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
|
import numpy as np
from scipy.sparse import csr_matrix
from sklearn.utils.testing import assert_array_equal, assert_equal, assert_true
from sklearn.utils.testing import assert_not_equal
from sklearn.utils.testing import assert_array_almost_equal, assert_raises
from sklearn.utils.testing import assert_less_equal
from sklearn.metrics.pairwise import kernel_metrics
from sklearn.kernel_approximation import RBFSampler
from sklearn.kernel_approximation import AdditiveChi2Sampler
from sklearn.kernel_approximation import SkewedChi2Sampler
from sklearn.kernel_approximation import Nystroem
from sklearn.metrics.pairwise import polynomial_kernel, rbf_kernel
# generate data
rng = np.random.RandomState(0)
X = rng.random_sample(size=(300, 50))
Y = rng.random_sample(size=(300, 50))
X /= X.sum(axis=1)[:, np.newaxis]
Y /= Y.sum(axis=1)[:, np.newaxis]
def test_additive_chi2_sampler():
# test that AdditiveChi2Sampler approximates kernel on random data
# compute exact kernel
# abbreviations for easier formula
X_ = X[:, np.newaxis, :]
Y_ = Y[np.newaxis, :, :]
large_kernel = 2 * X_ * Y_ / (X_ + Y_)
# reduce to n_samples_x x n_samples_y by summing over features
kernel = (large_kernel.sum(axis=2))
# approximate kernel mapping
transform = AdditiveChi2Sampler(sample_steps=3)
X_trans = transform.fit_transform(X)
Y_trans = transform.transform(Y)
kernel_approx = np.dot(X_trans, Y_trans.T)
assert_array_almost_equal(kernel, kernel_approx, 1)
X_sp_trans = transform.fit_transform(csr_matrix(X))
Y_sp_trans = transform.transform(csr_matrix(Y))
assert_array_equal(X_trans, X_sp_trans.A)
assert_array_equal(Y_trans, Y_sp_trans.A)
# test error is raised on negative input
Y_neg = Y.copy()
Y_neg[0, 0] = -1
assert_raises(ValueError, transform.transform, Y_neg)
# test error on invalid sample_steps
transform = AdditiveChi2Sampler(sample_steps=4)
assert_raises(ValueError, transform.fit, X)
# test that the sample interval is set correctly
sample_steps_available = [1, 2, 3]
for sample_steps in sample_steps_available:
# test that the sample_interval is initialized correctly
transform = AdditiveChi2Sampler(sample_steps=sample_steps)
assert_equal(transform.sample_interval, None)
# test that the sample_interval is changed in the fit method
transform.fit(X)
assert_not_equal(transform.sample_interval_, None)
# test that the sample_interval is set correctly
sample_interval = 0.3
transform = AdditiveChi2Sampler(sample_steps=4,
sample_interval=sample_interval)
assert_equal(transform.sample_interval, sample_interval)
transform.fit(X)
assert_equal(transform.sample_interval_, sample_interval)
def test_skewed_chi2_sampler():
# test that RBFSampler approximates kernel on random data
# compute exact kernel
c = 0.03
# abbreviations for easier formula
X_c = (X + c)[:, np.newaxis, :]
Y_c = (Y + c)[np.newaxis, :, :]
# we do it in log-space in the hope that it's more stable
# this array is n_samples_x x n_samples_y big x n_features
log_kernel = ((np.log(X_c) / 2.) + (np.log(Y_c) / 2.) + np.log(2.) -
np.log(X_c + Y_c))
# reduce to n_samples_x x n_samples_y by summing over features in log-space
kernel = np.exp(log_kernel.sum(axis=2))
# approximate kernel mapping
transform = SkewedChi2Sampler(skewedness=c, n_components=1000,
random_state=42)
X_trans = transform.fit_transform(X)
Y_trans = transform.transform(Y)
kernel_approx = np.dot(X_trans, Y_trans.T)
assert_array_almost_equal(kernel, kernel_approx, 1)
# test error is raised on negative input
Y_neg = Y.copy()
Y_neg[0, 0] = -1
assert_raises(ValueError, transform.transform, Y_neg)
def test_rbf_sampler():
# test that RBFSampler approximates kernel on random data
# compute exact kernel
gamma = 10.
kernel = rbf_kernel(X, Y, gamma=gamma)
# approximate kernel mapping
rbf_transform = RBFSampler(gamma=gamma, n_components=1000, random_state=42)
X_trans = rbf_transform.fit_transform(X)
Y_trans = rbf_transform.transform(Y)
kernel_approx = np.dot(X_trans, Y_trans.T)
error = kernel - kernel_approx
assert_less_equal(np.abs(np.mean(error)), 0.01) # close to unbiased
np.abs(error, out=error)
assert_less_equal(np.max(error), 0.1) # nothing too far off
assert_less_equal(np.mean(error), 0.05) # mean is fairly close
def test_input_validation():
# Regression test: kernel approx. transformers should work on lists
# No assertions; the old versions would simply crash
X = [[1, 2], [3, 4], [5, 6]]
AdditiveChi2Sampler().fit(X).transform(X)
SkewedChi2Sampler().fit(X).transform(X)
RBFSampler().fit(X).transform(X)
X = csr_matrix(X)
RBFSampler().fit(X).transform(X)
def test_nystroem_approximation():
# some basic tests
rnd = np.random.RandomState(0)
X = rnd.uniform(size=(10, 4))
# With n_components = n_samples this is exact
X_transformed = Nystroem(n_components=X.shape[0]).fit_transform(X)
K = rbf_kernel(X)
assert_array_almost_equal(np.dot(X_transformed, X_transformed.T), K)
trans = Nystroem(n_components=2, random_state=rnd)
X_transformed = trans.fit(X).transform(X)
assert_equal(X_transformed.shape, (X.shape[0], 2))
# test callable kernel
linear_kernel = lambda X, Y: np.dot(X, Y.T)
trans = Nystroem(n_components=2, kernel=linear_kernel, random_state=rnd)
X_transformed = trans.fit(X).transform(X)
assert_equal(X_transformed.shape, (X.shape[0], 2))
# test that available kernels fit and transform
kernels_available = kernel_metrics()
for kern in kernels_available:
trans = Nystroem(n_components=2, kernel=kern, random_state=rnd)
X_transformed = trans.fit(X).transform(X)
assert_equal(X_transformed.shape, (X.shape[0], 2))
def test_nystroem_singular_kernel():
# test that nystroem works with singular kernel matrix
rng = np.random.RandomState(0)
X = rng.rand(10, 20)
X = np.vstack([X] * 2) # duplicate samples
gamma = 100
N = Nystroem(gamma=gamma, n_components=X.shape[0]).fit(X)
X_transformed = N.transform(X)
K = rbf_kernel(X, gamma=gamma)
assert_array_almost_equal(K, np.dot(X_transformed, X_transformed.T))
assert_true(np.all(np.isfinite(Y)))
def test_nystroem_poly_kernel_params():
# Non-regression: Nystroem should pass other parameters beside gamma.
rnd = np.random.RandomState(37)
X = rnd.uniform(size=(10, 4))
K = polynomial_kernel(X, degree=3.1, coef0=.1)
nystroem = Nystroem(kernel="polynomial", n_components=X.shape[0],
degree=3.1, coef0=.1)
X_transformed = nystroem.fit_transform(X)
assert_array_almost_equal(np.dot(X_transformed, X_transformed.T), K)
def test_nystroem_callable():
# Test Nystroem on a callable.
rnd = np.random.RandomState(42)
n_samples = 10
X = rnd.uniform(size=(n_samples, 4))
def logging_histogram_kernel(x, y, log):
"""Histogram kernel that writes to a log."""
log.append(1)
return np.minimum(x, y).sum()
kernel_log = []
X = list(X) # test input validation
Nystroem(kernel=logging_histogram_kernel,
n_components=(n_samples - 1),
kernel_params={'log': kernel_log}).fit(X)
assert_equal(len(kernel_log), n_samples * (n_samples - 1) / 2)
|
