forked from uncertainty-toolbox/uncertainty-toolbox
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtest_recalibration.py
293 lines (243 loc) · 9.94 KB
/
test_recalibration.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
"""
Tests for recalibration procedures.
"""
import random
import numpy as np
import pytest
from uncertainty_toolbox.recalibration import (
iso_recal,
optimize_recalibration_ratio,
get_std_recalibrator,
get_quantile_recalibrator,
get_interval_recalibrator,
)
from uncertainty_toolbox.metrics_calibration import (
root_mean_squared_calibration_error,
mean_absolute_calibration_error,
miscalibration_area,
get_proportion_lists_vectorized,
get_prediction_interval,
get_quantile,
)
@pytest.fixture
def supply_test_set():
y_pred = np.arange(5000) / 5000.0
y_std = np.arange(50, 5050) / 1000.0
y_true = np.arange(5000) / 5000.0 + 0.5
return y_pred, y_std, y_true
def test_recal_model_mace_criterion_on_test_set(supply_test_set):
"""
Test recalibration on mean absolute calibration error on the test set
for some dummy values.
"""
test_mace = mean_absolute_calibration_error(
*supply_test_set, num_bins=100, vectorized=True, recal_model=None
)
test_exp_props, test_obs_props = get_proportion_lists_vectorized(
*supply_test_set, num_bins=100, recal_model=None
)
recal_model = iso_recal(test_exp_props, test_obs_props)
recal_test_mace = mean_absolute_calibration_error(
*supply_test_set, num_bins=100, vectorized=True, recal_model=recal_model
)
recal_exp_props = recal_model.predict(test_obs_props)
assert np.abs(test_mace - 0.24206060606060598) < 1e-2
assert np.abs(recal_test_mace - 0.003035353535353514) < 1e-2
for idx in range(1, recal_exp_props.shape[0]):
assert recal_exp_props[idx - 1] <= recal_exp_props[idx]
def test_recal_model_rmce_criterion_on_test_set(supply_test_set):
"""
Test recalibration on root mean squared calibration error on the test set
for some dummy values.
"""
test_rmsce = root_mean_squared_calibration_error(
*supply_test_set, num_bins=100, vectorized=True, recal_model=None
)
test_exp_props, test_obs_props = get_proportion_lists_vectorized(
*supply_test_set, num_bins=100, recal_model=None
)
recal_model = iso_recal(test_exp_props, test_obs_props)
recal_test_rmsce = root_mean_squared_calibration_error(
*supply_test_set, num_bins=100, vectorized=True, recal_model=recal_model
)
recal_exp_props = recal_model.predict(test_obs_props)
assert np.abs(test_rmsce - 0.28741418862839013) < 1e-2
assert np.abs(recal_test_rmsce - 0.003981861230030349) < 1e-2
for idx in range(1, recal_exp_props.shape[0]):
assert recal_exp_props[idx - 1] <= recal_exp_props[idx]
def test_recal_model_miscal_area_criterion_on_test_set(supply_test_set):
"""
Test recalibration on miscalibration area on the test set
for some dummy values.
"""
test_miscal_area = miscalibration_area(
*supply_test_set, num_bins=100, vectorized=True, recal_model=None
)
test_exp_props, test_obs_props = get_proportion_lists_vectorized(
*supply_test_set, num_bins=100, recal_model=None
)
recal_model = iso_recal(test_exp_props, test_obs_props)
recal_test_miscal_area = miscalibration_area(
*supply_test_set, num_bins=100, vectorized=True, recal_model=recal_model
)
recal_exp_props = recal_model.predict(test_obs_props)
assert np.abs(test_miscal_area - 0.24426139657444004) < 1e-2
assert np.abs(recal_test_miscal_area - 0.0029569160997732244) < 1e-2
for idx in range(1, recal_exp_props.shape[0]):
assert recal_exp_props[idx - 1] <= recal_exp_props[idx]
def test_optimize_recalibration_ratio_mace_criterion(supply_test_set):
"""
Test standard deviation recalibration on mean absolute calibration error
on the test set for some dummy values.
"""
random.seed(0)
np.random.seed(seed=0)
y_pred, y_std, y_true = supply_test_set
ma_cal_ratio = optimize_recalibration_ratio(
y_pred, y_std, y_true, criterion="ma_cal"
)
recal_ma_cal = mean_absolute_calibration_error(y_pred, ma_cal_ratio * y_std, y_true)
recal_rms_cal = root_mean_squared_calibration_error(
y_pred, ma_cal_ratio * y_std, y_true
)
recal_miscal = miscalibration_area(y_pred, ma_cal_ratio * y_std, y_true)
assert np.abs(ma_cal_ratio - 0.33215708813773176) < 1e-2
assert np.abs(recal_ma_cal - 0.06821616161616162) < 1e-2
assert np.abs(recal_rms_cal - 0.08800130087804929) < 1e-2
assert np.abs(recal_miscal - 0.06886262626262629) < 1e-2
def test_optimize_recalibration_ratio_rmce_criterion(supply_test_set):
"""
Test standard deviation recalibration on root mean squared calibration error
on the test set for some dummy values.
"""
random.seed(0)
np.random.seed(seed=0)
y_pred, y_std, y_true = supply_test_set
rms_cal_ratio = optimize_recalibration_ratio(
y_pred, y_std, y_true, criterion="rms_cal"
)
recal_ma_cal = mean_absolute_calibration_error(
y_pred, rms_cal_ratio * y_std, y_true
)
recal_rms_cal = root_mean_squared_calibration_error(
y_pred, rms_cal_ratio * y_std, y_true
)
recal_miscal = miscalibration_area(y_pred, rms_cal_ratio * y_std, y_true)
assert np.abs(rms_cal_ratio - 0.34900989073212507) < 1e-2
assert np.abs(recal_ma_cal - 0.06945555555555555) < 1e-2
assert np.abs(recal_rms_cal - 0.08570902541177935) < 1e-2
assert np.abs(recal_miscal - 0.07011706864564003) < 1e-2
def test_optimize_recalibration_ratio_miscal_area_criterion(supply_test_set):
"""
Test standard deviation recalibration on miscalibration area
on the test set for some dummy values.
"""
random.seed(0)
np.random.seed(seed=0)
y_pred, y_std, y_true = supply_test_set
miscal_ratio = optimize_recalibration_ratio(
y_pred, y_std, y_true, criterion="miscal"
)
recal_ma_cal = mean_absolute_calibration_error(y_pred, miscal_ratio * y_std, y_true)
recal_rms_cal = root_mean_squared_calibration_error(
y_pred, miscal_ratio * y_std, y_true
)
recal_miscal = miscalibration_area(y_pred, miscal_ratio * y_std, y_true)
assert np.abs(miscal_ratio - 0.3321912522557988) < 1e-2
assert np.abs(recal_ma_cal - 0.06821616161616162) < 1e-2
assert np.abs(recal_rms_cal - 0.08800130087804929) < 1e-2
assert np.abs(recal_miscal - 0.06886262626262629) < 1e-2
def test_get_prediction_interval_recalibrated(supply_test_set):
"""
Test standard deviation recalibration on miscalibration area
on the test set for some dummy values.
"""
random.seed(0)
np.random.seed(seed=0)
y_pred, y_std, y_true = supply_test_set
test_exp_props, test_obs_props = get_proportion_lists_vectorized(
y_pred, y_std, y_true, num_bins=100, recal_model=None
)
recal_model = iso_recal(test_exp_props, test_obs_props)
test_quantile_prop_list = [
(0.01, 0.0, 0.0),
(0.25, 0.69, 0.25),
(0.5, 0.86, 0.5),
(0.75, 0.92, 0.75),
(0.99, 0.97, 0.97),
]
for q, test_orig_prop, test_recal_prop in test_quantile_prop_list:
orig_bounds = get_prediction_interval(y_pred, y_std, q, None)
recal_bounds = get_prediction_interval(y_pred, y_std, q, recal_model)
orig_prop = np.mean(
(orig_bounds.lower <= y_true) * (y_true <= orig_bounds.upper)
)
recal_prop = np.mean(
(recal_bounds.lower <= y_true) * (y_true <= recal_bounds.upper)
)
assert np.max(np.abs(test_orig_prop - orig_prop)) < 1e-2
assert np.max(np.abs(test_recal_prop - recal_prop)) < 1e-2
def test_get_std_recalibrator(supply_test_set):
"""
Test get_std_recalibration on the test set for some dummy values.
"""
random.seed(0)
np.random.seed(seed=0)
y_pred, y_std, y_true = supply_test_set
test_quantile_prop_list = [
(0.01, 0.00, 0.00),
(0.25, 0.06, 0.00),
(0.50, 0.56, 0.00),
(0.75, 0.74, 0.56),
(0.99, 0.89, 0.88),
]
std_recalibrator = get_std_recalibrator(y_pred, y_std, y_true)
for q, test_prop_in_pi, test_prop_under_q in test_quantile_prop_list:
y_std_recal = std_recalibrator(y_std)
pi = get_prediction_interval(y_pred, y_std_recal, q)
prop_in_pi = ((pi.lower <= y_true) * (y_true <= pi.upper)).mean()
quantile_bound = get_quantile(y_pred, y_std_recal, q)
prop_under_q = (quantile_bound >= y_true).mean()
assert np.max(np.abs(test_prop_in_pi - prop_in_pi)) < 5e-2
assert np.max(np.abs(test_prop_under_q - prop_under_q)) < 5e-2
def test_get_quantile_recalibrator(supply_test_set):
"""
Test get_std_recalibration on the test set for some dummy values.
"""
random.seed(0)
np.random.seed(seed=0)
y_pred, y_std, y_true = supply_test_set
test_quantile_prop_list = [
(0.01, 0.00),
(0.25, 0.00),
(0.50, 0.00),
(0.75, 0.00),
(0.99, 0.83),
]
quantile_recalibrator = get_quantile_recalibrator(y_pred, y_std, y_true)
for q, test_prop_under_q in test_quantile_prop_list:
quantile_bound_recal = quantile_recalibrator(y_pred, y_std, q)
assert all(np.isfinite(quantile_bound_recal))
prop_under_q_recal = (quantile_bound_recal >= y_true).mean()
assert np.max(np.abs(test_prop_under_q - prop_under_q_recal)) < 1e-2
def test_get_interval_recalibrator(supply_test_set):
"""
Test get_std_recalibration on the test set for some dummy values.
"""
random.seed(0)
np.random.seed(seed=0)
y_pred, y_std, y_true = supply_test_set
test_quantile_prop_list = [
(0.01, 0.00),
(0.25, 0.25),
(0.50, 0.50),
(0.75, 0.75),
(0.99, 0.97),
]
interval_recalibrator = get_interval_recalibrator(y_pred, y_std, y_true)
for q, test_prop_in_interval in test_quantile_prop_list:
interval_recal = interval_recalibrator(y_pred, y_std, q)
prop_in_interval_recal = (
(interval_recal.lower <= y_true) * (y_true <= interval_recal.upper)
).mean()
assert np.max(np.abs(test_prop_in_interval - prop_in_interval_recal)) < 1e-2