-
Notifications
You must be signed in to change notification settings - Fork 4
/
Copy pathtest_minimax_da_mnistm.py
213 lines (167 loc) · 7.98 KB
/
test_minimax_da_mnistm.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
## Minimax version of domain-adpative NN
# Common filter for both mnist and mnistm
## Jihun Hamm, 2017
# Some parts of the codes are from https://github.com/pumpikano/tf-dann.
import tensorflow as tf
import numpy as np
import cPickle as pkl
import matplotlib.pyplot as plt
import keras
#from keras.datasets import cifar10, cifar100
from keras.datasets import mnist
from utils import *
#########################################################################################################################
## Lower-level net (perturbation/filter)
def NN_filt(ins,scope='filt',reuse=False):
with tf.variable_scope(scope,reuse=reuse):
W1 = tf.get_variable('W1',[5,5,3,32],initializer=tf.random_normal_initializer(stddev=0.1))
b1 = tf.get_variable('b1',[32],initializer=tf.constant_initializer(0.0))
c1 = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(ins, W1, strides=[1,1,1,1], padding='SAME'),b1))
p1 = tf.nn.max_pool(c1, ksize=[1,2,2,1],strides=[1,2,2,1], padding='SAME')
W2 = tf.get_variable('W2',[5,5,32,48],initializer=tf.random_normal_initializer(stddev=0.1))
b2 = tf.get_variable('b2',[48],initializer=tf.constant_initializer(0.0))
c2 = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(p1, W2, strides=[1,1,1,1], padding='SAME'), b2))
p2 = tf.nn.max_pool(c2, ksize=[1,2,2,1],strides=[1,2,2,1], padding='SAME')
out = tf.reshape(p2,[-1,7*7*48])
reg = tf.nn.l2_loss(W1) + tf.nn.l2_loss(W2)
return [out,reg]
## Upper-level net
## Utility classifier
def NN_util(ins,scope='util',reuse=False):
with tf.variable_scope(scope,reuse=reuse):
W1 = tf.get_variable('W1',[7*7*48,nh],initializer=tf.random_normal_initializer(stddev=0.1))
b1 = tf.get_variable('b1',[nh],initializer=tf.constant_initializer(0.0))
a1 = tf.nn.relu(tf.nn.bias_add(tf.matmul(ins,W1),b1))
W2 = tf.get_variable('W2',[nh,nh],initializer=tf.random_normal_initializer(stddev=0.1))
b2 = tf.get_variable('b2',[nh],initializer=tf.constant_initializer(0.0))
a2 = tf.nn.relu(tf.nn.bias_add(tf.matmul(a1,W2),b2))
W3 = tf.get_variable('W3',[nh,Ku],initializer=tf.random_normal_initializer(stddev=0.1))
b3 = tf.get_variable('b3',[Ku],initializer=tf.constant_initializer(0.0))
out = tf.nn.bias_add(tf.matmul(a2,W3),b3)
reg = tf.nn.l2_loss(W1) + tf.nn.l2_loss(W2) + tf.nn.l2_loss(W3)
return [out,reg]
## Privacy classifier
def NN_priv(ins,scope='priv',reuse=False):
with tf.variable_scope(scope,reuse=reuse):
W1 = tf.get_variable('W1',[7*7*48,nh],initializer=tf.random_normal_initializer(stddev=0.1))
b1 = tf.get_variable('b1',[nh],initializer=tf.constant_initializer(0.0))
a1 = tf.nn.relu(tf.nn.bias_add(tf.matmul(ins,W1),b1))
W2 = tf.get_variable('W2',[nh,Kp],initializer=tf.random_normal_initializer(stddev=0.1))
b2 = tf.get_variable('b2',[Kp],initializer=tf.constant_initializer(0.0))
out = tf.nn.bias_add(tf.matmul(a1,W2),b2)
reg = tf.nn.l2_loss(W1) + tf.nn.l2_loss(W2)
return [out,reg]
## minimax optimization
def minimax_kbeam(sess,feed_dict={}):
# min step
for it_min in range(min_step):
fs = sess.run(loss,feed_dict)
id_max = np.argmax(fs)
sess.run(optim_min[id_max],feed_dict)
# max step
for it_max in range(max_step):
sess.run(optim_max,feed_dict)
def evaluate(X1test,X2test,y1test,y2test):
batchsize = 1280
n = X1test.shape[0]
nbatch = np.int(np.ceil(np.float(n)/np.float(batchsize)))
err_test_src = 0.
err_test_tar = 0.
for i in range(nbatch):
ind = range(batchsize*i,min(batchsize*(1+i),n))
err_test_src += sess.run(sumu,feed_dict={x:X1test[ind],yutil:y1test[ind], batch_size:batchsize})
err_test_tar += sess.run(sumu,feed_dict={x:X2test[ind],yutil:y2test[ind], batch_size:batchsize})
err_test_src /= np.float(n)
err_test_tar /= np.float(n)
return [err_test_src,err_test_tar]
#########################################################################################################################
K = 2
max_step = 1
min_step = 1
ntrial = 10
max_iter = 10001
nskip = 100
Ku = 10
Kp = 2
nh = 100
rho = 1E0 # Utility accuracy drops too much for rho smaller than 1
lamb = 1E-12
batchsize = 128
lr = 1E-2
## Load data
(X1train, y1train), (X1test, y1test) = mnist.load_data()
X1train = np.tile(X1train.reshape((-1,28,28,1)),(1,1,1,3))
X1test = np.tile(X1test.reshape((-1,28,28,1)),(1,1,1,3))
y1train = keras.utils.to_categorical(y1train,10)
y1test = keras.utils.to_categorical(y1test,10)
## Load MNIST-M
mnistm = pkl.load(open('mnistm_data.pkl'))
X2train = mnistm['train']
X2test = mnistm['test']
X2valid = mnistm['valid']
## Create the model
x = tf.placeholder(tf.uint8, [None,28,28,3],'x') # Source domain
z = tf.placeholder(tf.uint8, [None,28,28,3],'z') # Target domain
yutil = tf.placeholder(tf.float32, [None, Ku],'yutil')
batch_size = tf.placeholder(tf.int32,[],'batch_size')
xfloat = tf.multiply(tf.cast(x, tf.float32),0.0039215686)
zfloat = tf.multiply(tf.cast(z, tf.float32),0.0039215686)
## Connect networks
filtx,reg_filt = NN_filt(xfloat)
filtz,_ = NN_filt(zfloat,reuse=True)
futilx,reg_util = NN_util(filtx)
futilz,_ = NN_util(filtz,reuse=True)
## y=0: real data, y=1: fake data
yneg = tf.concat([tf.ones([batch_size,1]),tf.zeros([batch_size,1])],1)
ypos = tf.concat([tf.zeros([batch_size,1]),tf.ones([batch_size,1])],1)
loss_utilx = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=futilx,labels=yutil))
loss_utilz = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=futilz,labels=yutil))
vars_filt = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,scope='filt')
vars_util = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,scope='util')
fprivx = [[] for i in range(K)]
fprivz = [[] for i in range(K)]
loss_privx = [[] for i in range(K)]
loss_privz = [[] for i in range(K)]
loss = [[] for i in range(K)]
vars_priv = [[] for i in range(K)]
optim_max = [[] for i in range(K)]
optim_min = [[] for i in range(K)]
optimizer_min = tf.train.MomentumOptimizer(lr,0.9)
for i in range(K):
fprivx[i],_ = NN_priv(filtx,'priv'+str(i),reuse=False)
fprivz[i],_ = NN_priv(filtz,'priv'+str(i),reuse=True)
loss_privx[i] = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=fprivx[i],labels=yneg))
loss_privz[i] = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=fprivz[i],labels=ypos))
loss[i] = rho*loss_utilx -0.5*loss_privx[i] -0.5*loss_privz[i] + lamb*(reg_filt+reg_util)
vars_priv[i] = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, 'priv'+str(i))
optim_min[i] = optimizer_min.minimize(loss[i],var_list=vars_filt+vars_util)
optim_max[i] = tf.train.MomentumOptimizer(lr,0.9).minimize(-loss[i],var_list=vars_priv[i])
## test accuracy
iswrongu = tf.not_equal(tf.argmax(futilx,1), tf.argmax(yutil, 1))
sumu = tf.reduce_sum(tf.cast(iswrongu, tf.float32))
## Misc
saver = tf.train.Saver()
#####################################################################################################################
sess = tf.Session()
sess.run(tf.global_variables_initializer())
errs_src = np.nan*np.ones((int(np.ceil(max_iter/np.float(nskip))),ntrial))
errs_tar = np.nan*np.ones((int(np.ceil(max_iter/np.float(nskip))),ntrial))
gen_source_batch = batch_generator([X1train, y1train], batchsize)
gen_target_batch = batch_generator([X2train, y1train], batchsize)
print '\nK=%d, J=%d'%(K,max_step)
## Train
for trial in range(ntrial):
#print '%d/%d'%(trial,ntrial)
cnt = 0
sess.run(tf.global_variables_initializer())
for it in range(max_iter):
X, Y = gen_source_batch.next()
Z, _ = gen_target_batch.next()
feed_dict = {x:X, z:Z, yutil:Y, batch_size:batchsize}
minimax_kbeam(sess, feed_dict)
if it%nskip == 0:
terr_src,terr_tar = evaluate(X1test,X2test,y1test,y1test)
errs_src[cnt,trial] = terr_src
errs_tar[cnt,trial] = terr_tar
cnt += 1
print 'trial %d/%d, step %d: test err src=%g, tar=%g'%(trial,ntrial,it,terr_src,terr_tar)