forked from run-youngjoo/SC-FEGAN
-
Notifications
You must be signed in to change notification settings - Fork 0
/
model.py
140 lines (121 loc) · 6.21 KB
/
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
import tensorflow as tf
import tensorflow.contrib.slim.nets as nets
import numpy as np
from ops import *
class Model(object):
def __init__(self, config):
self.input_size = config.INPUT_SIZE
self.batch_size = config.BATCH_SIZE
image_dims = [self.input_size, self.input_size, 3]
sk_dims = [self.input_size, self.input_size, 1]
color_dims = [self.input_size, self.input_size, 3]
masks_dims = [self.input_size, self.input_size, 1]
noises_dims = [self.input_size, self.input_size, 1]
self.dtype = tf.float32
self.images = tf.placeholder(
self.dtype, [self.batch_size] + image_dims, name='real_images')
self.sketches = tf.placeholder(
self.dtype, [self.batch_size] + sk_dims, name='sketches')
self.color = tf.placeholder(
self.dtype, [self.batch_size] + color_dims, name='color')
self.masks = tf.placeholder(
self.dtype, [self.batch_size] + masks_dims, name='masks')
self.noises = tf.placeholder(
self.dtype, [self.batch_size] + noises_dims, name='noises')
def build_gen(self, x, mask, name='generator',reuse=False, trainig=True):
cnum = 64
s_h, s_w = self.input_size, self.input_size
s_h2, s_w2 = int(self.input_size/2), int(self.input_size/2)
s_h4, s_w4 = int(self.input_size/4), int(self.input_size/4)
s_h8, s_w8 = int(self.input_size/8), int(self.input_size/8)
s_h16, s_w16 = int(self.input_size/16), int(self.input_size/16)
s_h32, s_w32 = int(self.input_size/32), int(self.input_size/32)
s_h64, s_w64 = int(self.input_size/64), int(self.input_size/64)
with tf.variable_scope(name, reuse=reuse):
# encoder
x_now = x
x1, mask1 = gate_conv(x,cnum,7,2,use_lrn=False,name='gconv1_ds')
x2, mask2 = gate_conv(x1,2*cnum,5,2,name='gconv2_ds')
x3, mask3 = gate_conv(x2,4*cnum,5,2,name='gconv3_ds')
x4, mask4 = gate_conv(x3,8*cnum,3,2,name='gconv4_ds')
x5, mask5 = gate_conv(x4,8*cnum,3,2,name='gconv5_ds')
x6, mask6 = gate_conv(x5,8*cnum,3,2,name='gconv6_ds')
x7, mask7 = gate_conv(x6,8*cnum,3,2,name='gconv7_ds')
# dilated conv
x7,_ = gate_conv(x7, 8*cnum, 3, 1, rate=2, name='co_conv1_dlt')
x7,_ = gate_conv(x7, 8*cnum, 3, 1, rate=4, name='co_conv2_dlt')
x7,_ = gate_conv(x7, 8*cnum, 3, 1, rate=8, name='co_conv3_dlt')
x7,_ = gate_conv(x7, 8*cnum, 3, 1, rate=16, name='co_conv4_dlt')
# decoder
x8, _ = gate_deconv(x7,[self.batch_size, s_h64, s_w64, 8*cnum], name='deconv1')
x8 = tf.concat([x6,x8],axis=3)
x8, mask8 = gate_conv(x8,8*cnum,3,1,name='gconv8')
x9, _ = gate_deconv(x8,[self.batch_size, s_h32, s_w32, 8*cnum], name='deconv2')
x9 = tf.concat([x5,x9],axis=3)
x9, mask9 = gate_conv(x9,8*cnum,3,1,name='gconv9')
x10, _ = gate_deconv(x9,[self.batch_size, s_h16, s_w16, 8*cnum], name='deconv3')
x10 = tf.concat([x4,x10],axis=3)
x10, mask10 = gate_conv(x10,8*cnum,3,1,name='gconv10')
x11, _ = gate_deconv(x10,[self.batch_size, s_h8, s_w8, 4*cnum], name='deconv4')
x11 = tf.concat([x3,x11],axis=3)
x11, mask11 = gate_conv(x11,4*cnum,3,1,name='gconv11')
x12, _ = gate_deconv(x11,[self.batch_size, s_h4, s_w4, 2*cnum], name='deconv5')
x12 = tf.concat([x2,x12],axis=3)
x12, mask12 = gate_conv(x12,2*cnum,3,1,name='gconv12')
x13, _ = gate_deconv(x12,[self.batch_size, s_h2, s_w2, cnum], name='deconv6')
x13 = tf.concat([x1,x13],axis=3)
x13, mask13 = gate_conv(x13,cnum,3,1,name='gconv13')
x14, _ = gate_deconv(x13,[self.batch_size, s_h, s_w, 3], name='deconv7')
x14 = tf.concat([x_now,x14],axis=3)
x14, mask14 = gate_conv(x14,3,3,1,activation=None,use_lrn=False,name='gconv14')
output = tf.tanh(x14)
return output, mask14
def build_demo_graph(self, config):
incom_imgs = self.images*(1-self.masks)
batch_data = tf.concat([incom_imgs,self.sketches,\
self.color,self.masks,self.noises],axis=3)
gen_img, output_mask = self.build_gen(batch_data,self.masks)
self.demo_output= gen_img*self.masks + incom_imgs
def load_demo_graph(self, config):
sess_config = tf.ConfigProto()
sess_config.gpu_options.allow_growth = True
self.sess = tf.Session(config=sess_config)
self.build_demo_graph(config)
init_op = tf.global_variables_initializer()
self.sess.run(init_op)
vars_list = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES)
assign_ops = []
ckpt_path = config.CKPT_DIR
if ckpt_path:
print('Model loaded from {}....start'.format(ckpt_path))
for var in vars_list:
vname = var.name
from_name = vname
# print(from_name)
var_value = tf.contrib.framework.load_variable(ckpt_path, from_name)
# self.sess.run(tf.assign(var, var_value))
assign_ops.append(tf.assign(var, var_value))
self.sess.run(assign_ops)
self.warmup(config)
print('Model loaded from {}....end'.format(ckpt_path))
else:
print('Model loading is fail')
def warmup(self,config):
size = config.INPUT_SIZE
bc = config.BATCH_SIZE
_ = self.sess.run(self.demo_output,
feed_dict={
self.images: np.zeros([bc,size,size,3]),
self.sketches: np.zeros([bc,size,size,1]),
self.color: np.zeros([bc,size,size,3]),
self.masks: np.zeros([bc,size,size,1]),
self.noises: np.zeros([bc,size,size,1])})
def demo(self, config, batch):
demo_output = self.sess.run(self.demo_output,
feed_dict={
self.images: batch[:,:,:,:3],
self.sketches: batch[:,:,:,3:4],
self.color: batch[:,:,:,4:7],
self.masks: batch[:,:,:,7:8],
self.noises: batch[:,:,:,8:9]})
return demo_output