-
Notifications
You must be signed in to change notification settings - Fork 1
/
distance_lib.py
301 lines (264 loc) · 11.5 KB
/
distance_lib.py
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
import logging
import numpy as np
from abc import ABC, abstractclassmethod
from scipy import stats
from scipy.stats._stats_py import PearsonRResult
from sklearn.utils import column_or_1d
from sklearn.utils.validation import check_consistent_length
from typing import Dict, List, Tuple, Union
class DistanceMetric(ABC):
def __init__(self, eps: float = 0.001):
self.eps = eps
@abstractclassmethod
def distance(
self,
expected: Union[List[float], List[List[float]]],
predicted: Union[List[float], List[List[float]]],
**kwargs,
) -> Union[Tuple[float, float], float]:
pass
def _add_eps_for_constant_array(self, arr: np.ndarray):
# Correlation coefficients return NaN when an array has 0 variance.
# This helper function adds eps to one element in the array (if the array
# has 0 variance) to deal with this issue.
if len(arr) == 0:
return arr
if np.all(np.equal(arr, arr[0])):
arr[-1] += self.eps
return arr
return arr
def _add_eps_for_constant_array_decorator(self, fn):
def wrapper_fn(expected, predicted, **kwargs):
e = self._add_eps_for_constant_array(expected)
p = self._add_eps_for_constant_array(predicted)
return fn(e, p, **kwargs)
return wrapper_fn
def _get_only_valid_values(self, expected, predicted):
valid_idx_predicted = np.logical_and(~np.isnan(predicted), ~np.isinf(predicted))
valid_idx_expected = np.logical_and(~np.isnan(expected), ~np.isinf(expected))
valid_idx = np.logical_and(valid_idx_expected, valid_idx_predicted)
return expected[valid_idx], predicted[valid_idx]
class RankingMetric(DistanceMetric):
def __init__(
self,
metric_type: str = "spearman", # choices are ['spearman', 'kendall']
):
super().__init__()
if metric_type == "spearman":
self.metric_fn = stats.spearmanr
elif metric_type == "kendall":
self.metric_fn = stats.kendalltau
else:
raise ValueError(
f"No metric type named {metric_type}. Possible choices: ['spearman', 'kendall]"
)
self.metric_fn = self._add_eps_for_constant_array_decorator(self.metric_fn)
def distance(
self,
expected: Union[List[float], List[List[float]]],
predicted: Union[List[float], List[List[float]]],
**kwargs,
) -> Union[Tuple[float, List[float], List[float]], Tuple[float, float]]:
"""
If |expected| and |predicted| are lists of lists, then we calculate the mean correlation
for each pair of rankings. If they are non-nested lists, then we calculate a single
correlation across the entire list.
If |expected| and |predicted| are non-nested lists, then return (statistic, list of all p-values).
Otherwise, return (mean statistic, list of all p-values, list of all statistics).
"""
if len(expected) == 0 or len(predicted) == 0:
raise ValueError(f"Expected and predicted must be non-empty.")
assert len(expected) == len(
predicted
), "Expected must be the same length as predicted."
if isinstance(expected[0], list):
resps = [
self.metric_fn(e, p, **kwargs) for e, p in zip(expected, predicted)
]
stat_means = np.array([r.statistic for r in resps])
stat_mean = np.mean(stat_means[~np.isnan(stat_means)])
pvalues = [r.pvalue for r in resps]
return (stat_mean, pvalues, list(stat_means))
else:
resp = self.metric_fn(expected, predicted, **kwargs)
stat_mean = resp.statistic
pvalues = [resp.pvalue]
return (stat_mean, pvalues)
class Correlation(DistanceMetric):
def __init__(self):
super().__init__()
self.metric_fn = self._add_eps_for_constant_array_decorator(
self._error_handler(self._return_nan_for_invalid_inputs(stats.pearsonr))
)
def _return_nan_for_invalid_inputs(self, fn):
def wrapper_fn(expected, predicted, **kwargs):
if np.any(np.logical_or(np.isnan(expected), np.isinf(expected))) or np.any(
np.logical_or(np.isnan(predicted), np.isinf(predicted))
):
return PearsonRResult(
statistic=np.nan,
pvalue=np.nan,
alternative="two-sided",
n=len(expected),
x=expected,
y=predicted,
)
return fn(expected, predicted, **kwargs)
return wrapper_fn
def _error_handler(self, fn):
def wrapper_fn(expected, predicted, **kwargs):
try:
return fn(expected, predicted, **kwargs)
except ValueError as err:
if np.any(np.isnan(expected)):
logging.warning(f"NaNs in expected: {expected}")
if np.any(np.isnan(predicted)):
logging.warning(f"NaNs in predicted: {predicted}")
if np.any(np.isinf(expected)):
logging.warning(f"Infs in expected: {expected}")
if np.any(np.isinf(predicted)):
logging.warning(f"Infs in predicted: {predicted}")
logging.warning(
f"ValueError while computing Pearson correlation. Expected: {expected}\nPredicted: {predicted}"
)
raise err
return wrapper_fn
def distance(
self,
expected: Union[List[float], List[List[float]]],
predicted: Union[List[float], List[List[float]]],
**kwargs,
) -> Union[Tuple[float, List[float], List[float]], Tuple[float, float]]:
"""
If |expected| and |predicted| are lists of lists, then we calculate the mean correlation
for each pair of list items. If they are non-nested lists, then we calculate a single
correlation across the entire list.
If |expected| and |predicted| are non-nested lists, then return (statistic, list of all p-values).
Otherwise, return (mean statistic, list of all p-values, list of all statistics).
"""
if len(expected) == 0 or len(predicted) == 0:
raise ValueError(f"Expected and predicted must be non-empty.")
assert len(expected) == len(
predicted
), "Expected must be the same length as predicted."
if isinstance(expected[0], list) or isinstance(expected[0], np.ndarray):
resps = [self.metric_fn(e, p) for e, p in zip(expected, predicted)]
stat_means = np.array([r.statistic for r in resps])
stat_mean = np.mean(stat_means[~np.isnan(stat_means)])
pvalues = [r.pvalue for r in resps]
return (stat_mean, pvalues, list(stat_means))
else:
resp = self.metric_fn(expected, predicted)
stat_mean = resp.statistic
pvalues = [resp.pvalue]
return (stat_mean, pvalues)
class MSE(DistanceMetric):
def distance(
self, expected: List[float], predicted: List[float], **kwargs
) -> Union[Tuple[float, float], float]:
"""
Returns (mean square error, std square error)
"""
squared_error = [(e - p) ** 2 for e, p in zip(expected, predicted)]
return (np.mean(squared_error), np.std(squared_error))
class Accuracy(DistanceMetric):
def distance(
self, expected: List[float], predicted: List[float], **kwargs
) -> Union[Tuple[float, float], float]:
"""
Returns mean accuracy and also individual values.
"""
is_correct = []
for e, p in zip(expected, predicted):
if (not isinstance(e, list) and np.isnan(e)) or (
not isinstance(p, list) and np.isnan(p)
):
is_correct.append(np.nan)
else:
is_correct.append(1 if e == p else 0)
return np.nanmean(is_correct), is_correct
class LogDifference(DistanceMetric):
def __init__(self, eps: float = 0.001) -> None:
self.eps = eps
def _replace_zeros_with_eps(self, arr: List[float]) -> List[float]:
return [self.eps if x == 0 else x for x in arr]
def distance(
self, expected: List[float], predicted: List[float], **kwargs
) -> Union[Tuple[float, float], float]:
"""
Returns (expected abs. log difference, std. dev. of abs. log difference)
"""
expected = self._replace_zeros_with_eps(expected)
predicted = self._replace_zeros_with_eps(predicted)
abs_log_diffs = [
np.abs(np.log(e) - np.log(p)) for e, p in zip(expected, predicted)
]
return (np.mean(abs_log_diffs), np.std(abs_log_diffs))
class ExpectedCalibrationError(DistanceMetric):
def distance(
self,
expected: List[float],
predicted: List[float],
n_bins: int = 10,
strategy: str = "uniform", # can be ['uniform', 'quantile']
**kwargs,
) -> Union[Tuple[float, float], float]:
"""
Returns expected calibration error.
|expected| contains the true labels, and |predicted| contains the predicted probabilities of the positive class.
"""
y_true = column_or_1d(expected)
y_prob = column_or_1d(predicted)
check_consistent_length(y_true, y_prob)
if y_prob.min() < 0 or y_prob.max() > 1:
raise ValueError("y_prob has values outside [0, 1].")
labels = np.unique(y_true)
if len(labels) > 2:
raise ValueError(
f"Only binary classification is supported. Provided labels {labels}."
)
y_true = y_true == 1
if strategy == "quantile": # Determine bin edges by distribution of data
quantiles = np.linspace(0, 1, n_bins + 1)
bins = np.percentile(y_prob, quantiles * 100)
elif strategy == "uniform":
bins = np.linspace(0.0, 1.0, n_bins + 1)
else:
raise ValueError(
"Invalid entry to 'strategy' input. Strategy "
"must be either 'quantile' or 'uniform'."
)
binids = np.searchsorted(bins[1:-1], y_prob)
bin_sums = np.bincount(binids, weights=y_prob, minlength=len(bins))
bin_true = np.bincount(binids, weights=y_true, minlength=len(bins))
bin_total = np.bincount(binids, minlength=len(bins))
nonzero = bin_total != 0
prob_true = bin_true[nonzero] / bin_total[nonzero]
prob_pred = bin_sums[nonzero] / bin_total[nonzero]
bin_sizes = bin_total[nonzero] / len(binids)
abs_errs = [
np.abs(p_true - p_pred) for p_true, p_pred in zip(prob_true, prob_pred)
]
binned_ece = np.sum(
[bin_size * abs_err for abs_err, bin_size in zip(abs_errs, bin_sizes)]
)
std_ece = np.sqrt(
np.sum(
[
bin_size * (abs_err - binned_ece) ** 2
for bin_size, abs_err in zip(bin_sizes, abs_errs)
]
)
)
return (binned_ece, std_ece)
class AbsoluteDifferenceMeans(DistanceMetric):
def distance(
self, expected: List[float], predicted: List[float], **kwargs
) -> Union[Tuple[float, float], float]:
"""
Returns |mean(expected) - mean(predicted)|
"""
expected, predicted = self._get_only_valid_values(
np.array(expected), np.array(predicted)
)
return np.abs(np.mean(expected) - np.mean(predicted))