-
Notifications
You must be signed in to change notification settings - Fork 273
/
calibration.py
339 lines (288 loc) · 11.7 KB
/
calibration.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
# Copyright 2019 Google LLC
#
# 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
#
# https://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.
"""Calibration related metrics."""
from typing import Any, Dict, Iterable, Optional
import apache_beam as beam
import numpy as np
from tensorflow_model_analysis.metrics import metric_types
from tensorflow_model_analysis.metrics import metric_util
from tensorflow_model_analysis.proto import config_pb2
CALIBRATION_NAME = 'calibration'
MEAN_LABEL_NAME = 'mean_label'
MEAN_PREDICTION_NAME = 'mean_prediction'
_WEIGHTED_LABELS_PREDICTIONS_EXAMPLES_NAME = (
'_weighted_labels_predictions_examples')
class MeanLabel(metric_types.Metric):
"""Mean label."""
def __init__(self, name: str = MEAN_LABEL_NAME):
"""Initializes mean label.
Args:
name: Metric name.
"""
super().__init__(
metric_util.merge_per_key_computations(_mean_label), name=name)
metric_types.register_metric(MeanLabel)
def _mean_label(
name: str = MEAN_LABEL_NAME,
eval_config: Optional[config_pb2.EvalConfig] = None,
model_name: str = '',
output_name: str = '',
sub_key: Optional[metric_types.SubKey] = None,
aggregation_type: Optional[metric_types.AggregationType] = None,
class_weights: Optional[Dict[int, float]] = None,
example_weighted: bool = False) -> metric_types.MetricComputations:
"""Returns metric computations for mean label."""
key = metric_types.MetricKey(
name=name,
model_name=model_name,
output_name=output_name,
sub_key=sub_key,
example_weighted=example_weighted)
# Make sure weighted_labels_predictions_examples are calculated.
computations = _weighted_labels_predictions_examples(
eval_config=eval_config,
model_name=model_name,
output_name=output_name,
sub_key=sub_key,
aggregation_type=aggregation_type,
class_weights=class_weights,
example_weighted=example_weighted)
weighted_labels_predictions_key = computations[-1].keys[-1]
def result(
metrics: Dict[metric_types.MetricKey, Any]
) -> Dict[metric_types.MetricKey, Any]:
"""Returns mean label."""
metric = metrics[weighted_labels_predictions_key]
if np.isclose(metric.total_weighted_examples, 0.0):
value = float('nan')
else:
value = metric.total_weighted_labels / metric.total_weighted_examples
return {key: value}
derived_computation = metric_types.DerivedMetricComputation(
keys=[key], result=result)
computations.append(derived_computation)
return computations
class MeanPrediction(metric_types.Metric):
"""Mean prediction."""
def __init__(self, name: str = MEAN_PREDICTION_NAME):
"""Initializes mean prediction.
Args:
name: Metric name.
"""
super().__init__(
metric_util.merge_per_key_computations(_mean_prediction), name=name)
metric_types.register_metric(MeanPrediction)
def _mean_prediction(
name: str = MEAN_PREDICTION_NAME,
eval_config: Optional[config_pb2.EvalConfig] = None,
model_name: str = '',
output_name: str = '',
sub_key: Optional[metric_types.SubKey] = None,
aggregation_type: Optional[metric_types.AggregationType] = None,
class_weights: Optional[Dict[int, float]] = None,
example_weighted: bool = False) -> metric_types.MetricComputations:
"""Returns metric computations for mean prediction."""
key = metric_types.MetricKey(
name=name,
model_name=model_name,
output_name=output_name,
sub_key=sub_key,
example_weighted=example_weighted)
# Make sure weighted_labels_predictions_examples are calculated.
computations = _weighted_labels_predictions_examples(
eval_config=eval_config,
model_name=model_name,
output_name=output_name,
sub_key=sub_key,
aggregation_type=aggregation_type,
class_weights=class_weights,
example_weighted=example_weighted)
weighted_labels_predictions_key = computations[-1].keys[-1]
def result(
metrics: Dict[metric_types.MetricKey, Any]
) -> Dict[metric_types.MetricKey, Any]:
"""Returns mean prediction."""
metric = metrics[weighted_labels_predictions_key]
if np.isclose(metric.total_weighted_examples, 0.0):
value = float('nan')
else:
value = metric.total_weighted_predictions / metric.total_weighted_examples
return {key: value}
derived_computation = metric_types.DerivedMetricComputation(
keys=[key], result=result)
computations.append(derived_computation)
return computations
class Calibration(metric_types.Metric):
"""Calibration.
Calibration in this context is defined as the total weighted predictions /
total weighted labels.
"""
def __init__(self, name: str = CALIBRATION_NAME):
"""Initializes calibration.
Args:
name: Metric name.
"""
super().__init__(
metric_util.merge_per_key_computations(_calibration), name=name)
metric_types.register_metric(Calibration)
def _calibration(
name: str = CALIBRATION_NAME,
eval_config: Optional[config_pb2.EvalConfig] = None,
model_name: str = '',
output_name: str = '',
sub_key: Optional[metric_types.SubKey] = None,
aggregation_type: Optional[metric_types.AggregationType] = None,
class_weights: Optional[Dict[int, float]] = None,
example_weighted: bool = False) -> metric_types.MetricComputations:
"""Returns metric computations for calibration."""
key = metric_types.MetricKey(
name=name,
model_name=model_name,
output_name=output_name,
sub_key=sub_key,
example_weighted=example_weighted)
# Make sure weighted_labels_predictions_examples are calculated.
computations = _weighted_labels_predictions_examples(
eval_config=eval_config,
model_name=model_name,
output_name=output_name,
sub_key=sub_key,
aggregation_type=aggregation_type,
class_weights=class_weights,
example_weighted=example_weighted)
weighted_labels_predictions_key = computations[-1].keys[-1]
def result(
metrics: Dict[metric_types.MetricKey, Any]
) -> Dict[metric_types.MetricKey, Any]:
"""Returns calibration."""
metric = metrics[weighted_labels_predictions_key]
if np.isclose(metric.total_weighted_labels, 0.0):
value = float('nan')
else:
value = metric.total_weighted_predictions / metric.total_weighted_labels
return {key: value}
derived_computation = metric_types.DerivedMetricComputation(
keys=[key], result=result)
computations.append(derived_computation)
return computations
def _weighted_labels_predictions_examples(
name: str = _WEIGHTED_LABELS_PREDICTIONS_EXAMPLES_NAME,
eval_config: Optional[config_pb2.EvalConfig] = None,
model_name: str = '',
output_name: str = '',
sub_key: Optional[metric_types.SubKey] = None,
aggregation_type: Optional[metric_types.AggregationType] = None,
class_weights: Optional[Dict[int, float]] = None,
example_weighted: bool = False) -> metric_types.MetricComputations:
"""Returns metric computations for weighted labels, predictions, and examples.
Args:
name: Metric name.
eval_config: Eval config.
model_name: Optional model name (if multi-model evaluation).
output_name: Optional output name (if multi-output model type).
sub_key: Optional sub key.
aggregation_type: Optional aggregation type.
class_weights: Optional class weights to apply to multi-class / multi-label
labels and predictions prior to flattening (when micro averaging is used).
example_weighted: True if example weights should be applied.
"""
key = metric_types.MetricKey(
name=name,
model_name=model_name,
output_name=output_name,
sub_key=sub_key,
example_weighted=example_weighted)
return [
metric_types.MetricComputation(
keys=[key],
preprocessors=None, # Use default
combiner=_WeightedLabelsPredictionsExamplesCombiner(
key,
eval_config=eval_config,
aggregation_type=aggregation_type,
class_weights=class_weights,
example_weighted=example_weighted))
]
class _WeightedLabelsPredictionsExamples:
"""Total weighted labels, predictions, and examples."""
__slots__ = [
'total_weighted_labels', 'total_weighted_predictions',
'total_weighted_examples'
]
def __init__(self):
"""Initializes accumulator."""
self.total_weighted_labels = 0.0
self.total_weighted_predictions = 0.0
self.total_weighted_examples = 0.0
class _WeightedLabelsPredictionsExamplesCombiner(beam.CombineFn):
"""Computes weighted labels, predictions, and examples."""
def __init__(self, key: metric_types.MetricKey,
eval_config: Optional[config_pb2.EvalConfig],
aggregation_type: Optional[metric_types.AggregationType],
class_weights: Optional[Dict[int,
float]], example_weighted: bool):
self._key = key
self._eval_config = eval_config
self._aggregation_type = aggregation_type
self._class_weights = class_weights
self._example_weighted = example_weighted
def create_accumulator(self) -> _WeightedLabelsPredictionsExamples:
return _WeightedLabelsPredictionsExamples()
def add_input(
self, accumulator: _WeightedLabelsPredictionsExamples,
element: metric_types.StandardMetricInputs
) -> _WeightedLabelsPredictionsExamples:
for label, prediction, example_weight in (
metric_util.to_label_prediction_example_weight(
element,
eval_config=self._eval_config,
model_name=self._key.model_name,
output_name=self._key.output_name,
sub_key=self._key.sub_key,
aggregation_type=self._aggregation_type,
class_weights=self._class_weights,
example_weighted=self._example_weighted,
allow_none=True)):
example_weight = float(example_weight)
accumulator.total_weighted_examples += example_weight
if label is not None and len(label):
if self._key.sub_key and self._key.sub_key.top_k is not None:
for i in range(self._key.sub_key.top_k):
weighted_label = label[i] * example_weight
else:
weighted_label = float(label) * example_weight
accumulator.total_weighted_labels += weighted_label
if prediction is not None and len(label):
if self._key.sub_key and self._key.sub_key.top_k is not None:
for i in range(self._key.sub_key.top_k):
weighted_prediction = prediction[i] * example_weight
else:
weighted_prediction = float(prediction) * example_weight
accumulator.total_weighted_predictions += weighted_prediction
return accumulator
def merge_accumulators(
self, accumulators: Iterable[_WeightedLabelsPredictionsExamples]
) -> _WeightedLabelsPredictionsExamples:
accumulators = iter(accumulators)
result = next(accumulators)
for accumulator in accumulators:
result.total_weighted_labels += accumulator.total_weighted_labels
result.total_weighted_predictions += (
accumulator.total_weighted_predictions)
result.total_weighted_examples += accumulator.total_weighted_examples
return result
def extract_output(
self, accumulator: _WeightedLabelsPredictionsExamples
) -> Dict[metric_types.MetricKey, _WeightedLabelsPredictionsExamples]:
return {self._key: accumulator}