_macro"
|
657 |
average = None # Report per class then aggregate by mean
|
658 |
classes_to_ignore = ["none"]
|
|
|
659 |
|
660 |
def prepare(self):
|
661 |
super().prepare()
|
662 |
-
self._metric = evaluate.load(
|
663 |
|
664 |
def add_str_to_id(self, str):
|
665 |
if str not in self.str_to_id:
|
@@ -683,22 +772,10 @@ class F1MultiLabel(GlobalMetric):
|
|
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}
|
@@ -732,19 +809,60 @@ class F1MultiLabel(GlobalMetric):
|
|
732 |
average=self.average,
|
733 |
labels=labels_param,
|
734 |
)
|
735 |
-
if isinstance(result[
|
736 |
from statistics import mean
|
737 |
|
738 |
-
assert
|
739 |
-
labels
|
740 |
-
), f
|
741 |
-
final_result = {self.main_score: mean(result[
|
742 |
for i, label in enumerate(labels):
|
743 |
-
final_result["
|
744 |
else:
|
745 |
-
final_result = {self.main_score: result[
|
746 |
return final_result
|
747 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
748 |
|
749 |
class F1MicroMultiLabel(F1MultiLabel):
|
750 |
main_score = "f1_micro"
|
@@ -868,27 +986,36 @@ class MatthewsCorrelation(HuggingfaceMetric):
|
|
868 |
|
869 |
class CustomF1(GlobalMetric):
|
870 |
main_score = "f1_micro"
|
871 |
-
|
872 |
zero_division = 0.0
|
873 |
|
874 |
@abstractmethod
|
875 |
-
def get_element_group(self, element):
|
876 |
pass
|
877 |
|
878 |
@abstractmethod
|
879 |
-
def get_element_representation(self, element):
|
880 |
pass
|
881 |
|
882 |
-
def
|
|
|
|
|
|
|
|
|
|
|
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 {
|
|
|
|
|
|
|
|
|
892 |
}
|
893 |
|
894 |
def calculate_groups_ratio(self, actual_group, total_group):
|
@@ -910,30 +1037,46 @@ class CustomF1(GlobalMetric):
|
|
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
|
|
|
|
|
|
|
|
|
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 |
-
|
928 |
-
|
929 |
-
|
930 |
-
}
|
931 |
else:
|
932 |
-
|
933 |
groups_statistics = {}
|
934 |
-
for references_batch, predictions_batch in zip(
|
935 |
-
|
936 |
-
|
|
|
|
|
|
|
|
|
937 |
all_groups = set(grouped_references.keys()).union(
|
938 |
grouped_predictions.keys()
|
939 |
)
|
@@ -976,7 +1119,7 @@ class CustomF1(GlobalMetric):
|
|
976 |
rn_total + rn,
|
977 |
rd_total + rd,
|
978 |
)
|
979 |
-
if group in
|
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)
|
@@ -995,7 +1138,7 @@ class CustomF1(GlobalMetric):
|
|
995 |
except ZeroDivisionError:
|
996 |
result["f1_macro"] = self.zero_division
|
997 |
result["recall_macro"] = self.zero_division
|
998 |
-
result["
|
999 |
|
1000 |
amount_of_predictions = pd_total
|
1001 |
if amount_of_predictions == 0:
|
@@ -1013,10 +1156,10 @@ class CustomF1(GlobalMetric):
|
|
1013 |
|
1014 |
|
1015 |
class NER(CustomF1):
|
1016 |
-
def get_element_group(self, element):
|
1017 |
return element[1]
|
1018 |
|
1019 |
-
def get_element_representation(self, element):
|
1020 |
return str(element)
|
1021 |
|
1022 |
|
@@ -1042,6 +1185,7 @@ def normalize_answer(s):
|
|
1042 |
class TokenOverlap(InstanceMetric):
|
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]
|
@@ -1075,6 +1219,7 @@ class BertScore(HuggingfaceBulkMetric):
|
|
1075 |
main_score = "f1"
|
1076 |
reduction_map = {"mean": ["f1", "precision", "recall"]}
|
1077 |
hf_metric_fields = ["f1", "precision", "recall"]
|
|
|
1078 |
model_name: str
|
1079 |
|
1080 |
def prepare(self):
|
@@ -1223,3 +1368,165 @@ class NDCG(GlobalMetric):
|
|
1223 |
]
|
1224 |
scores.append(self.eval([q_references], [q_predictions]))
|
1225 |
return {self.main_score: mean(scores) if len(scores) > 0 else np.nan}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import re
|
2 |
import string
|
3 |
import uuid
|
|
|
13 |
|
14 |
from .artifact import Artifact
|
15 |
from .dataclass import InternalField, OptionalField
|
16 |
+
from .logging_utils import get_logger
|
17 |
from .operator import (
|
18 |
MultiStreamOperator,
|
19 |
SingleStreamOperator,
|
|
|
23 |
from .operators import CopyFields
|
24 |
from .random_utils import get_seed
|
25 |
from .stream import MultiStream, Stream
|
26 |
+
from .type_utils import isoftype
|
27 |
|
28 |
+
logger = get_logger()
|
29 |
# The default number of resamples used to estimate the confidence intervals
|
30 |
# global and instances metrics. Use None to disable confidence interval computation by default.
|
31 |
_N_RESAMPLES_DEFAULT_FOR_INSTANCE_METRICS = 1000
|
|
|
63 |
# Use None to disable confidence interval computation.
|
64 |
n_resamples: int = None
|
65 |
confidence_level: float = 0.95
|
66 |
+
ci_scores: List[str] = None
|
67 |
|
68 |
@staticmethod
|
69 |
def new_random_generator():
|
|
|
82 |
and num_predictions > 1
|
83 |
)
|
84 |
|
85 |
+
def score_based_confidence_interval(self, instances):
|
86 |
"""Compute confidence intervals based on existing scores, already computed on the input instances.
|
87 |
|
88 |
score_names: List[str]
|
|
|
97 |
if not self._can_compute_confidence_intervals(num_predictions=len(instances)):
|
98 |
return result
|
99 |
|
100 |
+
score_names = (
|
101 |
+
self.ci_scores if self.ci_scores is not None else [self.main_score]
|
102 |
+
)
|
103 |
+
|
104 |
for score_name in score_names:
|
105 |
scores = [
|
106 |
instance["score"]["instance"][score_name] for instance in instances
|
|
|
138 |
except Exception as e:
|
139 |
# this happens in edge cases, for example, when the sampling creates a
|
140 |
# sample where all strings are empty and this fails bleu.
|
141 |
+
logger.info(f"Warning in {self.__class__.__name__}", e)
|
142 |
return np.nan
|
143 |
|
144 |
scores = numpy.apply_along_axis(
|
|
|
348 |
global_score["score_name"] = self.main_score
|
349 |
|
350 |
confidence_interval = self.score_based_confidence_interval(
|
351 |
+
instances=instances
|
352 |
)
|
353 |
global_score.update(confidence_interval)
|
354 |
|
|
|
418 |
global_score["score_name"] = self.main_score
|
419 |
|
420 |
confidence_interval = self.score_based_confidence_interval(
|
421 |
+
instances=instances
|
422 |
)
|
423 |
global_score.update(confidence_interval)
|
424 |
|
|
|
480 |
return result
|
481 |
|
482 |
|
483 |
+
class StringContainment(InstanceMetric):
|
484 |
+
reduction_map = {"mean": ["string_containment"]}
|
485 |
+
main_score = "string_containment"
|
486 |
+
|
487 |
+
def compute(
|
488 |
+
self, references: List[Any], prediction: Any, additional_inputs: List[Dict]
|
489 |
+
) -> dict:
|
490 |
+
result = {
|
491 |
+
self.main_score: float(
|
492 |
+
any(str(reference) in prediction for reference in references)
|
493 |
+
)
|
494 |
+
}
|
495 |
+
result["score"] = result[self.main_score]
|
496 |
+
result["score_name"] = self.main_score
|
497 |
+
return result
|
498 |
+
|
499 |
+
|
500 |
class MetricPipeline(MultiStreamOperator, Metric):
|
501 |
main_score: str = None
|
502 |
preprocess_steps: Optional[List[StreamingOperator]] = field(default_factory=list)
|
|
|
536 |
|
537 |
scale: float = 1.0 # optional scaling of main results
|
538 |
scaled_fields: list = None
|
539 |
+
# This are fixed arguments passed to compute method
|
540 |
hf_compute_args: Dict[str, Any] = OptionalField(default_factory=dict)
|
541 |
+
# These are additional input fields passed to HF compute method (a list with one value per instance)
|
542 |
+
hf_additional_input_fields: List = OptionalField(default_factory=list)
|
543 |
+
# These are additional input fields that are passed as one value
|
544 |
+
hf_additional_input_fields_pass_one_value: List = OptionalField(
|
545 |
+
default_factory=list
|
546 |
+
)
|
547 |
+
|
548 |
experiment_id: str = OptionalField(default_factory=lambda: str(uuid.uuid4()))
|
549 |
|
550 |
+
def verify(self):
|
551 |
+
assert (
|
552 |
+
self.hf_additional_input_fields is None
|
553 |
+
or isoftype(self.hf_additional_input_fields, List[str])
|
554 |
+
), f"Argument hf_additional_input_fields should be either None or List[str]. It is now: {self.hf_additional_input_fields}."
|
555 |
+
assert (
|
556 |
+
self.hf_additional_input_fields_pass_one_value is None
|
557 |
+
or isoftype(self.hf_additional_input_fields_pass_one_value, List[str])
|
558 |
+
), f"Argument hf_additional_input_fields_pass_one_value should be either None or List[str]. It is now: {self.hf_additional_input_fields_pass_one_value}."
|
559 |
+
|
560 |
+
return super().verify()
|
561 |
+
|
562 |
def prepare(self):
|
563 |
super().prepare()
|
564 |
self.metric = evaluate.load(
|
|
|
571 |
predictions: List[Any],
|
572 |
additional_inputs: List[Dict],
|
573 |
) -> dict:
|
574 |
+
passed_additional_inputs = {}
|
575 |
+
for additional_input_field in self.hf_additional_input_fields:
|
576 |
+
assert (
|
577 |
+
additional_input_field in additional_inputs[0]
|
578 |
+
), f"'{additional_input_field}' field required by {__class__.__name__} is not in passed in additional inputs: {additional_inputs[0]}"
|
579 |
+
passed_additional_inputs[additional_input_field] = [
|
580 |
+
additional_input[additional_input_field]
|
581 |
+
for additional_input in additional_inputs
|
582 |
+
]
|
583 |
+
for additional_input_field in self.hf_additional_input_fields_pass_one_value:
|
584 |
+
assert (
|
585 |
+
additional_input_field in additional_inputs[0]
|
586 |
+
), f"'{additional_input_field}' field required by {__class__.__name__} is not in passed in additional inputs: {additional_inputs[0]}"
|
587 |
+
|
588 |
+
values = {
|
589 |
+
additional_input[additional_input_field]
|
590 |
+
for additional_input in additional_inputs
|
591 |
+
}
|
592 |
+
assert (
|
593 |
+
len(values) == 1
|
594 |
+
), f"Values of '{additional_input_field}' field required by {__class__.__name__} should all be the same, but have multiple values {values}"
|
595 |
+
|
596 |
+
passed_additional_inputs[additional_input_field] = next(iter(values))
|
597 |
+
|
598 |
+
# add check that all required fields in self.metrics are in passed_additional_inputs print(passed_additional_inputs)
|
599 |
result = self.metric.compute(
|
600 |
+
predictions=predictions,
|
601 |
+
references=references,
|
602 |
+
**passed_additional_inputs,
|
603 |
+
**self.hf_compute_args,
|
604 |
)
|
605 |
if self.hf_main_score:
|
606 |
result[self.main_score] = result[self.hf_main_score]
|
|
|
631 |
|
632 |
hf_metric_fields: List[str]
|
633 |
hf_compute_args: dict = {}
|
634 |
+
hf_additional_input_fields: List = OptionalField(default_factory=list)
|
635 |
|
636 |
def prepare(self):
|
637 |
super().prepare()
|
|
|
643 |
predictions: List[str],
|
644 |
additional_inputs: List[Any],
|
645 |
) -> List[Dict[str, Any]]:
|
646 |
+
passed_additional_inputs = {}
|
647 |
+
passed_additional_inputs = {}
|
648 |
+
for additional_input_field in self.hf_additional_input_fields:
|
649 |
+
assert (
|
650 |
+
additional_input_field in additional_inputs[0]
|
651 |
+
), f"'{additional_input_field}' field required by {__class__.__name__} is not in passed in additional inputs: {additional_inputs[0]}"
|
652 |
+
passed_additional_inputs[additional_input_field] = [
|
653 |
+
additional_input[additional_input_field]
|
654 |
+
for additional_input in additional_inputs
|
655 |
+
]
|
656 |
+
# add check that all required fields in self.metrics are in passed_additional_inputs
|
657 |
+
|
658 |
scores = self.metric.compute(
|
659 |
+
predictions=predictions,
|
660 |
+
references=references,
|
661 |
+
**passed_additional_inputs,
|
662 |
+
**self.hf_compute_args,
|
663 |
)
|
664 |
|
665 |
# convert dict of lists to a list of dicts
|
|
|
744 |
main_score = "f1_macro"
|
745 |
average = None # Report per class then aggregate by mean
|
746 |
classes_to_ignore = ["none"]
|
747 |
+
metric = "f1"
|
748 |
|
749 |
def prepare(self):
|
750 |
super().prepare()
|
751 |
+
self._metric = evaluate.load(self.metric, "multilabel")
|
752 |
|
753 |
def add_str_to_id(self, str):
|
754 |
if str not in self.str_to_id:
|
|
|
772 |
) -> dict:
|
773 |
self.str_to_id = {}
|
774 |
self.id_to_str = {}
|
|
|
|
|
|
|
775 |
|
776 |
+
self._validate_references_and_prediction(references, predictions)
|
777 |
references = [reference[0] for reference in references]
|
778 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
779 |
labels = [
|
780 |
lbl
|
781 |
for lbl in {label for reference in references for label in reference}
|
|
|
809 |
average=self.average,
|
810 |
labels=labels_param,
|
811 |
)
|
812 |
+
if isinstance(result[self.metric], numpy.ndarray):
|
813 |
from statistics import mean
|
814 |
|
815 |
+
assert (
|
816 |
+
len(result[self.metric]) == len(labels)
|
817 |
+
), f"F1 result ({result[self.metric]}) has more entries than labels ({labels})"
|
818 |
+
final_result = {self.main_score: mean(result[self.metric])}
|
819 |
for i, label in enumerate(labels):
|
820 |
+
final_result[self.metric + "_" + label] = result[self.metric][i]
|
821 |
else:
|
822 |
+
final_result = {self.main_score: result[self.metric]}
|
823 |
return final_result
|
824 |
|
825 |
+
def _validate_references_and_prediction(self, references, predictions):
|
826 |
+
for reference in references:
|
827 |
+
if not len(reference) == 1:
|
828 |
+
raise ValueError(
|
829 |
+
f"Only a single reference per prediction is allowed in F1 multi label metric. Received reference: {reference}"
|
830 |
+
)
|
831 |
+
if not isoftype(reference[0], List[str]):
|
832 |
+
raise ValueError(
|
833 |
+
f"Each reference is expected to be a list of strings in F1 multi label metric. Received reference: '{reference[0]}'"
|
834 |
+
)
|
835 |
+
|
836 |
+
for prediction in predictions:
|
837 |
+
if not isoftype(prediction, List[str]):
|
838 |
+
raise ValueError(
|
839 |
+
f"Each prediction is expected to be a list of strings in F1 multi label metric. Received prediction: '{prediction}'"
|
840 |
+
)
|
841 |
+
|
842 |
+
|
843 |
+
class PrecisionMacroMultiLabel(F1MultiLabel):
|
844 |
+
main_score = "precision_macro"
|
845 |
+
metric = "precision"
|
846 |
+
average = "macro"
|
847 |
+
|
848 |
+
|
849 |
+
class PrecisionMicroMultiLabel(F1MultiLabel):
|
850 |
+
main_score = "precision_micro"
|
851 |
+
metric = "precision"
|
852 |
+
average = "micro"
|
853 |
+
|
854 |
+
|
855 |
+
class RecallMacroMultiLabel(F1MultiLabel):
|
856 |
+
main_score = "recall_macro"
|
857 |
+
metric = "recall"
|
858 |
+
average = "macro"
|
859 |
+
|
860 |
+
|
861 |
+
class RecallMicroMultiLabel(F1MultiLabel):
|
862 |
+
main_score = "recall_micro"
|
863 |
+
metric = "recall"
|
864 |
+
average = "micro"
|
865 |
+
|
866 |
|
867 |
class F1MicroMultiLabel(F1MultiLabel):
|
868 |
main_score = "f1_micro"
|
|
|
986 |
|
987 |
class CustomF1(GlobalMetric):
|
988 |
main_score = "f1_micro"
|
989 |
+
groups = None
|
990 |
zero_division = 0.0
|
991 |
|
992 |
@abstractmethod
|
993 |
+
def get_element_group(self, element, additional_input):
|
994 |
pass
|
995 |
|
996 |
@abstractmethod
|
997 |
+
def get_element_representation(self, element, additional_input):
|
998 |
pass
|
999 |
|
1000 |
+
def should_ignore_element(self, element, additional_input):
|
1001 |
+
return False
|
1002 |
+
|
1003 |
+
def group_elements(self, elements_list, additional_input):
|
1004 |
+
if not isinstance(elements_list, list):
|
1005 |
+
elements_list = [elements_list]
|
1006 |
return {
|
1007 |
k: Counter(
|
1008 |
[
|
1009 |
+
self.get_element_representation(value, additional_input)
|
1010 |
for value in elements_list
|
1011 |
+
if self.get_element_group(value, additional_input) == k
|
1012 |
]
|
1013 |
)
|
1014 |
+
for k in {
|
1015 |
+
self.get_element_group(e, additional_input)
|
1016 |
+
for e in elements_list
|
1017 |
+
if not self.should_ignore_element(e, additional_input)
|
1018 |
+
}
|
1019 |
}
|
1020 |
|
1021 |
def calculate_groups_ratio(self, actual_group, total_group):
|
|
|
1037 |
except ZeroDivisionError:
|
1038 |
return self.zero_division
|
1039 |
|
1040 |
+
def get_groups(self, elements, additional_inputs):
|
1041 |
+
groups = set()
|
1042 |
+
for sublist, additional_input in zip(elements, additional_inputs):
|
1043 |
+
for e in sublist:
|
1044 |
+
if self.should_ignore_element(e, additional_input):
|
1045 |
+
continue
|
1046 |
+
groups.add(self.get_element_group(e, additional_input))
|
1047 |
+
return groups
|
1048 |
+
|
1049 |
def compute(
|
1050 |
self,
|
1051 |
+
references: List[List[Any]],
|
1052 |
predictions: List[Any],
|
1053 |
additional_inputs: List[Dict],
|
1054 |
) -> dict:
|
1055 |
# in case reference are List[List[List[Any]]] and predictions are List[List[Any]]:
|
1056 |
+
if (
|
1057 |
+
isinstance(references[0], list)
|
1058 |
+
and len(references[0]) > 0
|
1059 |
+
and isinstance(references[0][0], list)
|
1060 |
+
):
|
1061 |
references = [element[0] for element in references]
|
1062 |
|
1063 |
assert len(references) == len(predictions), (
|
1064 |
f"references size ({len(references)})"
|
1065 |
f" doesn't mach predictions sise ({len(references)})."
|
1066 |
)
|
1067 |
+
|
1068 |
+
if self.groups is None:
|
1069 |
+
groups = self.get_groups(references, additional_inputs)
|
|
|
1070 |
else:
|
1071 |
+
groups = self.groups
|
1072 |
groups_statistics = {}
|
1073 |
+
for references_batch, predictions_batch, additional_input in zip(
|
1074 |
+
references, predictions, additional_inputs
|
1075 |
+
):
|
1076 |
+
grouped_references = self.group_elements(references_batch, additional_input)
|
1077 |
+
grouped_predictions = self.group_elements(
|
1078 |
+
predictions_batch, additional_input
|
1079 |
+
)
|
1080 |
all_groups = set(grouped_references.keys()).union(
|
1081 |
grouped_predictions.keys()
|
1082 |
)
|
|
|
1119 |
rn_total + rn,
|
1120 |
rd_total + rd,
|
1121 |
)
|
1122 |
+
if group in groups:
|
1123 |
f1_result[f"f1_{group}"] = self.f1(pn, pd, rn, rd)
|
1124 |
recall_result[f"recall_{group}"] = self.recall(pn, pd, rn, rd)
|
1125 |
precision_result[f"precision_{group}"] = self.precision(pn, pd, rn, rd)
|
|
|
1138 |
except ZeroDivisionError:
|
1139 |
result["f1_macro"] = self.zero_division
|
1140 |
result["recall_macro"] = self.zero_division
|
1141 |
+
result["precision_macro"] = self.zero_division
|
1142 |
|
1143 |
amount_of_predictions = pd_total
|
1144 |
if amount_of_predictions == 0:
|
|
|
1156 |
|
1157 |
|
1158 |
class NER(CustomF1):
|
1159 |
+
def get_element_group(self, element, additional_input):
|
1160 |
return element[1]
|
1161 |
|
1162 |
+
def get_element_representation(self, element, additional_input):
|
1163 |
return str(element)
|
1164 |
|
1165 |
|
|
|
1185 |
class TokenOverlap(InstanceMetric):
|
1186 |
reduction_map = {"mean": ["f1", "precision", "recall"]}
|
1187 |
main_score = "f1"
|
1188 |
+
ci_scores = ["f1", "precision", "recall"]
|
1189 |
|
1190 |
def compute(
|
1191 |
self, references: List[Any], prediction: Any, additional_inputs: List[Dict]
|
|
|
1219 |
main_score = "f1"
|
1220 |
reduction_map = {"mean": ["f1", "precision", "recall"]}
|
1221 |
hf_metric_fields = ["f1", "precision", "recall"]
|
1222 |
+
ci_scores = ["f1", "precision", "recall"]
|
1223 |
model_name: str
|
1224 |
|
1225 |
def prepare(self):
|
|
|
1368 |
]
|
1369 |
scores.append(self.eval([q_references], [q_predictions]))
|
1370 |
return {self.main_score: mean(scores) if len(scores) > 0 else np.nan}
|
1371 |
+
|
1372 |
+
|
1373 |
+
class RetrievalMetric(InstanceMetric):
|
1374 |
+
def compute(
|
1375 |
+
self, references: List[Any], prediction: Any, additional_inputs: Dict
|
1376 |
+
) -> dict:
|
1377 |
+
# digest input
|
1378 |
+
pred_ids: List[Any] = prediction
|
1379 |
+
ref_ids: List[Any] = list(dict.fromkeys(references))
|
1380 |
+
|
1381 |
+
# relevance_at_k: 1-based dictionary of indicators (0/1), telling whether
|
1382 |
+
# the doc id retrieved at position k (assuming it is 1-based, so k starts
|
1383 |
+
# from 1) is in the gold doc ids or not.
|
1384 |
+
# For example, assuming that in the retrieved docs we have correct predictions
|
1385 |
+
# at positions 2, 4 and 5 (1-based), the dict will look like:
|
1386 |
+
# {1: 0, 2: 1, 3: 0, 4: 1, 5: 1, ...}
|
1387 |
+
relevance_at_k = {
|
1388 |
+
k + 1: 1 if doc_id in ref_ids else 0 for k, doc_id in enumerate(pred_ids)
|
1389 |
+
}
|
1390 |
+
|
1391 |
+
# relevance_sum_at_k: 1-based dictionary of counts, where the value at k determines
|
1392 |
+
# how many gold doc ids have been observed up to index k.
|
1393 |
+
relevance_sum_at_k = {}
|
1394 |
+
for k, value in relevance_at_k.items():
|
1395 |
+
relevance_sum_at_k[k] = relevance_sum_at_k.get(k - 1, 0) + value
|
1396 |
+
|
1397 |
+
# precision_at_k: the precision of the top k retrieved documents. For example,
|
1398 |
+
# assuming that only 1 out of the first 4 retrieved documents is correct, the
|
1399 |
+
# value at 4 will be 1/4.
|
1400 |
+
precision_at_k = {k: value / k for k, value in relevance_sum_at_k.items()}
|
1401 |
+
|
1402 |
+
# recall_at_k: the recall of the top k retrieved documents. For example,
|
1403 |
+
# assuming that only 2 out of the 3 gold documents are in the top 5 results,
|
1404 |
+
# the value at 5 will be 2/3.
|
1405 |
+
n_refs = len(ref_ids)
|
1406 |
+
recall_at_k = {
|
1407 |
+
k: value / n_refs if n_refs > 0 else 0
|
1408 |
+
for k, value in relevance_sum_at_k.items()
|
1409 |
+
}
|
1410 |
+
|
1411 |
+
# rank - the 1-based index of the first hit of a gold doc id. So 1
|
1412 |
+
# means first position.
|
1413 |
+
rank = 0
|
1414 |
+
for k, relevance in relevance_at_k.items():
|
1415 |
+
if relevance == 1:
|
1416 |
+
rank = k
|
1417 |
+
break
|
1418 |
+
|
1419 |
+
# match_at_k: whether we have a match at the top k retrieved documents
|
1420 |
+
match_at_k = {
|
1421 |
+
k: 1.0 if value > 0 else 0.0 for k, value in relevance_sum_at_k.items()
|
1422 |
+
}
|
1423 |
+
|
1424 |
+
return self._compute(
|
1425 |
+
relevance_at_k,
|
1426 |
+
relevance_sum_at_k,
|
1427 |
+
precision_at_k,
|
1428 |
+
recall_at_k,
|
1429 |
+
match_at_k,
|
1430 |
+
rank,
|
1431 |
+
)
|
1432 |
+
|
1433 |
+
@abstractmethod
|
1434 |
+
def _compute(
|
1435 |
+
self,
|
1436 |
+
relevance_at_k,
|
1437 |
+
relevance_sum_at_k,
|
1438 |
+
precision_at_k,
|
1439 |
+
recall_at_k,
|
1440 |
+
match_at_k,
|
1441 |
+
rank,
|
1442 |
+
) -> dict:
|
1443 |
+
pass
|
1444 |
+
|
1445 |
+
|
1446 |
+
class MRR(RetrievalMetric):
|
1447 |
+
reduction_map = {"mean": ["mrr"]}
|
1448 |
+
main_score = "mrr"
|
1449 |
+
|
1450 |
+
def _compute(
|
1451 |
+
self,
|
1452 |
+
relevance_at_k,
|
1453 |
+
relevance_sum_at_k,
|
1454 |
+
precision_at_k,
|
1455 |
+
recall_at_k,
|
1456 |
+
match_at_k,
|
1457 |
+
rank,
|
1458 |
+
) -> dict:
|
1459 |
+
return {self.main_score: 1 / rank if rank > 0 else 0}
|
1460 |
+
|
1461 |
+
|
1462 |
+
class MAP(RetrievalMetric):
|
1463 |
+
reduction_map = {"mean": ["map"]}
|
1464 |
+
main_score = "map"
|
1465 |
+
|
1466 |
+
def _compute(
|
1467 |
+
self,
|
1468 |
+
relevance_at_k,
|
1469 |
+
relevance_sum_at_k,
|
1470 |
+
precision_at_k,
|
1471 |
+
recall_at_k,
|
1472 |
+
match_at_k,
|
1473 |
+
rank,
|
1474 |
+
) -> dict:
|
1475 |
+
result = 0
|
1476 |
+
if len(relevance_at_k) > 0:
|
1477 |
+
total = sum(relevance_at_k.values())
|
1478 |
+
if total > 0:
|
1479 |
+
dot = sum(relevance_at_k[k] * precision_at_k[k] for k in relevance_at_k)
|
1480 |
+
result = dot / total
|
1481 |
+
return {self.main_score: result}
|
1482 |
+
|
1483 |
+
|
1484 |
+
class RetrievalAtK(RetrievalMetric):
|
1485 |
+
k_list: List[int]
|
1486 |
+
main_score: str = None
|
1487 |
+
reduction_map: Dict[str, List[str]] = None
|
1488 |
+
|
1489 |
+
def prepare(self):
|
1490 |
+
super().prepare()
|
1491 |
+
self.main_score = self.score_name("match", self.k_list[0])
|
1492 |
+
self.ci_scores = [
|
1493 |
+
self.score_name(measure, k)
|
1494 |
+
for measure in ["precision", "recall", "match"]
|
1495 |
+
for k in self.k_list
|
1496 |
+
]
|
1497 |
+
self.reduction_map = {"mean": self.ci_scores}
|
1498 |
+
|
1499 |
+
@staticmethod
|
1500 |
+
def score_name(measure: str, k: int):
|
1501 |
+
return f"{measure}_at_{k}"
|
1502 |
+
|
1503 |
+
def _compute(
|
1504 |
+
self,
|
1505 |
+
relevance_at_k,
|
1506 |
+
relevance_sum_at_k,
|
1507 |
+
precision_at_k,
|
1508 |
+
recall_at_k,
|
1509 |
+
match_at_k,
|
1510 |
+
rank,
|
1511 |
+
) -> dict:
|
1512 |
+
result = {}
|
1513 |
+
for measure_array, measure_name in [
|
1514 |
+
(precision_at_k, "precision"),
|
1515 |
+
(recall_at_k, "recall"),
|
1516 |
+
(match_at_k, "match"),
|
1517 |
+
]:
|
1518 |
+
max_k = max(measure_array.keys())
|
1519 |
+
for k in self.k_list:
|
1520 |
+
result[self.score_name(measure_name, k)] = measure_array[min(k, max_k)]
|
1521 |
+
return result
|
1522 |
+
|
1523 |
+
|
1524 |
+
class KPA(CustomF1):
|
1525 |
+
def get_element_group(self, element, additional_input):
|
1526 |
+
return additional_input["keypoint"]
|
1527 |
+
|
1528 |
+
def get_element_representation(self, element, additional_input):
|
1529 |
+
return additional_input["keypoint"]
|
1530 |
+
|
1531 |
+
def should_ignore_element(self, element, additional_input):
|
1532 |
+
return element == "none"
|