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
|
"""
Testing for the gradient boosting loss functions and initial estimators.
"""
import numpy as np
from numpy.testing import assert_array_equal
from numpy.testing import assert_almost_equal
from numpy.testing import assert_equal
from sklearn.utils import check_random_state
from sklearn.utils.testing import assert_raises
from sklearn.ensemble.gradient_boosting import BinomialDeviance
from sklearn.ensemble.gradient_boosting import LogOddsEstimator
from sklearn.ensemble.gradient_boosting import LeastSquaresError
from sklearn.ensemble.gradient_boosting import RegressionLossFunction
from sklearn.ensemble.gradient_boosting import LOSS_FUNCTIONS
from sklearn.ensemble.gradient_boosting import _weighted_percentile
from sklearn.ensemble.gradient_boosting import QuantileLossFunction
def test_binomial_deviance():
# Check binomial deviance loss.
# Check against alternative definitions in ESLII.
bd = BinomialDeviance(2)
# pred has the same BD for y in {0, 1}
assert_equal(bd(np.array([0.0]), np.array([0.0])),
bd(np.array([1.0]), np.array([0.0])))
assert_almost_equal(bd(np.array([1.0, 1.0, 1.0]),
np.array([100.0, 100.0, 100.0])),
0.0)
assert_almost_equal(bd(np.array([1.0, 0.0, 0.0]),
np.array([100.0, -100.0, -100.0])), 0)
# check if same results as alternative definition of deviance (from ESLII)
alt_dev = lambda y, pred: np.mean(np.logaddexp(0.0, -2.0 *
(2.0 * y - 1) * pred))
test_data = [(np.array([1.0, 1.0, 1.0]), np.array([100.0, 100.0, 100.0])),
(np.array([0.0, 0.0, 0.0]), np.array([100.0, 100.0, 100.0])),
(np.array([0.0, 0.0, 0.0]),
np.array([-100.0, -100.0, -100.0])),
(np.array([1.0, 1.0, 1.0]),
np.array([-100.0, -100.0, -100.0]))]
for datum in test_data:
assert_almost_equal(bd(*datum), alt_dev(*datum))
# check the gradient against the
alt_ng = lambda y, pred: (2 * y - 1) / (1 + np.exp(2 * (2 * y - 1) * pred))
for datum in test_data:
assert_almost_equal(bd.negative_gradient(*datum), alt_ng(*datum))
def test_log_odds_estimator():
# Check log odds estimator.
est = LogOddsEstimator()
assert_raises(ValueError, est.fit, None, np.array([1]))
est.fit(None, np.array([1.0, 0.0]))
assert_equal(est.prior, 0.0)
assert_array_equal(est.predict(np.array([[1.0], [1.0]])),
np.array([[0.0], [0.0]]))
def test_sample_weight_smoke():
rng = check_random_state(13)
y = rng.rand(100)
pred = rng.rand(100)
# least squares
loss = LeastSquaresError(1)
loss_wo_sw = loss(y, pred)
loss_w_sw = loss(y, pred, np.ones(pred.shape[0], dtype=np.float32))
assert_almost_equal(loss_wo_sw, loss_w_sw)
def test_sample_weight_init_estimators():
# Smoke test for init estimators with sample weights.
rng = check_random_state(13)
X = rng.rand(100, 2)
sample_weight = np.ones(100)
reg_y = rng.rand(100)
clf_y = rng.randint(0, 2, size=100)
for Loss in LOSS_FUNCTIONS.values():
if Loss is None:
continue
if issubclass(Loss, RegressionLossFunction):
k = 1
y = reg_y
else:
k = 2
y = clf_y
if Loss.is_multi_class:
# skip multiclass
continue
loss = Loss(k)
init_est = loss.init_estimator()
init_est.fit(X, y)
out = init_est.predict(X)
assert_equal(out.shape, (y.shape[0], 1))
sw_init_est = loss.init_estimator()
sw_init_est.fit(X, y, sample_weight=sample_weight)
sw_out = init_est.predict(X)
assert_equal(sw_out.shape, (y.shape[0], 1))
# check if predictions match
assert_array_equal(out, sw_out)
def test_weighted_percentile():
y = np.empty(102, dtype=np.float64)
y[:50] = 0
y[-51:] = 2
y[-1] = 100000
y[50] = 1
sw = np.ones(102, dtype=np.float64)
sw[-1] = 0.0
score = _weighted_percentile(y, sw, 50)
assert score == 1
def test_weighted_percentile_equal():
y = np.empty(102, dtype=np.float64)
y.fill(0.0)
sw = np.ones(102, dtype=np.float64)
sw[-1] = 0.0
score = _weighted_percentile(y, sw, 50)
assert score == 0
def test_weighted_percentile_zero_weight():
y = np.empty(102, dtype=np.float64)
y.fill(1.0)
sw = np.ones(102, dtype=np.float64)
sw.fill(0.0)
score = _weighted_percentile(y, sw, 50)
assert score == 1.0
def test_quantile_loss_function():
# Non regression test for the QuantileLossFunction object
# There was a sign problem when evaluating the function
# for negative values of 'ytrue - ypred'
x = np.asarray([-1.0, 0.0, 1.0])
y_found = QuantileLossFunction(1, 0.9)(x, np.zeros_like(x))
y_expected = np.asarray([0.1, 0.0, 0.9]).mean()
np.testing.assert_allclose(y_found, y_expected)
def test_sample_weight_deviance():
# Test if deviance supports sample weights.
rng = check_random_state(13)
X = rng.rand(100, 2)
sample_weight = np.ones(100)
reg_y = rng.rand(100)
clf_y = rng.randint(0, 2, size=100)
mclf_y = rng.randint(0, 3, size=100)
for Loss in LOSS_FUNCTIONS.values():
if Loss is None:
continue
if issubclass(Loss, RegressionLossFunction):
k = 1
y = reg_y
p = reg_y
else:
k = 2
y = clf_y
p = clf_y
if Loss.is_multi_class:
k = 3
y = mclf_y
# one-hot encoding
p = np.zeros((y.shape[0], k), dtype=np.float64)
for i in range(k):
p[:, i] = y == i
loss = Loss(k)
deviance_w_w = loss(y, p, sample_weight)
deviance_wo_w = loss(y, p)
assert deviance_wo_w == deviance_w_w
|
