-
Notifications
You must be signed in to change notification settings - Fork 14
/
Copy pathATDA.py
265 lines (207 loc) · 13.2 KB
/
ATDA.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
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
from tensorlayer.layers import Conv2d,InputLayer,MaxPool2d,FlattenLayer,BatchNormLayer,DenseLayer,DropoutLayer
from tensorlayer.cost import cross_entropy
import tensorlayer as tl
import tensorflow as tf
import numpy as np
import random
import utils
def iterate_minibatches(inputs, targets, batchsize, shuffle=False):
assert len(inputs) == len(targets)
if shuffle:
indices = np.arange(len(inputs))
np.random.shuffle(indices)
for start_idx in range(0, len(inputs), batchsize):
end_idx = start_idx + batchsize
if end_idx > len(inputs):
end_idx = start_idx + (len(inputs) % batchsize)
if shuffle:
excerpt = indices[start_idx:end_idx]
else:
excerpt = slice(start_idx, end_idx)
yield inputs[excerpt], targets[excerpt]
def sample_candidatas(dataset,labels,candidates_num,shuffle=True):
if shuffle:
indices = np.arange(len(dataset))
np.random.shuffle(indices)
excerpt = indices[0:candidates_num]
candidates = dataset[excerpt]
true_label = labels[excerpt]
return candidates,true_label
def labeling_fuc(data,output1, output2, true_label, threshold=0.9):
id = np.equal(np.argmax(output1,1),np.argmax(output2,1))
output1 = output1[id,:]
output2 = output2[id, :]
data = data[id, :]
true_label = true_label[id, :]
max1 = np.max(output1,1)
max2 = np.max(output2,1)
id2=np.max(np.vstack((max1,max2)),0)>threshold
output1 = output1[id2,:]
data = data[id2, :]
pseudo_label =utils.dense_to_one_hot(np.argmax(output1,1),10)
true_label = true_label[id2, :]
return data,pseudo_label,true_label
class ATDA(object):
def __init__(self, sess,name='mnist-mnistm'):
self.name=name
self.sess=sess
def create_model(self):
if self.name == 'mnist-mnistm':
drop_prob={'Ft':0.2, 'F1':0.5, 'F2':0.5}
self.x = tf.placeholder(tf.float32, shape=[None, 28, 28, 3])
self.y_ = tf.placeholder(tf.float32, shape=[None,10])
self.istrain = tf.placeholder(tf.bool, shape=[])
_input = InputLayer(self.x,name='input_layer')
_shared_net = _input
_shared_net=Conv2d(_shared_net, n_filter=32, filter_size=(5, 5), strides=(1, 1),
act=tf.nn.relu, padding='SAME', name='cnn1')
_shared_net = MaxPool2d(_shared_net, filter_size=(2, 2), strides=(2, 2),
padding='SAME', name='pool_layer1')
_shared_net =Conv2d(_shared_net, n_filter=48, filter_size=(5, 5), strides=(1, 1),
act=tf.identity, padding='SAME', name='cnn2')
_shared_net = BatchNormLayer(_shared_net,is_train=True,act=tf.nn.relu)
_shared_net = MaxPool2d(_shared_net, filter_size=(2, 2), strides=(2, 2),
padding='SAME', name='pool_layer2')
_shared_net=FlattenLayer(_shared_net)
feature=_shared_net.outputs
_F1_net=_shared_net
_F2_net=_shared_net
_Ft_net=_shared_net
with tf.variable_scope("F1") as scope:
_F1_net = DropoutLayer(_F1_net, keep=drop_prob['F1'], name='drop1',is_fix=True,is_train=self.istrain)
_F1_net=DenseLayer(_F1_net,n_units=100,
act=tf.identity, name='relu1')
_F1_net = BatchNormLayer(_F1_net,is_train=True,act=tf.nn.relu,name='bn1')
_F1_net = DropoutLayer(_F1_net, keep=drop_prob['F1'], name='drop2',is_fix=True,is_train=self.istrain)
_F1_net = DenseLayer(_F1_net,n_units=100,
act=tf.identity, name='relu2')
_F1_net = BatchNormLayer(_F1_net, is_train=True, act=tf.nn.relu,name='bn2')
_F1_net = DenseLayer(_F1_net,n_units=10,
act=tf.nn.softmax, name='output')
self.F1_out=_F1_net.outputs
with tf.variable_scope("F2") as scope:
_F2_net = DropoutLayer(_F2_net, keep=drop_prob['F2'], name='drop1',is_fix=True,is_train=self.istrain)
_F2_net=DenseLayer(_F2_net,n_units=100,
act=tf.identity, name='relu1')
_F2_net = BatchNormLayer(_F2_net,is_train=True,act=tf.nn.relu,name='bn1')
_F2_net = DropoutLayer(_F2_net, keep=drop_prob['F2'], name='drop2',is_fix=True,is_train=self.istrain)
_F2_net = DenseLayer(_F2_net,n_units=100,
act=tf.identity, name='relu2')
_F2_net = BatchNormLayer(_F2_net, is_train=True, act=tf.nn.relu,name='bn2')
_F2_net = DenseLayer(_F2_net,n_units=10,
act=tf.nn.softmax, name='output')
self.F2_out = _F2_net.outputs
with tf.variable_scope("Ft") as scope:
_Ft_net = DropoutLayer(_Ft_net, keep=drop_prob['Ft'], name='drop1',is_fix=True,is_train=self.istrain)
_Ft_net=DenseLayer(_Ft_net,n_units=100,
act=tf.identity, name='relu1')
_Ft_net = BatchNormLayer(_Ft_net,is_train=True,act=tf.nn.relu,name='bn1')
_Ft_net = DropoutLayer(_Ft_net, keep=drop_prob['Ft'], name='drop2',is_fix=True,is_train=self.istrain)
_Ft_net = DenseLayer(_Ft_net,n_units=100,
act=tf.identity, name='relu2')
_Ft_net = BatchNormLayer(_Ft_net, is_train=True, act=tf.nn.relu,name='bn2')
_Ft_net = DenseLayer(_Ft_net,n_units=10,
act=tf.nn.softmax, name='output')
self.Ft_out = _Ft_net.outputs
#self.cost = cross_entropy(F1_out,self.y_,name='F1_loss')#+cross_entropy(F2_out,self.y_,name='F2_loss')+cross_entropy(Ft_out,self.y_,name='Ft_loss')
self.F1_loss = -tf.reduce_mean(self.y_ * tf.log(self.F1_out))
self.F2_loss = -tf.reduce_mean(self.y_*tf.log(self.F2_out))
self.Ft_loss = -tf.reduce_mean(self.y_*tf.log(self.Ft_out))
self.F1_acc = tf.reduce_mean(tf.cast(tf.equal(tf.argmax(self.F1_out,1),tf.argmax(self.y_,1)),tf.float32))
self.F2_acc = tf.reduce_mean(tf.cast(tf.equal(tf.argmax(self.F2_out, 1), tf.argmax(self.y_, 1)), tf.float32))
self.Ft_acc = tf.reduce_mean(tf.cast(tf.equal(tf.argmax(self.Ft_out, 1), tf.argmax(self.y_, 1)), tf.float32))
self.cost =self.F1_loss+self.F2_loss+self.Ft_loss+tf.reduce_sum(tf.abs(tf.multiply(tf.transpose(_F1_net.all_params[14]),_F2_net.all_params[14])))
self.labeling_cost = self.F1_loss+self.F2_loss+tf.reduce_sum(tf.abs(tf.multiply(tf.transpose(_F1_net.all_params[14]),_F2_net.all_params[14])))
self.targetspecific_cost = self.Ft_loss
self.F1F2Ft_op = tf.train.AdamOptimizer(learning_rate = 0.01).minimize(self.cost)
self.F1F2_op =tf.train.AdamOptimizer(learning_rate = 0.01).minimize(self.labeling_cost)
self.Ft_op = tf.train.AdamOptimizer(learning_rate = 0.01).minimize(self.targetspecific_cost)
tl.layers.initialize_global_variables(self.sess)
print '********************************************************************************************************************************************************'
_shared_net.print_params()
_shared_net.print_layers()
print '********************************************************************************************************************************************************'
_F1_net.print_params()
_F1_net.print_layers()
print '********************************************************************************************************************************************************'
_F2_net.print_params()
_F2_net.print_layers()
print '********************************************************************************************************************************************************'
_Ft_net.print_params()
_Ft_net.print_layers()
def fit_ATDA(self,source_train, y_train, target_val, y_val, target_data, target_label, nb_epoch=5, k_epoch=100, batch_size=128, shuffle=True,N_init=5000,N_max=40000):
n = source_train.shape[0]
print n
for e in range(nb_epoch):
n_batch = 0
for Xu_batch, Yu_batch in iterate_minibatches(source_train, y_train, batch_size, shuffle=shuffle):
feed_dict = { self.x: Xu_batch, self.y_: Yu_batch ,self.istrain:True}
_, cost, F1_loss, F2_loss, Ft_loss = self.sess.run([self.F1F2Ft_op, self.cost,self.F1_loss,self.F2_loss,self.Ft_loss], feed_dict=feed_dict)
n_batch += 1
#every 10 minibatch print loss
if n_batch % 10==0:
print("Epoch %d total_loss %f F1_loss %f F2_loss %f Ft_loss %f" % (e + 1, cost,F1_loss,F2_loss,Ft_loss))
#every 100 minibatch eval
if n_batch % 100==0:
print("**************************************************************************val_stage**************************************************************************")
feed_dict = {self.x: target_val, self.y_: y_val, self.istrain: False}
F1_acc, F2_acc, Ft_acc = self.sess.run(
[self.F1_acc,self.F2_acc,self.Ft_acc], feed_dict=feed_dict)
print("Epoch %d F1_acc %f F2_acc %f Ft_acc %f" % (
e + 1, F1_acc, F2_acc, Ft_acc))
print("\n*************************************************************************************************************************************************************")
Nt,true_label=sample_candidatas(target_data, target_label, N_init)
feed_dict = {self.x: Nt, self.istrain: False}
F1_out, F2_out = self.sess.run(
[self.F1_out, self.F2_out ], feed_dict=feed_dict)
Tl,pseudo_labels,true_labels=labeling_fuc(data=Nt,output1=F1_out,output2=F2_out,true_label=true_label)
print Tl.shape,pseudo_labels.shape,true_labels.shape
print Tl.dtype,pseudo_labels.dtype,true_labels.dtype
L=np.concatenate((source_train,Tl),0)
L_label=np.concatenate((y_train,pseudo_labels),0)
for k in range(k_epoch):
for e in range(nb_epoch):
n_batch = 0
for Xu_batch, Yu_batch in iterate_minibatches(L, L_label, batch_size, shuffle=shuffle):
feed_dict1 = {self.x: Xu_batch, self.y_: Yu_batch, self.istrain: True}
_, labeling_cost = self.sess.run([self.F1F2_op, self.labeling_cost],
feed_dict=feed_dict1)
Xu_batch,Yu_batch = sample_candidatas(Tl, pseudo_labels, batch_size)
feed_dict2 = {self.x: Xu_batch, self.y_: Yu_batch, self.istrain: True}
_ , targetspecific_cost = self.sess.run([self.Ft_op, self.targetspecific_cost],
feed_dict=feed_dict2)
n_batch += 1
if n_batch % 10 == 0:
print("Iter %d of Epoch %d labeling_cost %f targetspecific_cost %f" % (k + 1,
e + 1, labeling_cost, targetspecific_cost))
if n_batch % 100 == 0:
print(
"**************************************************************************val_stage**************************************************************************")
feed_dict = {self.x: target_val, self.y_: y_val, self.istrain: False}
F1_acc, F2_acc, Ft_acc = self.sess.run(
[self.F1_acc, self.F2_acc, self.Ft_acc], feed_dict=feed_dict)
print("Epoch %d F1_acc %f F2_acc %f Ft_acc %f" % (
e + 1, F1_acc, F2_acc, Ft_acc))
print(
"\n*************************************************************************************************************************************************************")
N=(k + 2) * n / 20
print N
if N>=N_max:
N=N_max
Nt, true_label = sample_candidatas(target_data, target_label, N)
n_batch = 0
for Xu_batch, _ in iterate_minibatches(Nt,true_label, batch_size, shuffle=False):
feed_dict = {self.x: Xu_batch, self.istrain: False}
F1_out_batch, F2_out_batch = self.sess.run([self.F1_out, self.F2_out], feed_dict=feed_dict)
n_batch += 1
if n_batch==1:
F1_out=F1_out_batch
F2_out=F2_out_batch
else:
F1_out=np.vstack((F1_out,F1_out_batch))
F2_out=np.vstack((F2_out,F2_out_batch))
Tl, pseudo_labels, true_labels = labeling_fuc(data=Nt, output1=F1_out, output2=F2_out,
true_label=true_label)
print Tl.shape, pseudo_labels.shape, true_labels.shape
L = np.concatenate((source_train, Tl), 0)
L_label = np.concatenate((y_train, pseudo_labels), 0)