-
Notifications
You must be signed in to change notification settings - Fork 13
/
Copy pathJPDA_compare_python.py
219 lines (191 loc) · 8.53 KB
/
JPDA_compare_python.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
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
217
218
219
# -*- coding: utf-8 -*-
# @Time : 2021/7/10 16:01
# @Author : Wen Zhang
# @File : JPDA_compare_python.py
# Reference: https://github.com/jindongwang/transferlearning/blob/master/code/traditional/JDA
import numpy as np
import scipy.io
import scipy.linalg
from sklearn import preprocessing
from sklearn.metrics import accuracy_score
from sklearn.metrics.pairwise import linear_kernel, rbf_kernel
from sklearn.neighbors import KNeighborsClassifier
from sklearn.preprocessing import OneHotEncoder
def kernel(ker, X1, X2, gamma):
K = None
if not ker or ker == 'primal':
K = X1
elif ker == 'linear':
if X2:
K = linear_kernel(np.asarray(X1).T, np.asarray(X2).T)
else:
K = linear_kernel(np.asarray(X1).T)
elif ker == 'rbf':
if X2:
K = rbf_kernel(np.asarray(X1).T, np.asarray(X2).T, gamma)
else:
K = rbf_kernel(np.asarray(X1).T, None, gamma)
return K
def get_matrix_M(Ys, Y_tar_pseudo, ns, nt, C, mu, type='djp-mmd'):
M = 0
if type == 'jmmd':
N = 0
n = ns + nt
e = np.vstack((1 / ns * np.ones((ns, 1)), -1 / nt * np.ones((nt, 1))))
M0 = e * e.T * C
if Y_tar_pseudo is not None and len(Y_tar_pseudo) == nt:
for c in range(1, C + 1):
e = np.zeros((n, 1))
tt = Ys == c
e[np.where(tt == True)] = 1 / len(Ys[np.where(Ys == c)])
yy = Y_tar_pseudo == c
ind = np.where(yy == True)
inds = [item + ns for item in ind]
e[tuple(inds)] = -1 / len(Y_tar_pseudo[np.where(Y_tar_pseudo == c)])
e[np.isinf(e)] = 0
N = N + np.dot(e, e.T)
M = M0 + N
M = M / np.linalg.norm(M, 'fro')
if type == 'jp-mmd':
ohe = OneHotEncoder()
ohe.fit(np.unique(Ys).reshape(-1, 1))
Ys_ohe = ohe.transform(Ys.reshape(-1, 1)).toarray().astype(np.int8)
# For transferability
Ns = 1 / ns * Ys_ohe
Nt = np.zeros([nt, C])
if Y_tar_pseudo is not None:
Yt_ohe = ohe.transform(Y_tar_pseudo.reshape(-1, 1)).toarray().astype(np.int8)
Nt = 1 / nt * Yt_ohe
Rmin = np.r_[np.c_[np.dot(Ns, Ns.T), np.dot(-Ns, Nt.T)], np.c_[np.dot(-Nt, Ns.T), np.dot(Nt, Nt.T)]]
M = Rmin / np.linalg.norm(Rmin, 'fro')
if type == 'djp-mmd':
ohe = OneHotEncoder()
ohe.fit(np.unique(Ys).reshape(-1, 1))
Ys_ohe = ohe.transform(Ys.reshape(-1, 1)).toarray().astype(np.int8)
# For transferability
Ns = 1 / ns * Ys_ohe
Nt = np.zeros([nt, C])
if Y_tar_pseudo is not None:
Yt_ohe = ohe.transform(Y_tar_pseudo.reshape(-1, 1)).toarray().astype(np.int8)
Nt = 1 / nt * Yt_ohe
Rmin = np.r_[np.c_[np.dot(Ns, Ns.T), np.dot(-Ns, Nt.T)], np.c_[np.dot(-Nt, Ns.T), np.dot(Nt, Nt.T)]]
Rmin = Rmin / np.linalg.norm(Rmin, 'fro')
# For discriminability
Ms = np.zeros([ns, (C - 1) * C])
Mt = np.zeros([nt, (C - 1) * C])
for i in range(C):
idx = np.arange((C - 1) * i, (C - 1) * (i + 1))
Ms[:, idx] = np.tile(Ns[:, i], (C - 1, 1)).T
tmp = np.arange(C)
Mt[:, idx] = Nt[:, tmp[tmp != i]]
Rmax = np.r_[np.c_[np.dot(Ms, Ms.T), np.dot(-Ms, Mt.T)], np.c_[np.dot(-Mt, Ms.T), np.dot(Mt, Mt.T)]]
Rmax = Rmax / np.linalg.norm(Rmax, 'fro')
M = Rmin - mu * Rmax
return M
class DA_statistics:
def __init__(self, kernel_type='primal', mmd_type='djp-mmd', dim=30, lamb=1, gamma=1, mu=0.1, T=5):
'''
Init func
:param kernel_type: kernel, values: 'primal' | 'linear' | 'rbf'
:param dim: dimension after transfer
:param lamb: lambda value in equation
:param gamma: kernel bandwidth for rbf kernel
:param T: iteration number
'''
self.kernel_type = kernel_type
self.mmd_type = mmd_type
self.dim = dim
self.lamb = lamb
self.gamma = gamma
self.mu = mu
self.T = T
def fit_predict(self, Xs, Ys, Xt, Yt):
'''
Transform and Predict using 1NN as JDA paper did
:param Xs: ns * n_feature, source feature
:param Ys: ns * 1, source label
:param Xt: nt * n_feature, target feature
:param Yt: nt * 1, target label
:return: acc, y_pred, list_acc
'''
X = np.hstack((Xs.T, Xt.T))
X = np.dot(X, np.diag(1. / np.linalg.norm(X, axis=0)))
m, n = X.shape # 800, 2081
ns, nt = len(Xs), len(Xt)
C = len(np.unique(Ys))
H = np.eye(n) - 1 / n * np.ones((n, n))
Y_tar_pseudo = None
list_acc = []
for itr in range(self.T):
M = get_matrix_M(Ys, Y_tar_pseudo, ns, nt, C, self.mu, type=self.mmd_type)
K = kernel(self.kernel_type, X, None, gamma=self.gamma)
n_eye = m if self.kernel_type == 'primal' else n
a, b = np.linalg.multi_dot([K, M, K.T]) + self.lamb * np.eye(n_eye), np.linalg.multi_dot([K, H, K.T])
w, V = scipy.linalg.eig(a, b)
ind = np.argsort(w)
A = V[:, ind[:self.dim]]
Z = np.dot(A.T, K)
Z /= np.linalg.norm(Z, axis=0)
Xs_new, Xt_new = Z[:, :ns].T, Z[:, ns:].T
clf = KNeighborsClassifier(n_neighbors=1)
clf.fit(Xs_new, Ys.ravel())
Y_tar_pseudo = clf.predict(Xt_new)
acc = accuracy_score(Yt, Y_tar_pseudo)
list_acc.append(acc)
print('iteration [{}/{}]: acc: {:.4f}'.format(itr + 1, self.T, acc))
return list_acc[-1], Y_tar_pseudo, list_acc
if __name__ == '__main__':
domains = ['caltech_SURF_L10.mat', 'amazon_SURF_L10.mat', 'webcam_SURF_L10.mat', 'dslr_SURF_L10.mat']
name_list = [name[0].upper() for name in domains]
mmd_list = ['jmmd', 'jp-mmd', 'djp-mmd']
num_domain = len(domains)
acc_all = np.zeros([len(name_list) * (len(name_list) - 1), len(mmd_list)])
itr_idx = 0
for s in range(num_domain): # source
for t in range(num_domain): # target
if s != t:
print('%s: %s --> %s' % (itr_idx, name_list[s], name_list[t]))
src, tar = 'data/Office/' + domains[s], 'data/Office/' + domains[t]
src_domain, tar_domain = scipy.io.loadmat(src), scipy.io.loadmat(tar)
Xs, Ys, Xt, Yt = src_domain['fts'], src_domain['labels'], tar_domain['fts'], tar_domain['labels']
# can only be added in offline learning, follow JDA original code
Xs = preprocessing.scale(Xs)
Xt = preprocessing.scale(Xt)
# linear kernel for office-caltech as the original JDA paper
ker_type = 'linear'
# # I: joint MMD
mmd_type = mmd_list[0]
traditional_tl = DA_statistics(kernel_type=ker_type, mmd_type=mmd_type, dim=100, lamb=1, gamma=1)
acc_all[itr_idx, 0], _, _ = traditional_tl.fit_predict(Xs, Ys, Xt, Yt)
print('type: {} -- acc: {:.4f}\n'.format(mmd_type, acc_all[itr_idx, 0]))
# # II: joint probability MMD
mmd_type = mmd_list[1]
traditional_tl = DA_statistics(kernel_type=ker_type, mmd_type=mmd_type, dim=100, lamb=1, gamma=1)
acc_all[itr_idx, 1], _, _ = traditional_tl.fit_predict(Xs, Ys, Xt, Yt)
print('type: {} -- acc: {:.4f}\n'.format(mmd_type, acc_all[itr_idx, 1]))
# # III: discriminative joint probability MMD
mmd_type = mmd_list[2]
traditional_tl = DA_statistics(kernel_type=ker_type, mmd_type=mmd_type, dim=100, lamb=1, gamma=1)
acc_all[itr_idx, 2], _, _ = traditional_tl.fit_predict(Xs, Ys, Xt, Yt)
print('type: {} -- acc: {:.4f}\n'.format(mmd_type, acc_all[itr_idx, 2]))
itr_idx += 1
print('mean acc...')
print(np.round(np.mean(acc_all, axis=0), 4))
print('\n', mmd_list)
print(np.round(acc_all, 4))
# mean acc...
# [0.4549 0.4634 0.4798]
#
# ['jmmd', 'jp-mmd', 'djp-mmd']
# [[0.4551 0.4499 0.4854]
# [0.4102 0.4203 0.4441]
# [0.3822 0.3822 0.4268]
# [0.4025 0.39 0.3954]
# [0.4034 0.3966 0.4441]
# [0.4204 0.4459 0.4459]
# [0.3161 0.3197 0.3179]
# [0.3173 0.2954 0.3163]
# [0.8917 0.9108 0.8981]
# [0.2965 0.3375 0.3571]
# [0.3257 0.3351 0.3486]
# [0.8373 0.878 0.878]]