-
Notifications
You must be signed in to change notification settings - Fork 387
/
calibration_error.py
365 lines (297 loc) · 16.5 KB
/
calibration_error.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
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
# Copyright The Lightning team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from typing import Optional, Tuple, Union
import torch
from torch import Tensor
from typing_extensions import Literal
from torchmetrics.functional.classification.confusion_matrix import (
_binary_confusion_matrix_format,
_binary_confusion_matrix_tensor_validation,
_multiclass_confusion_matrix_format,
_multiclass_confusion_matrix_tensor_validation,
)
from torchmetrics.utilities.enums import ClassificationTaskNoMultilabel
def _binning_bucketize(
confidences: Tensor, accuracies: Tensor, bin_boundaries: Tensor
) -> Tuple[Tensor, Tensor, Tensor]:
"""Compute calibration bins using ``torch.bucketize``. Use for ``pytorch >=1.6``.
Args:
confidences: The confidence (i.e. predicted prob) of the top1 prediction.
accuracies: 1.0 if the top-1 prediction was correct, 0.0 otherwise.
bin_boundaries: Bin boundaries separating the ``linspace`` from 0 to 1.
Returns:
tuple with binned accuracy, binned confidence and binned probabilities
"""
accuracies = accuracies.to(dtype=confidences.dtype)
acc_bin = torch.zeros(len(bin_boundaries), device=confidences.device, dtype=confidences.dtype)
conf_bin = torch.zeros(len(bin_boundaries), device=confidences.device, dtype=confidences.dtype)
count_bin = torch.zeros(len(bin_boundaries), device=confidences.device, dtype=confidences.dtype)
indices = torch.bucketize(confidences, bin_boundaries, right=True) - 1
count_bin.scatter_add_(dim=0, index=indices, src=torch.ones_like(confidences))
conf_bin.scatter_add_(dim=0, index=indices, src=confidences)
conf_bin = torch.nan_to_num(conf_bin / count_bin)
acc_bin.scatter_add_(dim=0, index=indices, src=accuracies)
acc_bin = torch.nan_to_num(acc_bin / count_bin)
prop_bin = count_bin / count_bin.sum()
return acc_bin, conf_bin, prop_bin
def _ce_compute(
confidences: Tensor,
accuracies: Tensor,
bin_boundaries: Union[Tensor, int],
norm: str = "l1",
debias: bool = False,
) -> Tensor:
"""Compute the calibration error given the provided bin boundaries and norm.
Args:
confidences: The confidence (i.e. predicted prob) of the top1 prediction.
accuracies: 1.0 if the top-1 prediction was correct, 0.0 otherwise.
bin_boundaries: Bin boundaries separating the ``linspace`` from 0 to 1.
norm: Norm function to use when computing calibration error. Defaults to "l1".
debias: Apply debiasing to L2 norm computation as in
`Verified Uncertainty Calibration`_. Defaults to False.
Raises:
ValueError: If an unsupported norm function is provided.
Returns:
Tensor: Calibration error scalar.
"""
if isinstance(bin_boundaries, int):
bin_boundaries = torch.linspace(0, 1, bin_boundaries + 1, dtype=confidences.dtype, device=confidences.device)
if norm not in {"l1", "l2", "max"}:
raise ValueError(f"Argument `norm` is expected to be one of 'l1', 'l2', 'max' but got {norm}")
with torch.no_grad():
acc_bin, conf_bin, prop_bin = _binning_bucketize(confidences, accuracies, bin_boundaries)
if norm == "l1":
return torch.sum(torch.abs(acc_bin - conf_bin) * prop_bin)
if norm == "max":
ce = torch.max(torch.abs(acc_bin - conf_bin))
if norm == "l2":
ce = torch.sum(torch.pow(acc_bin - conf_bin, 2) * prop_bin)
# NOTE: debiasing is disabled in the wrapper functions. This implementation differs from that in sklearn.
if debias:
# the order here (acc_bin - 1 ) vs (1 - acc_bin) is flipped from
# the equation in Verified Uncertainty Prediction (Kumar et al 2019)/
debias_bins = (acc_bin * (acc_bin - 1) * prop_bin) / (prop_bin * accuracies.size()[0] - 1)
ce += torch.sum(torch.nan_to_num(debias_bins)) # replace nans with zeros if nothing appeared in a bin
return torch.sqrt(ce) if ce > 0 else torch.tensor(0)
return ce
def _binary_calibration_error_arg_validation(
n_bins: int,
norm: Literal["l1", "l2", "max"] = "l1",
ignore_index: Optional[int] = None,
) -> None:
if not isinstance(n_bins, int) or n_bins < 1:
raise ValueError(f"Expected argument `n_bins` to be an integer larger than 0, but got {n_bins}")
allowed_norm = ("l1", "l2", "max")
if norm not in allowed_norm:
raise ValueError(f"Expected argument `norm` to be one of {allowed_norm}, but got {norm}.")
if ignore_index is not None and not isinstance(ignore_index, int):
raise ValueError(f"Expected argument `ignore_index` to either be `None` or an integer, but got {ignore_index}")
def _binary_calibration_error_tensor_validation(
preds: Tensor, target: Tensor, ignore_index: Optional[int] = None
) -> None:
_binary_confusion_matrix_tensor_validation(preds, target, ignore_index)
if not preds.is_floating_point():
raise ValueError(
"Expected argument `preds` to be floating tensor with probabilities/logits"
f" but got tensor with dtype {preds.dtype}"
)
def _binary_calibration_error_update(preds: Tensor, target: Tensor) -> Tuple[Tensor, Tensor]:
confidences, accuracies = preds, target
return confidences, accuracies
def binary_calibration_error(
preds: Tensor,
target: Tensor,
n_bins: int = 15,
norm: Literal["l1", "l2", "max"] = "l1",
ignore_index: Optional[int] = None,
validate_args: bool = True,
) -> Tensor:
r"""`Top-label Calibration Error`_ for binary tasks.
The expected calibration error can be used to quantify how well a given model is calibrated e.g. how well the
predicted output probabilities of the model matches the actual probabilities of the ground truth distribution.
Three different norms are implemented, each corresponding to variations on the calibration error metric.
.. math::
\text{ECE} = \sum_i^N b_i \|(p_i - c_i)\|, \text{L1 norm (Expected Calibration Error)}
.. math::
\text{MCE} = \max_{i} (p_i - c_i), \text{Infinity norm (Maximum Calibration Error)}
.. math::
\text{RMSCE} = \sqrt{\sum_i^N b_i(p_i - c_i)^2}, \text{L2 norm (Root Mean Square Calibration Error)}
Where :math:`p_i` is the top-1 prediction accuracy in bin :math:`i`, :math:`c_i` is the average confidence of
predictions in bin :math:`i`, and :math:`b_i` is the fraction of data points in bin :math:`i`. Bins are constructed
in an uniform way in the [0,1] range.
Accepts the following input tensors:
- ``preds`` (float tensor): ``(N, ...)``. Preds should be a tensor containing probabilities or logits for each
observation. If preds has values outside [0,1] range we consider the input to be logits and will auto apply
sigmoid per element.
- ``target`` (int tensor): ``(N, ...)``. Target should be a tensor containing ground truth labels, and therefore
only contain {0,1} values (except if `ignore_index` is specified). The value 1 always encodes the positive class.
Additional dimension ``...`` will be flattened into the batch dimension.
Args:
preds: Tensor with predictions
target: Tensor with true labels
n_bins: Number of bins to use when computing the metric.
norm: Norm used to compare empirical and expected probability bins.
ignore_index:
Specifies a target value that is ignored and does not contribute to the metric calculation
validate_args: bool indicating if input arguments and tensors should be validated for correctness.
Set to ``False`` for faster computations.
Example:
>>> from torchmetrics.functional.classification import binary_calibration_error
>>> preds = torch.tensor([0.25, 0.25, 0.55, 0.75, 0.75])
>>> target = torch.tensor([0, 0, 1, 1, 1])
>>> binary_calibration_error(preds, target, n_bins=2, norm='l1')
tensor(0.2900)
>>> binary_calibration_error(preds, target, n_bins=2, norm='l2')
tensor(0.2918)
>>> binary_calibration_error(preds, target, n_bins=2, norm='max')
tensor(0.3167)
"""
if validate_args:
_binary_calibration_error_arg_validation(n_bins, norm, ignore_index)
_binary_calibration_error_tensor_validation(preds, target, ignore_index)
preds, target = _binary_confusion_matrix_format(
preds, target, threshold=0.0, ignore_index=ignore_index, convert_to_labels=False
)
confidences, accuracies = _binary_calibration_error_update(preds, target)
return _ce_compute(confidences, accuracies, n_bins, norm)
def _multiclass_calibration_error_arg_validation(
num_classes: int,
n_bins: int,
norm: Literal["l1", "l2", "max"] = "l1",
ignore_index: Optional[int] = None,
) -> None:
if not isinstance(num_classes, int) or num_classes < 2:
raise ValueError(f"Expected argument `num_classes` to be an integer larger than 1, but got {num_classes}")
if not isinstance(n_bins, int) or n_bins < 1:
raise ValueError(f"Expected argument `n_bins` to be an integer larger than 0, but got {n_bins}")
allowed_norm = ("l1", "l2", "max")
if norm not in allowed_norm:
raise ValueError(f"Expected argument `norm` to be one of {allowed_norm}, but got {norm}.")
if ignore_index is not None and not isinstance(ignore_index, int):
raise ValueError(f"Expected argument `ignore_index` to either be `None` or an integer, but got {ignore_index}")
def _multiclass_calibration_error_tensor_validation(
preds: Tensor, target: Tensor, num_classes: int, ignore_index: Optional[int] = None
) -> None:
_multiclass_confusion_matrix_tensor_validation(preds, target, num_classes, ignore_index)
if not preds.is_floating_point():
raise ValueError(
"Expected argument `preds` to be floating tensor with probabilities/logits"
f" but got tensor with dtype {preds.dtype}"
)
def _multiclass_calibration_error_update(
preds: Tensor,
target: Tensor,
) -> Tuple[Tensor, Tensor]:
if not torch.all((preds >= 0) * (preds <= 1)):
preds = preds.softmax(1)
confidences, predictions = preds.max(dim=1)
accuracies = predictions.eq(target)
return confidences.float(), accuracies.float()
def multiclass_calibration_error(
preds: Tensor,
target: Tensor,
num_classes: int,
n_bins: int = 15,
norm: Literal["l1", "l2", "max"] = "l1",
ignore_index: Optional[int] = None,
validate_args: bool = True,
) -> Tensor:
r"""`Top-label Calibration Error`_ for multiclass tasks.
The expected calibration error can be used to quantify how well a given model is calibrated e.g. how well the
predicted output probabilities of the model matches the actual probabilities of the ground truth distribution.
Three different norms are implemented, each corresponding to variations on the calibration error metric.
.. math::
\text{ECE} = \sum_i^N b_i \|(p_i - c_i)\|, \text{L1 norm (Expected Calibration Error)}
.. math::
\text{MCE} = \max_{i} (p_i - c_i), \text{Infinity norm (Maximum Calibration Error)}
.. math::
\text{RMSCE} = \sqrt{\sum_i^N b_i(p_i - c_i)^2}, \text{L2 norm (Root Mean Square Calibration Error)}
Where :math:`p_i` is the top-1 prediction accuracy in bin :math:`i`, :math:`c_i` is the average confidence of
predictions in bin :math:`i`, and :math:`b_i` is the fraction of data points in bin :math:`i`. Bins are constructed
in an uniform way in the [0,1] range.
Accepts the following input tensors:
- ``preds`` (float tensor): ``(N, C, ...)``. Preds should be a tensor containing probabilities or logits for each
observation. If preds has values outside [0,1] range we consider the input to be logits and will auto apply
softmax per sample.
- ``target`` (int tensor): ``(N, ...)``. Target should be a tensor containing ground truth labels, and therefore
only contain values in the [0, n_classes-1] range (except if `ignore_index` is specified).
Additional dimension ``...`` will be flattened into the batch dimension.
Args:
preds: Tensor with predictions
target: Tensor with true labels
num_classes: Integer specifing the number of classes
n_bins: Number of bins to use when computing the metric.
norm: Norm used to compare empirical and expected probability bins.
ignore_index:
Specifies a target value that is ignored and does not contribute to the metric calculation
validate_args: bool indicating if input arguments and tensors should be validated for correctness.
Set to ``False`` for faster computations.
Example:
>>> from torchmetrics.functional.classification import multiclass_calibration_error
>>> preds = torch.tensor([[0.25, 0.20, 0.55],
... [0.55, 0.05, 0.40],
... [0.10, 0.30, 0.60],
... [0.90, 0.05, 0.05]])
>>> target = torch.tensor([0, 1, 2, 0])
>>> multiclass_calibration_error(preds, target, num_classes=3, n_bins=3, norm='l1')
tensor(0.2000)
>>> multiclass_calibration_error(preds, target, num_classes=3, n_bins=3, norm='l2')
tensor(0.2082)
>>> multiclass_calibration_error(preds, target, num_classes=3, n_bins=3, norm='max')
tensor(0.2333)
"""
if validate_args:
_multiclass_calibration_error_arg_validation(num_classes, n_bins, norm, ignore_index)
_multiclass_calibration_error_tensor_validation(preds, target, num_classes, ignore_index)
preds, target = _multiclass_confusion_matrix_format(preds, target, ignore_index, convert_to_labels=False)
confidences, accuracies = _multiclass_calibration_error_update(preds, target)
return _ce_compute(confidences, accuracies, n_bins, norm)
def calibration_error(
preds: Tensor,
target: Tensor,
task: Literal["binary", "multiclass"],
n_bins: int = 15,
norm: Literal["l1", "l2", "max"] = "l1",
num_classes: Optional[int] = None,
ignore_index: Optional[int] = None,
validate_args: bool = True,
) -> Tensor:
r"""`Top-label Calibration Error`_.
The expected calibration error can be used to quantify how well a given model is calibrated e.g. how well the
predicted output probabilities of the model matches the actual probabilities of the ground truth distribution.
Three different norms are implemented, each corresponding to variations on the calibration error metric.
.. math::
\text{ECE} = \sum_i^N b_i \|(p_i - c_i)\|, \text{L1 norm (Expected Calibration Error)}
.. math::
\text{MCE} = \max_{i} (p_i - c_i), \text{Infinity norm (Maximum Calibration Error)}
.. math::
\text{RMSCE} = \sqrt{\sum_i^N b_i(p_i - c_i)^2}, \text{L2 norm (Root Mean Square Calibration Error)}
Where :math:`p_i` is the top-1 prediction accuracy in bin :math:`i`, :math:`c_i` is the average confidence of
predictions in bin :math:`i`, and :math:`b_i` is the fraction of data points in bin :math:`i`. Bins are constructed
in an uniform way in the [0,1] range.
This function is a simple wrapper to get the task specific versions of this metric, which is done by setting the
``task`` argument to either ``'binary'`` or ``'multiclass'``. See the documentation of
:func:`~torchmetrics.functional.classification.binary_calibration_error` and
:func:`~torchmetrics.functional.classification.multiclass_calibration_error` for the specific details of
each argument influence and examples.
"""
task = ClassificationTaskNoMultilabel.from_str(task)
assert norm is not None # noqa: S101 # needed for mypy
if task == ClassificationTaskNoMultilabel.BINARY:
return binary_calibration_error(preds, target, n_bins, norm, ignore_index, validate_args)
if task == ClassificationTaskNoMultilabel.MULTICLASS:
if not isinstance(num_classes, int):
raise ValueError(f"`num_classes` is expected to be `int` but `{type(num_classes)} was passed.`")
return multiclass_calibration_error(preds, target, num_classes, n_bins, norm, ignore_index, validate_args)
raise ValueError(f"Expected argument `task` to either be `'binary'` or `'multiclass'` but got {task}")