-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathinpaint_model.py
211 lines (181 loc) · 9.48 KB
/
inpaint_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
import tensorflow as tf
from inpaint_ops import *
import cv2 as cv
import numpy as np
from vgg.vgg16 import *
from region_conv import *
def RW_generator(x,mask,padding='SAME',name='inpaint_net',reuse=False):
'''
Region-wise generator
Args:
x: incomplete image
mask: mask region {0,1}
returns:
predicted image
'''
x1 = SINet(x, mask, reuse=reuse)
x_combine = x * mask + x1 * (1 - mask)
x2 = GPNet(x_combine,mask, reuse=reuse)
return x1, x2
def SINet(x,mask,padding='SAME',name='inpaint_net',reuse=False):
'''
Semantic inferring network
Args:
x: incomplete image
mask: mask region {0,1}
returns:
image predicted by semantic inferring network
'''
xin=x;
mask_in=mask
cnum=32
ones_x = tf.ones_like(x)[:, :, :, 0:1] #一层
x = tf.concat([x, mask], axis=3) #拼接
with tf.variable_scope(name,reuse=reuse):
x1=standard_conv(x,mask,cnum,5,1,name='conv1')
x2=standard_conv(x1,mask,2*cnum,3,2,name='conv2_downsample')
x3=standard_conv(x2,mask,2*cnum,3,1,name='conv3')
x4=standard_conv(x3,mask,4*cnum,3,2,name='conv4_downsample')
x5=standard_conv(x4,mask,4*cnum,3,1,name='conv5')
x6=standard_conv(x5,mask,4*cnum,3,1,name='conv6')
#dilated conv
x7=standard_conv(x6,mask,4*cnum,3,rate=2,name='conv7_atrous')
x8=standard_conv(x7,mask,4*cnum,3,rate=4,name='conv8_atrous')
x9=standard_conv(x8,mask,4*cnum,3,rate=8,name='conv9_atrous')
x10=standard_conv(x9,mask,4*cnum,3,rate=16,name='conv10_atrous')
x11=standard_conv(tf.concat([x10,x6],axis=-1), mask,4*cnum,3,1,name='conv11')
x12=standard_conv(tf.concat([x11,x5],axis=-1),mask,4*cnum,3,1,name='conv12')
x_complete, x_missing = tf.concat([x12,x4],axis=-1),x12
x13 = region_deconv(x_complete, x_missing, mask,name = 'com_13')
x_complete, x_missing = tf.concat([x13,x3],axis=-1),x13
x14 = region_conv(x_complete,x_missing,mask,name='com_14')
x_complete, x_missing = tf.concat([x14,x2],axis=-1),x14
x15 = region_deconv(x_complete, x_missing, mask, name = 'com_15')
x16 = standard_conv(x15,mask,cnum,3,1,name='conv16')
x17=standard_conv(x16,mask,cnum//2,3,1,name='conv17')
x18=standard_conv(x17,mask,3,3,1,name='conv18')
x18=tf.clip_by_value(x18,-1.,1.)
return x18
def GPNet(x,mask,padding='SAME',name='inpaint_net_1',reuse=False):
'''
Global perceiving network
Args:
x: incomplete image
mask: mask region {0,1}
returns:
image predicted by global perceiving network
'''
xin=x;
mask_in=mask
cnum=32
x = tf.concat([x, mask], axis=3) #concat
with tf.variable_scope(name,reuse=reuse):
x1=standard_conv(x,mask,cnum,5,1,name='conv1')
x2=standard_conv(x1,mask,2*cnum,3,2,name='conv2_downsample')
x3=standard_conv(x2,mask,2*cnum,3,1,name='conv3')
x4=standard_conv(x3,mask,4*cnum,3,2,name='conv4_downsample')
x5=standard_conv(x4,mask,4*cnum,3,1,name='conv5')
x6=standard_conv(x5,mask,4*cnum,3,1,name='conv6')
#dilated conv
x7=standard_conv(x6,mask,4*cnum,3,rate=2,name='conv7_atrous')
x8=standard_conv(x7,mask,4*cnum,3,rate=4,name='conv8_atrous')
x9=standard_conv(x8,mask,4*cnum,3,rate=8,name='conv9_atrous')
x10=standard_conv(x9,mask,4*cnum,3,rate=16,name='conv10_atrous')
x11=standard_conv(tf.concat([x10,x6],axis=-1), mask,4*cnum,3,1,name='conv11')
x12=standard_conv(tf.concat([x11,x5],axis=-1),mask,4*cnum,3,1,name='conv12')
x13=standard_dconv(tf.concat([x12,x4],axis=-1),mask,2*cnum,name='conv13_upsample')
x14=standard_conv(tf.concat([x13,x3],axis=-1),mask,2*cnum,3,1,name='conv14')
x15=standard_dconv(tf.concat([x14,x2],axis=-1),mask,cnum,name='conv15_upsample')
x16=standard_conv(tf.concat([x15,x1],axis=-1),mask,cnum//2,3,1,name='conv16')
x17=standard_conv(x16,mask,3,3,1,name='conv17')
x18=tf.clip_by_value(x17,-1.,1.)
return x18
def RW_discriminator(x, mask, batch_size, activation = 'leaky_relu',reuse=False):
'''
Region-wise discriminator
Args:
x: input images
mask: mask region {0,1}
returns:
matrix {real, fake}
'''
with tf.variable_scope('discriminator',reuse=reuse):
cnum=64
x=tf.concat([x,mask],axis=3)
x=dis_conv(x,cnum,name='d_conv1')
x=dis_conv(x,2*cnum,name='d_conv2')
x=dis_conv(x,4*cnum,name='d_conv3')
x=dis_conv(x,4*cnum,name='d_conv4')
x=dis_conv(x,4*cnum,name='d_conv5')
x=dis_conv(x,4*cnum,name='d_conv6', activation = activation)
return x
def build_graph_with_loss(batch_data, batch_size, mask, vgg_path, adv_type, stage = 0,
lambda_style = 0.001, lambda_cor = 0.00001, alpha = 0.01 ,lambda_adv = 1.0,
reuse=False, training = True):
image_gt=tf.subtract(tf.divide(batch_data,127.5),1.)
date_shape=batch_data.get_shape().as_list()
batch_incomplete=image_gt*mask
image_p1, image_p2 = RW_generator(batch_incomplete, mask)
image_c1 = image_p1 * (1 - mask) + image_gt * mask
image_c2 = image_p2 * (1 - mask) + image_gt * mask
rec_loss = tf.reduce_sum(tf.abs(image_gt - image_p1)) + tf.reduce_sum(tf.abs(image_gt - image_p2))
vgg = Vgg16(vgg_path)
vgg.build(image_gt)
vgg_pos = [vgg.pool1,vgg.pool2, vgg.pool3]
vgg.build(image_c1)
vgg_x1 = [vgg.pool1,vgg.pool2, vgg.pool3]
vgg.build(image_c2)
vgg_x2 = [vgg.pool1,vgg.pool2, vgg.pool3]
cor_loss = loss_cor(vgg_x1, vgg_pos)
style_loss= loss_style(vgg_x2, vgg_pos)
cor_loss = cor_loss * lambda_cor
style_loss = style_loss * lambda_style
cor_style = cor_loss + style_loss
if stage == 1:
activation = None
if adv_type == ' ':
activation = 'leaky_relu'
d_pred = RW_discriminator(image_c1 * (1 - mask), mask, batch_size, activation)
d_pred2 = RW_discriminator(image_c2 * (1 - mask), mask, batch_size, activation, reuse = True)
d_real = RW_discriminator(image_gt * (1 - mask), mask, batch_size, activation, reuse = True)
mask_label = 1 - mask
shape = d_pred.get_shape().as_list()
mask_label = tf.image.resize_nearest_neighbor(mask_label, [shape[1],shape[2]])
if adv_type == 'wgan_gp':
penalty_img = random_interpolates(image_gt, image_c2)
dout_penalty = RW_discriminator(penalty_img, mask, batch_size, activation,reuse = True)
penalty_loss = gradients_penalty(penalty_img, dout_penalty, mask = mask)
d_loss = tf.reduce_mean(d_pred * mask_label) + tf.reduce_mean(d_pred2 * mask_label) - 0.01 * tf.reduce_mean(d_real * mask_label) +penalty_loss
d_g_loss = -1 * tf.reduce_mean(d_pred * mask_label) - tf.reduce_mean(d_pred2 * mask_label)
elif adv_type == 'gan':
adv_d_loss = 0.01 * tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=d_real, labels=tf.ones_like(d_real)) * mask_label)
+ tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=d_pred, labels=tf.zeros_like(d_pred)) * mask_label)
+ tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=d_pred2, labels=tf.zeros_like(d_pred2))* mask_label)
adv_g_loss = tf.reduce_mean( tf.nn.sigmoid_cross_entropy_with_logits(logits=d_pred, labels=tf.ones_like(d_real)) * mask_label) + tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=d_pred2, labels=tf.ones_like(d_real)) * mask_label)
elif adv_type == 'hinge':
adv_d_loss = 0.01 * tf.reduce_mean(tf.nn.relu(1 - d_real) * mask_label) + tf.reduce_mean(tf.nn.relu(1 + d_pred) * mask_label)
+ tf.reduce_mean(tf.nn.relu(1 + d_pred2) * mask_label)
adv_g_loss = -1 * tf.reduce_mean(tf.nn.relu(d_pred) * mask_label) - tf.reduce_mean(tf.nn.relu(d_pred2) * mask_label)
else:
adv_d_loss = alpha*tf.reduce_sum(tf.abs(mask_label - d_real)) + tf.reduce_sum(tf.abs(0 - d_pred)) + tf.reduce_sum(tf.abs(0 - d_pred2))
adv_g_loss = tf.reduce_sum(tf.abs(mask_label - d_pred)) + tf.reduce_sum(tf.abs(mask_label - d_pred2))
else:
adv_d_loss = None
adv_g_loss = None
if stage == 1:
g_loss= rec_loss+cor_style + lambda_adv * adv_g_loss
else:
g_loss = rec_loss + cor_style
g_vars=tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,'inpaint_net')
d_vars=tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,'discriminator') if stage == 1 else None
tf.summary.image('incomplete',batch_incomplete, max_outputs=7)
tf.summary.image('image_p1',image_p1, max_outputs=7)
tf.summary.image('image_p2',image_p2, max_outputs=7)
tf.summary.image('image_c2',image_c2, max_outputs=7)
tf.summary.scalar('rec_loss',rec_loss)
tf.summary.scalar('correlation loss', cor_loss)
tf.summary.scalar('style loss', style_loss)
if stage == 1:
tf.summary.scalar('adv_g_loss', adv_g_loss)
tf.summary.scalar('adv_d_loss', adv_d_loss)
return g_vars,d_vars,g_loss,adv_d_loss,rec_loss,cor_loss,style_loss