-
Notifications
You must be signed in to change notification settings - Fork 0
/
module.py
185 lines (148 loc) · 8.47 KB
/
module.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
from __future__ import division
import tensorflow as tf
from ops import *
from utils import *
def discriminator(image, options, reuse=False, name="discriminator"):
with tf.variable_scope(name):
# image is 256 x 256 x input_c_dim
if reuse:
tf.get_variable_scope().reuse_variables()
else:
assert tf.get_variable_scope().reuse is False
h0 = lrelu(conv2d(image, options.df_dim, name='d_h0_conv'))
# h0 is (128 x 128 x self.df_dim)
h1 = lrelu(instance_norm(conv2d(h0, options.df_dim*2, name='d_h1_conv'), 'd_bn1'))
# h1 is (64 x 64 x self.df_dim*2)
h2 = lrelu(instance_norm(conv2d(h1, options.df_dim*4, name='d_h2_conv'), 'd_bn2'))
# h2 is (32x 32 x self.df_dim*4)
h3 = lrelu(instance_norm(conv2d(h2, options.df_dim*8, s=1, name='d_h3_conv'), 'd_bn3'))
# h3 is (32 x 32 x self.df_dim*8)
h4 = conv2d(h3, 1, s=1, name='d_h3_pred')
# h4 is (32 x 32 x 1)
return h4
def generator_unet(image, options, reuse=False, name="generator"):
dropout_rate = 0.5 if options.is_training else 1.0
with tf.variable_scope(name):
# image is 256 x 256 x input_c_dim
if reuse:
tf.get_variable_scope().reuse_variables()
else:
assert tf.get_variable_scope().reuse is False
# image is (256 x 256 x input_c_dim)
e1 = instance_norm(conv2d(image, options.gf_dim, name='g_e1_conv'))
# e1 is (128 x 128 x self.gf_dim)
e2 = instance_norm(conv2d(lrelu(e1), options.gf_dim*2, name='g_e2_conv'), 'g_bn_e2')
# e2 is (64 x 64 x self.gf_dim*2)
e3 = instance_norm(conv2d(lrelu(e2), options.gf_dim*4, name='g_e3_conv'), 'g_bn_e3')
# e3 is (32 x 32 x self.gf_dim*4)
e4 = instance_norm(conv2d(lrelu(e3), options.gf_dim*8, name='g_e4_conv'), 'g_bn_e4')
# e4 is (16 x 16 x self.gf_dim*8)
e5 = instance_norm(conv2d(lrelu(e4), options.gf_dim*8, name='g_e5_conv'), 'g_bn_e5')
# e5 is (8 x 8 x self.gf_dim*8)
e6 = instance_norm(conv2d(lrelu(e5), options.gf_dim*8, name='g_e6_conv'), 'g_bn_e6')
# e6 is (4 x 4 x self.gf_dim*8)
e7 = instance_norm(conv2d(lrelu(e6), options.gf_dim*8, name='g_e7_conv'), 'g_bn_e7')
# e7 is (2 x 2 x self.gf_dim*8)
e8 = instance_norm(conv2d(lrelu(e7), options.gf_dim*8, name='g_e8_conv'), 'g_bn_e8')
# e8 is (1 x 1 x self.gf_dim*8)
'''
#for concatenation
temp_o_s1 = [options.batch_size, s128, s128, self.gf_dim * 8]
temp_d1_w = tf.get_variable('temp_d1_w', [4, 4, temp_o_s1[-1], e8_for_concat.get_shape()[-1]],
initializer=tf.truncated_normal_initializer(stddev=0.02))
'''
d1 = deconv2d(tf.nn.relu(e8), options.gf_dim*8, name='g_d1')
d1 = tf.nn.dropout(d1, dropout_rate)
d1 = tf.concat([instance_norm(d1, 'g_bn_d1'), e7], 3)
# d1 is (2 x 2 x self.gf_dim*8*2)
d2 = deconv2d(tf.nn.relu(d1), options.gf_dim*8, name='g_d2')
d2 = tf.nn.dropout(d2, dropout_rate)
d2 = tf.concat([instance_norm(d2, 'g_bn_d2'), e6], 3)
# d2 is (4 x 4 x self.gf_dim*8*2)
d3 = deconv2d(tf.nn.relu(d2), options.gf_dim*8, name='g_d3')
d3 = tf.nn.dropout(d3, dropout_rate)
d3 = tf.concat([instance_norm(d3, 'g_bn_d3'), e5], 3)
# d3 is (8 x 8 x self.gf_dim*8*2)
d4 = deconv2d(tf.nn.relu(d3), options.gf_dim*8, name='g_d4')
d4 = tf.concat([instance_norm(d4, 'g_bn_d4'), e4], 3)
# d4 is (16 x 16 x self.gf_dim*8*2)
d5 = deconv2d(tf.nn.relu(d4), options.gf_dim*4, name='g_d5')
d5 = tf.concat([instance_norm(d5, 'g_bn_d5'), e3], 3)
# d5 is (32 x 32 x self.gf_dim*4*2)
d6 = deconv2d(tf.nn.relu(d5), options.gf_dim*2, name='g_d6')
d6 = tf.concat([instance_norm(d6, 'g_bn_d6'), e2], 3)
# d6 is (64 x 64 x self.gf_dim*2*2)
d7 = deconv2d(tf.nn.relu(d6), options.gf_dim, name='g_d7')
d7 = tf.concat([instance_norm(d7, 'g_bn_d7'), e1], 3)
# d7 is (128 x 128 x self.gf_dim*1*2)
d8 = deconv2d(tf.nn.relu(d7), options.output_c_dim, name='g_d8')
# d8 is (256 x 256 x output_c_dim)
return tf.nn.tanh(d8)
def channel_wise_fc_layer(input, name): # bottom: (7x7x512)
_, width, height, n_feat_map = input.get_shape().as_list()
input_reshape = tf.reshape( input, [-1, width*height, n_feat_map] )
input_transpose = tf.transpose( input_reshape, [2,0,1] )
with tf.variable_scope(name):
W = tf.get_variable(
"W",
shape=[n_feat_map,width*height, width*height], # (512,49,49)
initializer=tf.truncated_normal_initializer(stddev=0.005))
output = tf.matmul(input_transpose, W)
output_transpose = tf.transpose(output, [1,2,0])
output_reshape = tf.reshape( output_transpose, [-1, height, width, n_feat_map] )
return output_reshape
def generator_resnet(image, options, reuse=False, name="generator"):
with tf.variable_scope(name):
# image is 256 x 256 x input_c_dim
if reuse:
tf.get_variable_scope().reuse_variables()
else:
assert tf.get_variable_scope().reuse is False
# This is a block composed of 2 convolution layer and residual connection.
def residule_block(x, dim, ks=3, s=1, name='res'):
p = int((ks - 1) / 2)
y = tf.pad(x, [[0, 0], [p, p], [p, p], [0, 0]], "REFLECT")
y = instance_norm(conv2d(y, dim, ks, s, padding='VALID', name=name+'_c1'), name+'_bn1')
y = tf.pad(tf.nn.relu(y), [[0, 0], [p, p], [p, p], [0, 0]], "REFLECT")
y = instance_norm(conv2d(y, dim, ks, s, padding='VALID', name=name+'_c2'), name+'_bn2')
return y + x
def depthwise_residual_block(x, dim, ks=1, s=1, name='res'):
p = int((ks - 1) / 2)
y = tf.pad(x, [[0, 0], [p, p], [p, p], [0, 0]], "REFLECT")
y = instance_norm(depthwiseconv2d(y, filters=[ks, ks, 256, 1], padding='VALID', name=name+'_c1'), name+'_bn1')
y = tf.nn.relu(y)
#y = tf.pad(tf.nn.relu(y), [[0, 0], [p, p], [p, p], [0, 0]], "REFLECT")
#y = instance_norm(conv2d(y, dim, ks, s, padding='VALID', name=name+'_c2'), name+'_bn2')
return y + x
# c0 - c3 and r1 - r5 are encoder of generator networks.
c0 = tf.pad(image, [[0, 0], [3, 3], [3, 3], [0, 0]], "REFLECT")
c1 = tf.nn.relu(instance_norm(conv2d(c0, options.gf_dim, 7, 1, padding='VALID', name='g_e1_c'), 'g_e1_bn'))
c2 = tf.nn.relu(instance_norm(conv2d(c1, options.gf_dim*2, 3, 2, name='g_e2_c'), 'g_e2_bn'))
c3 = tf.nn.relu(instance_norm(conv2d(c2, options.gf_dim*4, 3, 2, name='g_e3_c'), 'g_e3_bn'))
r1 = residule_block(c3, options.gf_dim*4, name='g_r1')
r2 = residule_block(r1, options.gf_dim*4, name='g_r2')
r3 = residule_block(r2, options.gf_dim*4, name='g_r3')
r4 = residule_block(r3, options.gf_dim*4, name='g_r4')
r5 = residule_block(r4, options.gf_dim * 4, name='g_r5')
# it is a feature mapping layer used to connect encoder and decoder network.
temp = tf.nn.relu(instance_norm(conv2d(r5, options.gf_dim * 4, ks=1, s=1, padding='SAME', name='temp')))
temp = temp + r5
# r6 - r10, d1 - pred are decoder of generator networks.
r6 = residule_block(temp, options.gf_dim*4, name='g_r6')
r7 = residule_block(r6, options.gf_dim*4, name='g_r7')
r8 = residule_block(r7, options.gf_dim*4, name='g_r8')
r9 = residule_block(r8, options.gf_dim*4, name='g_r9')
r10 = residule_block(r9, options.gf_dim*4, name='g_r10')
d1 = deconv2d(r10, options.gf_dim*2, 3, 2, name='g_d1_dc')
d1 = tf.nn.relu(instance_norm(d1, 'g_d1_bn'))
d2 = deconv2d(d1, options.gf_dim, 3, 2, name='g_d2_dc')
d2 = tf.nn.relu(instance_norm(d2, 'g_d2_bn'))
d2 = tf.pad(d2, [[0, 0], [3, 3], [3, 3], [0, 0]], "REFLECT")
pred = tf.nn.tanh(conv2d(d2, options.output_c_dim, 7, 1, padding='VALID', name='g_pred_c'))
return pred
def abs_criterion(in_, target):
return tf.reduce_mean(tf.abs(in_ - target))
def mae_criterion(in_, target):
return tf.reduce_mean((in_-target)**2)
def sce_criterion(logits, labels):
return tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=logits, labels=labels))