|
1
|
|
|
#!/usr/bin/env python
|
|
2
|
|
|
# -*- coding: utf-8 -*-
|
|
3
|
|
|
|
|
4
|
1 |
|
import numpy as np
|
|
5
|
1 |
|
import scipy.stats as st
|
|
6
|
1 |
|
from scipy.sparse.linalg import eigs
|
|
7
|
1 |
|
from scipy.spatial.distance import cdist
|
|
8
|
1 |
|
import sklearn as sk
|
|
9
|
1 |
|
from sklearn.decomposition import PCA
|
|
10
|
1 |
|
from sklearn.svm import LinearSVC
|
|
11
|
1 |
|
from sklearn.linear_model import LogisticRegression, LinearRegression
|
|
12
|
1 |
|
from sklearn.model_selection import cross_val_predict
|
|
13
|
1 |
|
from os.path import basename
|
|
14
|
|
|
|
|
15
|
1 |
|
from .util import is_pos_def
|
|
16
|
|
|
|
|
17
|
|
|
|
|
18
|
1 |
|
class SubspaceAlignedClassifier(object):
|
|
19
|
|
|
"""
|
|
20
|
|
|
Class of classifiers based on Subspace Alignment.
|
|
21
|
|
|
|
|
22
|
|
|
Methods contain the alignment itself, classifiers and general utilities.
|
|
23
|
|
|
"""
|
|
24
|
|
|
|
|
25
|
1 |
|
def __init__(self, loss='logistic', l2=1.0, num_components=1):
|
|
26
|
|
|
"""
|
|
27
|
|
|
Select a particular type of subspace aligned classifier.
|
|
28
|
|
|
|
|
29
|
|
|
INPUT (1) str 'loss': loss function for weighted classifier, options:
|
|
30
|
|
|
'logistic', 'quadratic', 'hinge' (def: 'logistic')
|
|
31
|
|
|
(2) float 'l2': l2-regularization parameter value (def:0.01)
|
|
32
|
|
|
(3) int 'num_components': number of transfer components to
|
|
33
|
|
|
maintain (def: 1)
|
|
34
|
|
|
"""
|
|
35
|
1 |
|
self.loss = loss
|
|
36
|
1 |
|
self.l2 = l2
|
|
37
|
1 |
|
self.num_components = num_components
|
|
38
|
|
|
|
|
39
|
|
|
# Initialize untrained classifiers
|
|
40
|
1 |
|
if self.loss == 'logistic':
|
|
41
|
|
|
# Logistic regression model
|
|
42
|
1 |
|
self.clf = LogisticRegression()
|
|
43
|
|
|
elif self.loss == 'quadratic':
|
|
44
|
|
|
# Least-squares model
|
|
45
|
|
|
self.clf = LinearRegression()
|
|
46
|
|
|
elif self.loss == 'hinge':
|
|
47
|
|
|
# Linear support vector machine
|
|
48
|
|
|
self.clf = LinearSVC()
|
|
49
|
|
|
else:
|
|
50
|
|
|
# Other loss functions are not implemented
|
|
51
|
|
|
raise NotImplementedError
|
|
52
|
|
|
|
|
53
|
|
|
# Maintain target principal component coefficients
|
|
54
|
1 |
|
self.CZ = ''
|
|
55
|
|
|
|
|
56
|
|
|
# Whether model has been trained
|
|
57
|
1 |
|
self.is_trained = False
|
|
58
|
|
|
|
|
59
|
|
|
# Dimensionality of training data
|
|
60
|
1 |
|
self.train_data_dim = ''
|
|
61
|
|
|
|
|
62
|
1 |
|
def subspace_alignment(self, X, Z, num_components=1):
|
|
63
|
|
|
"""
|
|
64
|
|
|
Compute subspace and alignment matrix.
|
|
65
|
|
|
|
|
66
|
|
|
INPUT (1) array 'X': source data set (N samples by D features)
|
|
67
|
|
|
(2) array 'Z': target data set (M samples by D features)
|
|
68
|
|
|
(3) int 'num_components': number of components (def: 1)
|
|
69
|
|
|
OUTPUT (1) array 'V': transformation matrix (D features by D features)
|
|
70
|
|
|
(2) array 'CX': source principal component coefficients
|
|
71
|
|
|
(3) array 'CZ': target principal component coefficients
|
|
72
|
|
|
"""
|
|
73
|
|
|
# Data shapes
|
|
74
|
1 |
|
N, DX = X.shape
|
|
75
|
1 |
|
M, DZ = Z.shape
|
|
76
|
|
|
|
|
77
|
|
|
# Assert equivalent dimensionalities
|
|
78
|
1 |
|
assert DX == DZ
|
|
79
|
|
|
|
|
80
|
|
|
# Compute principal components
|
|
81
|
1 |
|
CX = PCA(n_components=num_components, whiten=True).fit(X).components_.T
|
|
82
|
1 |
|
CZ = PCA(n_components=num_components, whiten=True).fit(Z).components_.T
|
|
83
|
|
|
|
|
84
|
|
|
# Aligned source components
|
|
85
|
1 |
|
V = np.dot(CX.T, CZ)
|
|
86
|
|
|
|
|
87
|
|
|
# Return transformation matrix and principal component coefficients
|
|
88
|
1 |
|
return V, CX, CZ
|
|
89
|
|
|
|
|
90
|
1 |
View Code Duplication |
def fit(self, X, y, Z):
|
|
|
|
|
|
|
91
|
|
|
"""
|
|
92
|
|
|
Fit/train a classifier on data mapped onto transfer components.
|
|
93
|
|
|
|
|
94
|
|
|
INPUT (1) array 'X': source data (N samples by D features)
|
|
95
|
|
|
(2) array 'y': source labels (N samples by 1)
|
|
96
|
|
|
(3) array 'Z': target data (M samples by D features)
|
|
97
|
|
|
OUTPUT None
|
|
98
|
|
|
"""
|
|
99
|
|
|
# Data shapes
|
|
100
|
1 |
|
N, DX = X.shape
|
|
101
|
1 |
|
M, DZ = Z.shape
|
|
102
|
|
|
|
|
103
|
|
|
# Assert equivalent dimensionalities
|
|
104
|
1 |
|
assert DX == DZ
|
|
105
|
|
|
|
|
106
|
|
|
# Transfer component analysis (store target subspace)
|
|
107
|
1 |
|
V, CX, self.CZ = self.subspace_alignment(X, Z, num_components=self.
|
|
108
|
|
|
num_components)
|
|
109
|
|
|
|
|
110
|
|
|
# Map source data onto source principal components
|
|
111
|
1 |
|
X = np.dot(X, CX)
|
|
112
|
|
|
|
|
113
|
|
|
# Align source data to target subspace
|
|
114
|
1 |
|
X = np.dot(X, V)
|
|
115
|
|
|
|
|
116
|
|
|
# Train a weighted classifier
|
|
117
|
1 |
|
if self.loss == 'logistic':
|
|
118
|
|
|
# Logistic regression model with sample weights
|
|
119
|
1 |
|
self.clf.fit(X, y)
|
|
120
|
|
|
elif self.loss == 'quadratic':
|
|
121
|
|
|
# Least-squares model with sample weights
|
|
122
|
|
|
self.clf.fit(X, y)
|
|
123
|
|
|
elif self.loss == 'hinge':
|
|
124
|
|
|
# Linear support vector machine with sample weights
|
|
125
|
|
|
self.clf.fit(X, y)
|
|
126
|
|
|
else:
|
|
127
|
|
|
# Other loss functions are not implemented
|
|
128
|
|
|
raise NotImplementedError
|
|
129
|
|
|
|
|
130
|
|
|
# Mark classifier as trained
|
|
131
|
1 |
|
self.is_trained = True
|
|
132
|
|
|
|
|
133
|
|
|
# Store training data dimensionality
|
|
134
|
1 |
|
self.train_data_dim = DX
|
|
135
|
|
|
|
|
136
|
1 |
|
def predict(self, Z_, whiten=False):
|
|
137
|
|
|
"""
|
|
138
|
|
|
Make predictions on new dataset.
|
|
139
|
|
|
|
|
140
|
|
|
INPUT (1) array 'Z_': new data set (M samples by D features)
|
|
141
|
|
|
(2) boolean 'whiten': whether to whiten new data (def: false)
|
|
142
|
|
|
OUTPUT (1) array 'preds': label predictions (M samples by 1)
|
|
143
|
|
|
"""
|
|
144
|
|
|
# Data shape
|
|
145
|
1 |
|
M, D = Z_.shape
|
|
146
|
|
|
|
|
147
|
|
|
# If classifier is trained, check for same dimensionality
|
|
148
|
1 |
|
if self.is_trained:
|
|
149
|
1 |
|
assert self.train_data_dim == D
|
|
150
|
|
|
|
|
151
|
|
|
# Check for need to whiten data beforehand
|
|
152
|
1 |
|
if whiten:
|
|
153
|
|
|
Z_ = st.zscore(Z_)
|
|
154
|
|
|
|
|
155
|
|
|
# Map new target data onto target subspace
|
|
156
|
1 |
|
Z_ = np.dot(Z_, self.CZ)
|
|
157
|
|
|
|
|
158
|
|
|
# Call scikit's predict function
|
|
159
|
1 |
|
preds = self.clf.predict(Z_)
|
|
160
|
|
|
|
|
161
|
|
|
# For quadratic loss function, correct predictions
|
|
162
|
1 |
|
if self.loss == 'quadratic':
|
|
163
|
|
|
preds = (np.sign(preds)+1)/2.
|
|
164
|
|
|
|
|
165
|
|
|
# Return predictions array
|
|
166
|
1 |
|
return preds
|
|
167
|
|
|
|
|
168
|
1 |
|
def get_params(self):
|
|
169
|
|
|
"""Get classifier parameters."""
|
|
170
|
|
|
return self.clf.get_params()
|
|
171
|
|
|
|
|
172
|
1 |
|
def is_trained(self):
|
|
173
|
|
|
"""Check whether classifier is trained."""
|
|
174
|
|
|
return self.is_trained
|
|
175
|
|
|
|
wmkouw /
libTLDA
Loading history...