File size: 11,032 Bytes
256a159 |
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 |
import os
import random
from typing import List
import evaluate
import numpy as np
from datasets import Dataset
from opencompass.registry import ICL_EVALUATORS
from .icl_base_evaluator import BaseEvaluator
class HuggingfaceEvaluator(BaseEvaluator):
"""Use huggingface evaluate module to calculate the target metrics.
Args:
metric (str): Metric name in evaluate module.
seed (int): There exists some randomness during the calculation of some
metrics, thus we set a fixed random seed for reproducing. Defaults
to 0.
"""
def __init__(self, metric: str, seed: int = 0) -> None:
self.metric = metric
self.seed = seed
super().__init__()
def _preprocess(self, predictions: List, references: List) -> dict:
"""Preprocess the final predictions and references to needed format.
Args:
predictions (List): List of predictions of each sample.
references (List): List of targets for each sample.
Returns:
dict: preprocessed results.
"""
return {
'predictions': predictions,
'references': references,
}
def _postprocess(self, scores: dict) -> dict:
"""Postprocess for final scores.
Args:
scores (dict): Dict of calculated scores of metrics.
Returns:
dict: postprocessed scores.
"""
return scores
def score(self, predictions: List, references: List) -> dict:
"""Calculate scores.
Args:
predictions (List): List of predictions of each sample.
references (List): List of targets for each sample.
Returns:
dict: calculated scores.
"""
random_state = random.getstate()
np_random_state = np.random.get_state()
random.seed(self.seed)
np.random.seed(self.seed)
if len(predictions) != len(references):
return {
'error':
'predictions and references have different '
f'length. len(predictions): {len(predictions)}, '
f'len(references): {len(references)}'
}
# use codes pre-downloaded to opencompass repo, avoid downloading
local_path = os.path.join(os.path.dirname(os.path.abspath(__file__)),
'hf_metrics', self.metric + '.py')
if os.path.exists(local_path):
metric = evaluate.load(local_path)
else:
metric = evaluate.load(self.metric)
scores = metric.compute(**self._preprocess(predictions, references))
result = self._postprocess(scores)
random.setstate(random_state)
np.random.set_state(np_random_state)
return result
@ICL_EVALUATORS.register_module()
class AccEvaluator(HuggingfaceEvaluator):
"""Accuracy evaluator."""
def __init__(self) -> None:
super().__init__(metric='accuracy')
def _preprocess(self, predictions: List, references: List) -> dict:
"""Preprocess the final predictions and references to needed format.
Args:
predictions (List): List of predictions of each sample.
references (List): List of targets for each sample.
Returns:
dict: preprocessed results.
"""
mapping_to_int_dict = {
label: idx
for idx, label in enumerate(set(map(str, references)))
}
pred_set = set(predictions)
for pred in pred_set:
if str(pred) not in mapping_to_int_dict.keys():
mapping_to_int_dict[str(pred)] = len(mapping_to_int_dict)
golds = [mapping_to_int_dict[str(gold)] for gold in references]
preds = [mapping_to_int_dict[str(pred)] for pred in predictions]
return {
'predictions': preds,
'references': golds,
}
def _postprocess(self, scores: dict) -> dict:
"""Postprocess for final scores.
Args:
scores (dict): Dict of calculated scores of metrics.
Returns:
dict: postprocessed scores.
"""
scores['accuracy'] *= 100
return scores
@ICL_EVALUATORS.register_module()
class AccContaminationEvaluator(AccEvaluator):
"""Accuracy evaluator."""
def score(self, predictions: List, references: List,
test_set: Dataset) -> dict:
# group the predictions and references by their contamination status
clean_predictions, clean_references = [], []
input_contaminated_predictions, input_contaminated_references = [], []
input_and_label_contaminated_predictions, \
input_and_label_contaminated_references = [], []
for pred, ref, is_clean in zip(predictions, references,
test_set['is_clean']):
if is_clean == 'clean':
clean_predictions.append(pred)
clean_references.append(ref)
elif is_clean == 'input contamination':
input_contaminated_predictions.append(pred)
input_contaminated_references.append(ref)
elif is_clean == 'input-and-label contamination':
input_and_label_contaminated_predictions.append(pred)
input_and_label_contaminated_references.append(ref)
clean_results = super().score(clean_predictions, clean_references)
input_contaminated_results = super().score(
input_contaminated_predictions, input_contaminated_references)
input_and_label_contaminated_results = super().score(
input_and_label_contaminated_predictions,
input_and_label_contaminated_references)
# rename the keys of the results, add 'clean, 'input contaminated',
# 'input-and-label contaminated' as prefixes
clean_results = {f'{k} - clean': v for k, v in clean_results.items()}
input_contaminated_results = {
f'{k} - input contaminated': v
for k, v in input_contaminated_results.items()
}
input_and_label_contaminated_results = {
f'{k} - input-and-label contaminated': v
for k, v in input_and_label_contaminated_results.items()
}
return {
**clean_results,
**input_contaminated_results,
**input_and_label_contaminated_results
}
@ICL_EVALUATORS.register_module()
class RougeEvaluator(HuggingfaceEvaluator):
"""Rouge evaluator.
Note: this evaluator is not suitable for chinese datasets.
"""
def __init__(self) -> None:
super().__init__(metric='rouge')
def _postprocess(self, scores: dict) -> dict:
"""Postprocess for final scores.
Args:
scores (dict): Dict of calculated scores of metrics.
Returns:
dict: postprocessed scores.
"""
return {k: v * 100 for k, v in scores.items()}
@ICL_EVALUATORS.register_module()
class BleuEvaluator(HuggingfaceEvaluator):
"""Bleu evaluator."""
def __init__(self) -> None:
super().__init__(metric='sacrebleu')
class BleuFloresEvaluator(HuggingfaceEvaluator):
"""Bleu evaluator using flores200 tokenize."""
def __init__(self) -> None:
super().__init__(metric='sacrebleu')
def _preprocess(self, predictions: List, references: List) -> dict:
return {
'predictions': predictions,
'references': references,
'tokenize': 'flores200',
}
@ICL_EVALUATORS.register_module()
class MccEvaluator(AccEvaluator):
"""Matthews correlation evaluator."""
def __init__(self) -> None:
super(AccEvaluator, self).__init__(metric='matthews_correlation')
def _postprocess(self, scores: dict) -> dict:
"""Postprocess for final scores.
Args:
scores (dict): Dict of calculated scores of metrics.
Returns:
dict: postprocessed scores.
"""
scores['matthews_correlation'] *= 100
return scores
@ICL_EVALUATORS.register_module()
class SquadEvaluator(HuggingfaceEvaluator):
"""Squad evaluator."""
def __init__(self) -> None:
super().__init__(metric='squad')
def _preprocess(self, predictions: List, references: List) -> dict:
"""Preprocess the final predictions and references to needed format.
Args:
predictions (List): List of predictions of each sample.
references (List): List of targets for each sample.
Returns:
dict: preprocessed results.
"""
p_list = [{
'prediction_text': pred.split('\n')[0],
'id': str(i)
} for i, pred in enumerate(predictions)]
r_list = [{
'answers': {
'answer_start': [0],
'text': [ref]
},
'id': str(i)
} for i, ref in enumerate(references)]
return {
'predictions': p_list,
'references': r_list,
}
def _postprocess(self, scores: dict) -> dict:
"""Postprocess for final scores.
Args:
scores (dict): Dict of calculated scores of metrics.
Returns:
dict: postprocessed scores.
"""
return scores['f1']
@ICL_EVALUATORS.register_module()
class EDAccEvaluator(AccEvaluator):
"""Edit distance based accuracy evaluator.
This implementation requires the un-postprocessed outputs from the model,
and the reference list where each item is structured as:
.. code-block:: python
{
'candidates': [], # a list of informative answer candidates
'label': 0, # the index of the gold answer
}
It always matches the model's output to a valid answer with the citerion
as the minimum editing distance.
"""
def __init__(self) -> None:
super().__init__()
from rapidfuzz.distance import Levenshtein
self.dist = Levenshtein.distance
def _preprocess(self, predictions: List, references: List) -> dict:
"""Preprocess the final predictions and references to needed format.
Args:
predictions (List): List of predictions of each sample.
references (List): List of targets for each sample.
Returns:
dict: preprocessed results.
"""
preds = []
golds = []
for i in range(len(predictions)):
pred, ref = predictions[i], references[i]
dists = []
for cands in ref['candidates']:
if isinstance(cands, str):
d = self.dist(pred, cands)
else:
d = np.min([self.dist(pred, cand) for cand in cands])
dists.append(d)
preds.append(np.argmin(dists))
golds.append(ref['label'])
return {
'predictions': preds,
'references': golds,
}
|