|
1
|
|
|
"""
|
|
2
|
|
|
Utility functions necessary for different classifiers.
|
|
3
|
|
|
|
|
4
|
|
|
Contains algebraic operations, and label encodings.
|
|
5
|
|
|
"""
|
|
6
|
|
|
|
|
7
|
1 |
|
import numpy as np
|
|
8
|
1 |
|
import numpy.linalg as al
|
|
9
|
1 |
|
import scipy.stats as st
|
|
10
|
|
|
|
|
11
|
|
|
|
|
12
|
1 |
|
def one_hot(y, fill_k=False, one_not=False):
|
|
13
|
|
|
"""Map to one-hot encoding."""
|
|
14
|
|
|
# Check for negative labels
|
|
15
|
|
|
assert np.all(y >= 0)
|
|
16
|
|
|
|
|
17
|
|
|
# Number of samples
|
|
18
|
|
|
N = y.shape[0]
|
|
19
|
|
|
|
|
20
|
|
|
# Number of classes
|
|
21
|
|
|
K = len(np.unique(y))
|
|
22
|
|
|
|
|
23
|
|
|
# Preallocate array
|
|
24
|
|
|
if one_not:
|
|
25
|
|
|
Y = -np.ones((N, K))
|
|
26
|
|
|
else:
|
|
27
|
|
|
Y = np.zeros((N, K))
|
|
28
|
|
|
|
|
29
|
|
|
# Set k-th column to 1 for n-th sample
|
|
30
|
|
|
for n in range(N):
|
|
31
|
|
|
|
|
32
|
|
|
if fill_k:
|
|
33
|
|
|
Y[n, y[n]] = y[n]
|
|
34
|
|
|
else:
|
|
35
|
|
|
Y[n, y[n]] = 1
|
|
36
|
|
|
|
|
37
|
|
|
return Y
|
|
38
|
|
|
|
|
39
|
|
|
|
|
40
|
1 |
|
def regularize_matrix(A, a=0.0):
|
|
41
|
|
|
"""
|
|
42
|
|
|
Regularize matrix by ensuring minimum eigenvalues.
|
|
43
|
|
|
|
|
44
|
|
|
INPUT (1) array 'A': square matrix
|
|
45
|
|
|
(2) float 'a': constraint on minimum eigenvalue
|
|
46
|
|
|
OUTPUT (1) array 'B': constrained matrix
|
|
47
|
|
|
"""
|
|
48
|
|
|
# Check for square matrix
|
|
49
|
|
|
N, M = A.shape
|
|
50
|
|
|
assert N == M
|
|
51
|
|
|
|
|
52
|
|
|
# Check for valid matrix entries
|
|
53
|
|
|
assert not np.any(np.isnan(A)) or np.any(np.isinf(A))
|
|
54
|
|
|
|
|
55
|
|
|
# Check for non-negative minimum eigenvalue
|
|
56
|
|
|
if a < 0:
|
|
57
|
|
|
raise ValueError('minimum eigenvalue cannot be negative.')
|
|
58
|
|
|
|
|
59
|
|
|
elif a == 0:
|
|
60
|
|
|
return A
|
|
61
|
|
|
|
|
62
|
|
|
else:
|
|
63
|
|
|
# Ensure symmetric matrix
|
|
64
|
|
|
A = (A + A.T) / 2
|
|
65
|
|
|
|
|
66
|
|
|
# Eigenvalue decomposition
|
|
67
|
|
|
E, V = al.eig(A)
|
|
68
|
|
|
|
|
69
|
|
|
# Regularization matrix
|
|
70
|
|
|
aI = a * np.eye(N)
|
|
71
|
|
|
|
|
72
|
|
|
# Subtract regularization
|
|
73
|
|
|
E = np.diag(E) + aI
|
|
74
|
|
|
|
|
75
|
|
|
# Cap negative eigenvalues at zero
|
|
76
|
|
|
E = np.maximum(0, E)
|
|
77
|
|
|
|
|
78
|
|
|
# Reconstruct matrix
|
|
79
|
|
|
B = np.dot(np.dot(V, E), V.T)
|
|
80
|
|
|
|
|
81
|
|
|
# Add back subtracted regularization
|
|
82
|
|
|
return B + aI
|
|
83
|
|
|
|
|
84
|
|
|
|
|
85
|
1 |
|
def is_pos_def(X):
|
|
86
|
|
|
"""Check for positive definiteness."""
|
|
87
|
1 |
|
return np.all(np.linalg.eigvals(X) > 0)
|
|
88
|
|
|
|
|
89
|
|
|
|
|
90
|
1 |
|
def nullspace(A, atol=1e-13, rtol=0):
|
|
91
|
|
|
"""
|
|
92
|
|
|
Compute an approximate basis for the nullspace of A.
|
|
93
|
|
|
|
|
94
|
|
|
INPUT (1) array 'A': 1-D array with length k will be treated
|
|
95
|
|
|
as a 2-D with shape (1, k).
|
|
96
|
|
|
(2) float 'atol': the absolute tolerance for a zero singular value.
|
|
97
|
|
|
Singular values smaller than `atol` are considered to be zero.
|
|
98
|
|
|
(3) float 'rtol': relative tolerance. Singular values less than
|
|
99
|
|
|
rtol*smax are considered to be zero, where smax is the largest
|
|
100
|
|
|
singular value.
|
|
101
|
|
|
|
|
102
|
|
|
If both `atol` and `rtol` are positive, the combined tolerance
|
|
103
|
|
|
is the maximum of the two; tol = max(atol, rtol * smax)
|
|
104
|
|
|
Singular values smaller than `tol` are considered to be zero.
|
|
105
|
|
|
OUTPUT (1) array 'B': if A is an array with shape (m, k), then B will be
|
|
106
|
|
|
an array with shape (k, n), where n is the estimated dimension
|
|
107
|
|
|
of the nullspace of A. The columns of B are a basis for the
|
|
108
|
|
|
nullspace; each element in np.dot(A, B) will be
|
|
109
|
|
|
approximately zero.
|
|
110
|
|
|
"""
|
|
111
|
|
|
# Expand A to a matrix
|
|
112
|
|
|
A = np.atleast_2d(A)
|
|
113
|
|
|
|
|
114
|
|
|
# Singular value decomposition
|
|
115
|
|
|
u, s, vh = al.svd(A)
|
|
116
|
|
|
|
|
117
|
|
|
# Set tolerance
|
|
118
|
|
|
tol = max(atol, rtol * s[0])
|
|
119
|
|
|
|
|
120
|
|
|
# Compute the number of non-zero entries
|
|
121
|
|
|
nnz = (s >= tol).sum()
|
|
122
|
|
|
|
|
123
|
|
|
# Conjugate and transpose to ensure real numbers
|
|
124
|
|
|
ns = vh[nnz:].conj().T
|
|
125
|
|
|
|
|
126
|
|
|
return ns
|
|
127
|
|
|
|
wmkouw /
libTLDA
