-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathtest_metrics.py
170 lines (143 loc) · 5.97 KB
/
test_metrics.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
import unittest
import numpy as np
from functools import partial
from pycalib.metrics import (accuracy, cross_entropy, brier_score,
binary_ECE, conf_ECE, classwise_ECE, full_ECE,
MCE, pECE)
from sklearn.preprocessing import label_binarize
# TODO add more test cases
class TestFunctions(unittest.TestCase):
def test_accuracy(self):
S = np.array([[0.1, 0.9], [0.6, 0.4]])
Y = np.array([[0, 1], [0, 1]])
acc = accuracy(Y, S)
self.assertAlmostEqual(acc, 0.5)
S = np.array([[0.1, 0.9], [0.6, 0.4]])
Y = np.array([[1, 0], [0, 1]])
acc = accuracy(Y, S)
self.assertAlmostEqual(acc, 0.0)
def test_cross_entropy(self):
S = np.array([[0.1, 0.9], [0.6, 0.4]])
Y = np.array([[0, 1], [0, 1]])
ce = cross_entropy(Y, S)
expected = - (np.log(0.9) + np.log(0.4))/2
self.assertAlmostEqual(ce, expected)
S = np.array([[0.1, 0.9], [0.6, 0.4]])
Y = np.array([[1, 0], [0, 1]])
ce = cross_entropy(Y, S)
expected = - (np.log(0.1) + np.log(0.4))/2
self.assertAlmostEqual(ce, expected)
def test_brier_score(self):
S = np.array([[0.1, 0.9], [0.6, 0.4]])
Y = np.array([[0, 1], [0, 1]])
bs = brier_score(Y, S)
expected = np.mean(np.abs(S - Y)**2)
self.assertAlmostEqual(bs, expected)
S = np.array([[0.1, 0.9], [0.6, 0.4]])
Y = np.array([[1, 0], [0, 1]])
bs = brier_score(Y, S)
expected = np.mean(np.abs(S - Y)**2)
self.assertAlmostEqual(bs, expected)
def test_binary_ece(self):
S = np.array([.6, .6, .6, .6, .6, .6, .6, .6, .6, .6])
y = np.array([1, 1, 1, 1, 1, 1, 0, 0, 0, 0])
ece = binary_ECE(y, S)
self.assertAlmostEqual(ece, 0)
def test_conf_ece(self):
S = np.array([[0.6, 0.2, 0.2]]*10)
y = [0, 0, 0, 0, 0, 0, 1, 1, 2, 2]
Y = label_binarize(y, classes=range(3))
cece = conf_ECE(Y, S)
self.assertAlmostEqual(cece, 0)
# TODO Add more tests
def test_classwise_ece(self):
S = np.array([[0.6, 0.2, 0.2]]*10)
Y = label_binarize([0, 0, 0, 0, 0, 0, 1, 1, 2, 2], classes=range(3))
ece = classwise_ECE(Y, S)
self.assertAlmostEqual(ece, 0)
# TODO Add more tests
def test_full_ece(self):
S = np.array([[0.6, 0.2, 0.2]]*10)
Y = label_binarize([0, 0, 0, 0, 0, 0, 1, 1, 2, 2], classes=range(3))
ece = full_ECE(Y, S)
self.assertAlmostEqual(ece, 0)
# TODO Add more tests
def test_conf_mce(self):
S = np.ones((2, 3))/3.0
y = np.array([0, 0])
mce = MCE(y, S)
self.assertAlmostEqual(mce, 2.0/3)
y = np.array([0, 1, 2])
S = np.array([[1/3, 0.3, 0.3],
[1/3, 0.3, 0.3],
[1/3, 0.3, 0.3]])
mce = MCE(y, S)
self.assertAlmostEqual(mce, 0.0)
y = np.array([0, 1, 2])
S = np.array([[0.3, 1/3, 0.3],
[0.3, 1/3, 0.3],
[0.3, 1/3, 0.3]])
mce = MCE(y, S)
self.assertAlmostEqual(mce, 0.0)
y = np.array([0, 1, 2])
S = np.array([[0.3, 0.3, 1/3],
[0.3, 0.3, 1/3],
[0.3, 0.3, 1/3]])
mce = MCE(y, S)
self.assertAlmostEqual(mce, 0.0)
Y = np.array([[1, 0, 0], [0, 1, 0], [0, 0, 1]])
S = np.array([[0.3, 0.3, 1/3],
[0.3, 0.3, 1/3],
[0.3, 0.3, 1/3]])
mce = MCE(Y, S)
self.assertAlmostEqual(mce, 0.0)
Y = np.array([[1, 0, 0], [1, 0, 0], [1, 0, 0], [0, 1, 0],
[1, 0, 0], [1, 0, 0], [1, 0, 0], [0, 1, 0]])
S = np.array([[0.4, 0.3, 0.3], # correct
[0.3, 0.4, 0.3], # incorrect
[0.3, 0.3, 0.4], # incorrect
[0.3, 0.3, 0.4], # incorrect
[0.1, 0.7, 0.2], # incorrect mean conf 0.75
[0.2, 0.1, 0.7], # incorrect
[0.2, 0.8, 0.2], # incorrect
[0.8, 0.1, 0.1] # incorrect
])
mce = MCE(Y, S, bins=2)
self.assertEqual(mce, 0.75)
Y = np.array([[1, 0, 0], [1, 0, 0], [1, 0, 0], [0, 1, 0],
[1, 0, 0], [1, 0, 0], [1, 0, 0], [0, 1, 0]])
S = np.array([[0.4, 0.3, 0.3], # correct # conf 0.4
[0.3, 0.4, 0.3], # incorrect
[0.3, 0.3, 0.4], # incorrect
[0.3, 0.3, 0.4], # incorrect
[0.1, 0.7, 0.2], # incorrect
[0.7, 0.1, 0.2], # correct
[0.8, 0.0, 0.2], # correct
[0.1, 0.8, 0.1] # correct
])
mce = MCE(Y, S, bins=2)
self.assertAlmostEqual(mce, 0.4 - 1/4)
def test_calibrated_p_ece(self):
p = np.random.rand(5000, 3)
p /= p.sum(axis=1)[:, None]
multinomial = partial(np.random.multinomial, 1)
y = np.apply_along_axis(multinomial, 1, p)
calibrated_pECE = pECE(y, p, samples=2000, ece_function=classwise_ECE)
# FIXME Reduce computation and increase threshold to 0.04
self.assertGreater(calibrated_pECE, 0.02)
calibrated_pECE = pECE(y, p, samples=2000, ece_function=conf_ECE)
# FIXME Reduce computation and increase threshold to 0.04
self.assertGreater(calibrated_pECE, 0.02)
def test_uncalibrated_p_ece(self):
p = np.random.rand(1000, 3)
p /= p.sum(axis=1)[:, None]
y = np.eye(3)[np.random.choice([0, 1, 2], size=p.shape[0])]
uncalibrated_pECE = pECE(y, p, samples=1000,
ece_function=classwise_ECE)
self.assertLess(uncalibrated_pECE, 0.04)
uncalibrated_pECE = pECE(y, p, samples=1000, ece_function=conf_ECE)
self.assertLess(uncalibrated_pECE, 0.04)
def main():
unittest.main()
if __name__ == '__main__':
main()