-
Notifications
You must be signed in to change notification settings - Fork 0
/
net.py
177 lines (137 loc) · 6.73 KB
/
net.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
import os, sys
#os.environ['CUDA_VISIBLE_DEVICES'] = '-1'
import numpy as np
import cv2 as cv
import time
import tensorflow as tf
assert tf.__version__.startswith('2.'), 'This tutorial assumes Tensorflow 2.0+'
from tensorflow.keras import layers
from tensorflow.keras.applications import VGG16
from tensorflow.keras.optimizers import Adam
from utils import InstanceNormalization
import argparse
h,w = 256, 256
inshape = (h, w, 3)
def Conv2DNormLReLU(x, k_num, k_size, padding_type, name_id):
x = layers.Conv2D(k_num, k_size, strides=1, padding=padding_type,use_bias=None,kernel_initializer='he_normal',
activation=None, name=name_id, trainable=True)(x)
x = InstanceNormalization()(x)
x = layers.LeakyReLU(alpha=0.2)(x)
return x
def SeparableConv2D(x, k_num, k_size, strides,padding_type, name_id):
x = layers.SeparableConv2D(k_num,k_size,strides=strides,padding=padding_type,dilation_rate=(1, 1),depth_multiplier=1,name=name_id)(x)
x = InstanceNormalization()(x)
x = layers.LeakyReLU(alpha=0.2)(x)
return x
def InvertedRes_block(x, k_num, k_size, padding_type, name_id):
x1 = Conv2DNormLReLU(x, k_num*2, 3, "same", None)
x2 = layers.DepthwiseConv2D(k_size,strides=(1, 1),padding="same")(x1)
x2 = InstanceNormalization()(x2)
x2 = layers.LeakyReLU(alpha=0.2)(x2)
x3 = layers.Conv2D(k_num, k_size, strides=1, padding= "same",use_bias=None,kernel_initializer='he_normal')(x2)
x3 = InstanceNormalization()(x3)
y = layers.Add(name = name_id)([x,x3])
return y
def vgg_net(inshape):
rawvgg=VGG16(weights='imagenet',include_top=False,input_shape=inshape)
vgg_model=tf.keras.models.Model(inputs=rawvgg.input,outputs=rawvgg.layers[13].output)
return vgg_model
def G_net(inshape):
inputlayer = layers.Input(shape=inshape)
#block 1
conv1 = Conv2DNormLReLU(inputlayer, 64, 3, "same", "conv1")
conv2 = Conv2DNormLReLU(conv1, 64, 3, "same", "conv2")
downcon1 = SeparableConv2D(conv2, 128, 3, 2,"same", "downcon1")
#block 2
conv3 = Conv2DNormLReLU(downcon1, 128, 3, "same", "conv3")
conv4 = SeparableConv2D(conv3, 128, 3, 1,"same", "conv4")
downcon2 = SeparableConv2D(conv4, 256, 3, 2,"same", "downcon2")
#core
conv5 = Conv2DNormLReLU(downcon2, 256, 3, "same", "conv5")
irb1 = InvertedRes_block(conv5, 256, 3, "same", "irb1")
irb2 = InvertedRes_block(irb1, 256, 3, "same", "irb2")
irb3 = InvertedRes_block(irb2, 256, 3, "same", "irb3")
irb4 = InvertedRes_block(irb3, 256, 3, "same", "irb4")
donv5 = Conv2DNormLReLU(irb4, 256, 3, "same", "donv5")
#up_block2
upcon2 = layers.UpSampling2D(size=2, interpolation="nearest", name='upcon2')(donv5)
donv4 = SeparableConv2D(upcon2, 128, 3, 1,"same", "donv4")
donv3 = Conv2DNormLReLU(donv4, 128, 3, "same", "donv3")
#up_block1
upcon1 = layers.UpSampling2D(size=2, interpolation="nearest", name='upcon1')(donv3)
donv2_0 = SeparableConv2D(upcon1, 128, 3, 1,"same", "donv2_0")
donv2 = Conv2DNormLReLU(upcon1, 64, 3, "same", "donv2")
donv1 = Conv2DNormLReLU(donv2, 64, 3, "same", "donv1")
output = layers.Conv2D(3, 3, strides=1, padding= "same",activation='tanh',kernel_initializer='he_normal',name='output')(donv1)
return tf.keras.models.Model(inputs=inputlayer, outputs=output)
def inited_G_net(inshape):
G_net_model = G_net(inshape)#load the gen_net
real_img = G_net_model.input#get the input_img
gen_img = G_net_model.output#layers[75] is the G_net's output
vgg_model = vgg_net(inshape)#layers[78] is the vggmodel
#get the vgg_maps
real_vgg_map = vgg_model(real_img)
gen_vgg_map = vgg_model(gen_img)
output=layers.concatenate([real_vgg_map, gen_vgg_map], axis=-1)
return tf.keras.models.Model(inputs=real_img, outputs=output)
def inited_G_loss(_, y_pred):
real_vgg_map = y_pred[:,:,:,0:512]
gen_vgg_map = y_pred[:,:,:,512:]
mae = tf.keras.losses.MeanAbsoluteError()
loss = mae(real_vgg_map, gen_vgg_map)
return loss
def gram(x):
shape_x = tf.shape(x)
b = shape_x[0]
c = shape_x[3]
x = tf.reshape(x, [b, -1, c])
return tf.matmul(tf.transpose(x, [0, 2, 1]), x) / tf.cast((tf.size(x) // b), tf.float32)
def style_loss(_, y_pred):
real_vgg_map = y_pred[:,:,:,:512]
gen_vgg_map = y_pred[:,:,:,512:1024]
gray_vgg_map = y_pred[:,:,:,1024:]
#
mae = tf.keras.losses.MeanAbsoluteError()
c_loss = mae(real_vgg_map, gen_vgg_map)
s_loss = mae(gram(gray_vgg_map), gram(gen_vgg_map))
return 1.5*c_loss + 2.8*s_loss
def color_loss(_, y_pred):
real_img = y_pred[:,:,:,:3]
gen_img = y_pred[:,:,:,3:6]
smo_img = y_pred[:,:,:,6:]
mae = tf.keras.losses.MeanAbsoluteError()
hub = tf.keras.losses.Huber(delta=1.0, reduction="auto", name="huber_loss")
#
yuv_real = tf.image.rgb_to_yuv((real_img + 1.0)/2.0)
yuv_gen = tf.image.rgb_to_yuv((gen_img + 1.0)/2.0)
color_loss = mae(yuv_real[:,:,:,0], yuv_gen[:,:,:,0]) + hub(yuv_real[:,:,:,1],yuv_gen[:,:,:,1]) + hub(yuv_real[:,:,:,2],yuv_gen[:,:,:,2])
return 10*color_loss
def D_net(inshape):
inputlayer = layers.Input(shape=inshape)
#block 1
conv1 = layers.Conv2D(32, 3, strides=1, padding= "same",use_bias=None,kernel_initializer='he_normal')(inputlayer)
conv1 = layers.LeakyReLU(alpha=0.2)(conv1)
#block 2
conv2 = layers.Conv2D(64, 3, strides=2, padding= "same",use_bias=None,kernel_initializer='he_normal')(conv1)
conv2 = layers.LeakyReLU(alpha=0.2)(conv2)
conv3 = Conv2DNormLReLU(conv2, 128, 3, "same", "desr_conv3")
#block 3
conv4 = layers.Conv2D(128, 3, strides=2, padding= "same",use_bias=None,kernel_initializer='he_normal')(conv3)
conv4 = layers.LeakyReLU(alpha=0.2)(conv4)
conv5 = Conv2DNormLReLU(conv4, 256, 3, "same", "desr_conv5")
#last block
conv6 = Conv2DNormLReLU(conv5, 256, 3, "same", "desr_conv6")
output = layers.Conv2D(1, 3, strides=1, padding= "same",activation=None,kernel_initializer='he_normal',name='output')(conv6)
return tf.keras.models.Model(inputs=inputlayer, outputs=output)
#Dnet
def train_D_net(inshape):
D_net_model = D_net(inshape)#load the des_net
anie_img = layers.Input(shape=inshape)
anie_gray = layers.Input(shape=inshape)
gen_img = layers.Input(shape=inshape)
anie_smooth = layers.Input(shape=inshape)
anie_logit = D_net_model(anie_img)
anie_gray_logit = D_net_model(anie_gray)
gen_img_logit = D_net_model(gen_img)
anie_smooth_logit = D_net_model(anie_smooth)
return tf.keras.models.Model(inputs=[anie_img,anie_gray,anie_smooth,gen_img], outputs=[anie_logit,anie_gray_logit,anie_smooth_logit,gen_img_logit])