-
Notifications
You must be signed in to change notification settings - Fork 9
/
Copy pathdfc_vae_model.py
234 lines (186 loc) · 9.92 KB
/
dfc_vae_model.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
import numpy as np
import tensorflow as tf
from vgg16 import vgg16
class dfc_vae_model(object):
def __init__(self, shape, inputs, alpha = 1, beta = 0.5, vgg_layers = [], learning_rate = 0.0005):
self.shape = shape
self.img_input = inputs
self.alpha = alpha
self.beta = beta
self.gstep = tf.Variable(0, dtype=tf.int32, trainable=False, name='global_step')
self.vgg_layers = vgg_layers
self.learning_rate = learning_rate
def _get_weights(self, name, shape):
with tf.variable_scope("weights", reuse=tf.AUTO_REUSE) as scope:
w = tf.get_variable(name=name + '_W',
shape=shape,
initializer=tf.truncated_normal_initializer(stddev=0.1))
return w
def _get_biases(self, name, shape):
with tf.variable_scope("biases", reuse=tf.AUTO_REUSE) as scope:
b = tf.get_variable(name=name + '_b',
shape=shape,
initializer=tf.truncated_normal_initializer(stddev=0.1))
return b
def _conv2d_bn_relu(self, inputs, name, kernel_size, in_channel, out_channel, stride, activation=True,bn=True):
with tf.variable_scope(name) as scope:
### setup weights and biases
filters = self._get_weights(name, shape=[kernel_size, kernel_size, in_channel, out_channel])
biases = self._get_biases(name, shape=[out_channel])
### convolutional neural network
conv2d = tf.nn.conv2d(input=inputs,
filter=filters,
strides=[1,stride,stride,1],
padding='SAME',
name=name + '_conv')
conv2d = tf.nn.bias_add(conv2d, biases, name=name+'_add')
### in case of batch normalization
if bn == True:
conv2d = tf.contrib.layers.batch_norm(conv2d,
center=True, scale=True,
is_training=True,
scope='bn')
### in case of leaky relu activation
if activation == True:
conv2d = tf.nn.leaky_relu(conv2d, alpha=0.1, name=name)
return conv2d
def encoder(self, reuse=False):
with tf.variable_scope("encoder", reuse = reuse):
### Conv2d_bn_relu Layer 1
conv1 = self._conv2d_bn_relu(self.img_input,
name="conv1",
kernel_size=4,
in_channel=3,
out_channel=32,
stride=2)
### Conv2d_bn_relu Layer 2
conv2 = self._conv2d_bn_relu(conv1,
name="conv2",
kernel_size=4,
in_channel=32,
out_channel=64,
stride=2)
### Conv2d_bn_relu Layer 3
conv3 = self._conv2d_bn_relu(conv2,
name="conv3",
kernel_size=4,
in_channel=64,
out_channel=128,
stride=2)
### Conv2d_bn_relu Layer 4
conv4 = self._conv2d_bn_relu(conv3,
name="conv4",
kernel_size=4,
in_channel=128,
out_channel=256,
stride=2)
### flatten the output
conv4_flat = tf.reshape(conv4, [-1, 256*4*4])
### FC Layer for mean
fcmean = tf.layers.dense(inputs=conv4_flat,
units=100,
activation=None,
name="fcmean")
### FC Layer for standard deviation
fcstd = tf.layers.dense(inputs=conv4_flat,
units=100,
activation=None,
name="fcstd")
### fcmean and fcstd will be used for sample z value (latent variables)
return fcmean, fcstd + 1e-6
def decoder(self,inputs, reuse=False):
with tf.variable_scope("decoder", reuse = reuse):
### FC Layer for z
fc = tf.layers.dense(inputs=inputs,
units = 4096,
activation = None)
fc = tf.reshape(fc, [-1, 4, 4, 256])
### Layer 1
deconv1 = tf.image.resize_nearest_neighbor(fc, size=(8,8))
deconv1 = self._conv2d_bn_relu(deconv1,
name="deconv1",
kernel_size=3,
in_channel=256,
out_channel=128,
stride=1)
### Layer 2
deconv2 = tf.image.resize_nearest_neighbor(deconv1, size=(16,16))
deconv2 = self._conv2d_bn_relu(deconv2,
name="deconv2",
kernel_size=3,
in_channel=128,
out_channel=64,
stride=1)
### Layer 3
deconv3 = tf.image.resize_nearest_neighbor(deconv2, size=(32,32))
deconv3 = self._conv2d_bn_relu(deconv3,
name="deconv3",
kernel_size=3,
in_channel=64,
out_channel=32,
stride=1)
### Layer 4
deconv4 = tf.image.resize_nearest_neighbor(deconv3, size=(64,64))
deconv4 = self._conv2d_bn_relu(deconv4,
name="deconv4",
kernel_size=3,
in_channel=32,
out_channel=3,
stride=1,
activation=False,
bn=False)
return deconv4
def load_vgg(self):
### pass the input image to VGG model
#self.resize_input_img = tf.image.resize_images(self.img_input, [224,224])
#self.vgg_input = VGG(self.resize_input_img)
#self.l1_r, self.l2_r, self.l3_r = self.vgg_input.load(reuse=False)
### pass the generated image to VGG model
#self.resize_gen_img = tf.image.resize_images(self.gen_img, [224,224])
#self.vgg_gen = VGG(self.resize_gen_img)
#self.l1_g, self.l2_g, self.l3_g = self.vgg_gen.load(reuse=True)
self.resize_input_img = tf.image.resize_images(self.img_input, [224,224])
self.vgg_real = vgg16(self.resize_input_img, 'vgg16_weights.npz')
self.l1_r, self.l2_r, self.l3_r = self.vgg_real.get_layers()
self.resize_gen_img = tf.image.resize_images(self.gen_img, [224,224])
self.vgg_gen = vgg16(self.resize_gen_img, 'vgg16_weights.npz')
self.l1_g, self.l2_g, self.l3_g = self.vgg_gen.get_layers()
def calculate_loss(self):
### calculate perception loss
#l1_loss = (tf.reduce_sum(tf.square(self.l1_r-self.l1_g)))/tf.cast(tf.size(self.l1_r), tf.float32)
#l2_loss = (tf.reduce_sum(tf.square(self.l2_r-self.l2_g)))/tf.cast(tf.size(self.l2_r), tf.float32)
#l3_loss = (tf.reduce_sum(tf.square(self.l3_r-self.l3_g)))/tf.cast(tf.size(self.l3_r), tf.float32)
l1_loss = tf.reduce_sum(tf.square(self.l1_r-self.l1_g), [1,2,3])
l2_loss = tf.reduce_sum(tf.square(self.l2_r-self.l2_g), [1,2,3])
l3_loss = tf.reduce_sum(tf.square(self.l3_r-self.l3_g), [1,2,3])
self.pct_loss = tf.reduce_mean(l1_loss + l2_loss + l3_loss)
### calculate KL loss
self.kl_loss = tf.reduce_mean(-0.5*tf.reduce_sum(
1 + self.std - tf.square(self.mean) - tf.exp(self.std), 1))
### calculate total loss
self.loss = tf.add(self.beta*self.pct_loss,self.alpha*self.kl_loss)
def optimize(self):
### create optimizer
var_list = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='encoder') + tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='decoder')
self.optimizer = tf.train.AdamOptimizer(self.learning_rate).minimize(self.loss,global_step=self.gstep, var_list=var_list)
def build_model(self,reuse=tf.AUTO_REUSE):
### get mean and std from encoder
self.mean, self.std = self.encoder(reuse)
### sampling z and use reparameterization trick
epsilon = tf.random_normal((tf.shape(self.mean)[0],100), mean = 0.0, stddev=1.0)
self.z = self.mean + epsilon * tf.exp(.5*self.std)
### decode to get a generated image
self.gen_img = self.decoder(self.z,reuse)
### load vgg
self.load_vgg()
### calculate loss
self.calculate_loss()
### setup optimizer
self.optimize()
### generate random latent variable for random images
self.random_latent = tf.random_normal((tf.shape(self.mean)[0], 100))
self.ran_img = self.decoder(self.random_latent,reuse)
### load VGG weight
def load_vgg_weight(self, weight_file, sess):
self.vgg_real.load_weights(weight_file,sess)
self.vgg_gen.load_weights(weight_file,sess)