model.py
11.1 KB
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
import tensorflow as tf
import tensorflow.contrib.slim as slim
class DTN(object):
"""Domain Transfer Network
"""
def __init__(self, mode='train', learning_rate=0.0003):
self.mode = mode
self.learning_rate = learning_rate
def content_extractor(self, images, reuse=False):
# images: (batch, 32, 32, 3) or (batch, 32, 32, 1)
if images.get_shape()[3] == 1:
# For mnist dataset, replicate the gray scale image 3 times.
images = tf.image.grayscale_to_rgb(images)
with tf.variable_scope('content_extractor', reuse=reuse):
with slim.arg_scope([slim.conv2d], padding='SAME', activation_fn=None,
stride=2, weights_initializer=tf.contrib.layers.xavier_initializer()):
with slim.arg_scope([slim.batch_norm], decay=0.95, center=True, scale=True,
activation_fn=tf.nn.relu, is_training=(self.mode=='train' or self.mode=='pretrain')):
net = slim.conv2d(images, 64, [3, 3], scope='conv1') # (batch_size, 16, 16, 64)
net = slim.batch_norm(net, scope='bn1')
net = slim.conv2d(net, 128, [3, 3], scope='conv2') # (batch_size, 8, 8, 128)
net = slim.batch_norm(net, scope='bn2')
net = slim.conv2d(net, 256, [3, 3], scope='conv3') # (batch_size, 4, 4, 256)
net = slim.batch_norm(net, scope='bn3')
net = slim.conv2d(net, 128, [4, 4], padding='VALID', scope='conv4') # (batch_size, 1, 1, 128)
net = slim.batch_norm(net, activation_fn=tf.nn.tanh, scope='bn4')
if self.mode == 'pretrain':
net = slim.conv2d(net, 10, [1, 1], padding='VALID', scope='out')
net = slim.flatten(net)
return net
def generator(self, inputs, reuse=False):
# inputs: (batch, 1, 1, 128)
with tf.variable_scope('generator', reuse=reuse):
with slim.arg_scope([slim.conv2d_transpose], padding='SAME', activation_fn=None,
stride=2, weights_initializer=tf.contrib.layers.xavier_initializer()):
with slim.arg_scope([slim.batch_norm], decay=0.95, center=True, scale=True,
activation_fn=tf.nn.relu, is_training=(self.mode=='train')):
net = slim.conv2d_transpose(inputs, 512, [4, 4], padding='VALID', scope='conv_transpose1') # (batch_size, 4, 4, 512)
net = slim.batch_norm(net, scope='bn1')
net = slim.conv2d_transpose(net, 256, [3, 3], scope='conv_transpose2') # (batch_size, 8, 8, 256)
net = slim.batch_norm(net, scope='bn2')
net = slim.conv2d_transpose(net, 128, [3, 3], scope='conv_transpose3') # (batch_size, 16, 16, 128)
net = slim.batch_norm(net, scope='bn3')
net = slim.conv2d_transpose(net, 1, [3, 3], activation_fn=tf.nn.tanh, scope='conv_transpose4') # (batch_size, 32, 32, 1)
return net
def discriminator(self, images, reuse=False):
# images: (batch, 32, 32, 1)
with tf.variable_scope('discriminator', reuse=reuse):
with slim.arg_scope([slim.conv2d], padding='SAME', activation_fn=None,
stride=2, weights_initializer=tf.contrib.layers.xavier_initializer()):
with slim.arg_scope([slim.batch_norm], decay=0.95, center=True, scale=True,
activation_fn=tf.nn.relu, is_training=(self.mode=='train')):
net = slim.conv2d(images, 128, [3, 3], activation_fn=tf.nn.relu, scope='conv1') # (batch_size, 16, 16, 128)
net = slim.batch_norm(net, scope='bn1')
net = slim.conv2d(net, 256, [3, 3], scope='conv2') # (batch_size, 8, 8, 256)
net = slim.batch_norm(net, scope='bn2')
net = slim.conv2d(net, 512, [3, 3], scope='conv3') # (batch_size, 4, 4, 512)
net = slim.batch_norm(net, scope='bn3')
net = slim.conv2d(net, 1, [4, 4], padding='VALID', scope='conv4') # (batch_size, 1, 1, 1)
net = slim.flatten(net)
return net
def build_model(self):
if self.mode == 'pretrain':
self.images = tf.placeholder(tf.float32, [None, 32, 32, 3], 'svhn_images')
self.labels = tf.placeholder(tf.int64, [None], 'svhn_labels')
# logits and accuracy
self.logits = self.content_extractor(self.images)
self.pred = tf.argmax(self.logits, 1)
self.correct_pred = tf.equal(self.pred, self.labels)
self.accuracy = tf.reduce_mean(tf.cast(self.correct_pred, tf.float32))
# loss and train op
self.loss = slim.losses.sparse_softmax_cross_entropy(self.logits, self.labels)
self.optimizer = tf.train.AdamOptimizer(self.learning_rate)
self.train_op = slim.learning.create_train_op(self.loss, self.optimizer)
# summary op
loss_summary = tf.summary.scalar('classification_loss', self.loss)
accuracy_summary = tf.summary.scalar('accuracy', self.accuracy)
self.summary_op = tf.summary.merge([loss_summary, accuracy_summary])
elif self.mode == 'eval':
self.images = tf.placeholder(tf.float32, [None, 32, 32, 3], 'svhn_images')
# source domain (svhn to mnist)
self.fx = self.content_extractor(self.images)
self.sampled_images = self.generator(self.fx)
elif self.mode == 'train':
self.src_images = tf.placeholder(tf.float32, [None, 32, 32, 3], 'svhn_images')
self.trg_images = tf.placeholder(tf.float32, [None, 32, 32, 1], 'mnist_images')
# source domain (svhn to mnist)
self.fx = self.content_extractor(self.src_images)
self.fake_images = self.generator(self.fx)
self.logits = self.discriminator(self.fake_images)
self.fgfx = self.content_extractor(self.fake_images, reuse=True)
# loss
self.d_loss_src = slim.losses.sigmoid_cross_entropy(self.logits, tf.zeros_like(self.logits))
self.g_loss_src = slim.losses.sigmoid_cross_entropy(self.logits, tf.ones_like(self.logits))
self.f_loss_src = tf.reduce_mean(tf.square(self.fx - self.fgfx)) * 15.0
# optimizer
self.d_optimizer_src = tf.train.AdamOptimizer(self.learning_rate)
self.g_optimizer_src = tf.train.AdamOptimizer(self.learning_rate)
self.f_optimizer_src = tf.train.AdamOptimizer(self.learning_rate)
t_vars = tf.trainable_variables()
d_vars = [var for var in t_vars if 'discriminator' in var.name]
g_vars = [var for var in t_vars if 'generator' in var.name]
f_vars = [var for var in t_vars if 'content_extractor' in var.name]
# train op
with tf.name_scope('source_train_op'):
self.d_train_op_src = slim.learning.create_train_op(self.d_loss_src, self.d_optimizer_src, variables_to_train=d_vars)
self.g_train_op_src = slim.learning.create_train_op(self.g_loss_src, self.g_optimizer_src, variables_to_train=g_vars)
self.f_train_op_src = slim.learning.create_train_op(self.f_loss_src, self.f_optimizer_src, variables_to_train=f_vars)
# summary op
d_loss_src_summary = tf.summary.scalar('src_d_loss', self.d_loss_src)
g_loss_src_summary = tf.summary.scalar('src_g_loss', self.g_loss_src)
f_loss_src_summary = tf.summary.scalar('src_f_loss', self.f_loss_src)
origin_images_summary = tf.summary.image('src_origin_images', self.src_images)
sampled_images_summary = tf.summary.image('src_sampled_images', self.fake_images)
self.summary_op_src = tf.summary.merge([d_loss_src_summary, g_loss_src_summary,
f_loss_src_summary, origin_images_summary,
sampled_images_summary])
# target domain (mnist)
self.fx = self.content_extractor(self.trg_images, reuse=True)
self.reconst_images = self.generator(self.fx, reuse=True)
self.logits_fake = self.discriminator(self.reconst_images, reuse=True)
self.logits_real = self.discriminator(self.trg_images, reuse=True)
# loss
self.d_loss_fake_trg = slim.losses.sigmoid_cross_entropy(self.logits_fake, tf.zeros_like(self.logits_fake))
self.d_loss_real_trg = slim.losses.sigmoid_cross_entropy(self.logits_real, tf.ones_like(self.logits_real))
self.d_loss_trg = self.d_loss_fake_trg + self.d_loss_real_trg
self.g_loss_fake_trg = slim.losses.sigmoid_cross_entropy(self.logits_fake, tf.ones_like(self.logits_fake))
self.g_loss_const_trg = tf.reduce_mean(tf.square(self.trg_images - self.reconst_images)) * 15.0
self.g_loss_trg = self.g_loss_fake_trg + self.g_loss_const_trg
# optimizer
self.d_optimizer_trg = tf.train.AdamOptimizer(self.learning_rate)
self.g_optimizer_trg = tf.train.AdamOptimizer(self.learning_rate)
# train op
with tf.name_scope('target_train_op'):
self.d_train_op_trg = slim.learning.create_train_op(self.d_loss_trg, self.d_optimizer_trg, variables_to_train=d_vars)
self.g_train_op_trg = slim.learning.create_train_op(self.g_loss_trg, self.g_optimizer_trg, variables_to_train=g_vars)
# summary op
d_loss_fake_trg_summary = tf.summary.scalar('trg_d_loss_fake', self.d_loss_fake_trg)
d_loss_real_trg_summary = tf.summary.scalar('trg_d_loss_real', self.d_loss_real_trg)
d_loss_trg_summary = tf.summary.scalar('trg_d_loss', self.d_loss_trg)
g_loss_fake_trg_summary = tf.summary.scalar('trg_g_loss_fake', self.g_loss_fake_trg)
g_loss_const_trg_summary = tf.summary.scalar('trg_g_loss_const', self.g_loss_const_trg)
g_loss_trg_summary = tf.summary.scalar('trg_g_loss', self.g_loss_trg)
origin_images_summary = tf.summary.image('trg_origin_images', self.trg_images)
sampled_images_summary = tf.summary.image('trg_reconstructed_images', self.reconst_images)
self.summary_op_trg = tf.summary.merge([d_loss_trg_summary, g_loss_trg_summary,
d_loss_fake_trg_summary, d_loss_real_trg_summary,
g_loss_fake_trg_summary, g_loss_const_trg_summary,
origin_images_summary, sampled_images_summary])
for var in tf.trainable_variables():
tf.summary.histogram(var.op.name, var)