-
Notifications
You must be signed in to change notification settings - Fork 5
/
model.py
139 lines (109 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 numpy as np
import tensorlayer as tl
from tensorlayer.layers import Input,Dense,Dropout,Conv2d,DeConv2d,BatchNorm2d,InstanceNorm2d,flatten_reshape,MeanPool2d
from tensorlayer.act import lrelu,tanh
from tensorlayer.models import Model
class BicycleGAN(object):
model_name='BicycleGAN'
def __init__(self,sess,epoch,batch_size,Z_dim,image_size,dataset_name,checkpoint_dir,result_dir,log_dir):
self.sess=sess
self.epoch=epoch
self.batch_size=batch_size
self.image_size=image_size
self.Z_dim=Z_dim
self.dataset_name=dataset_name
self.checkpoint_dir=checkpoint_dir
self.result_dir=result_dir
self.log_dir=log_dir
if dataset_name=='cityscapes'
self.input_width=256
self.input_height=256
self.output_width=256
self.output_height=256
self.channels=3
self.learning_rate=0.0002
self.beta1=0.5
self.beta2=0.999
self.reconst_C=10
self.latent_C=0.5
self.kl_C=0.01
self.sample_num=64
self.train_A=glob.glob('city/scapes/train/*.jpg')
self.num_batches=len(self.train_A)
def residual(R,n_f,f_s):
R_tmp=R
R=BatchNorm2d(act=lrelu)(Conv2d(n_f,(f_s,f_s),(2,2))(R))
R=BatchNorm2d(act=None)(Conv2d(n_f,(3,3),(1,1))(R))
R_tmp=Conv2d(n_f,(1,1),(1,1))(R_tmp)
return tf.nn.relu(R_tmp+R)
def Discriminator(D):
D=Conv2d(64,(4,4),(2,2),padding='SAME',act=lrelu,name='D_conv_1')(D)
D=InstanceNorm2d(act=lrelu)(Conv2d(128,(4,4),(2,2),padding='SAME',name='D_conv_2')(D))
D=InstanceNorm2d(act=lrelu)(Conv2d(256,(4,4),(2,2),padding='SAME',name='D_conv_3')(D))
D=InstanceNorm2d(act=lrelu)(Conv2d(512,(4,4),(2,2),padding='SAME',name='D_conv_4')(D))
D=InstanceNorm2d(act=lrelu)(Conv2d(512,(4,4),(2,2),padding='SAME',name='D_conv_5')(D))
D=InstanceNorm2d(act=lrelu)(Conv2d(512,(4,4),(2,2),padding='SAME',name='D_conv_6')(D))
D=Conv2d(1,(4,4),(2,2),name='D_conv_7')(D)
D=tf.reduce_mean(D,axis=[1,2,3])
return D
def Generator(self,G,z):
z=tf.reshape(z,[self.batch_size,1,1,self.Z_dim])
z=tf.tile(z,[1,self.image_size,self.image_size,1])
G=tf.concat([G,z],axis=3)
G=Conv2d(64,(4,4),(2,2),padding='SAME',act=lrelu,name='G_conv_1')(I)
G=BatchNorm2d(act=lrelu)(Conv2d(128,(4,4),(2,2),padding='SAME',name='G_conv_2')(G))
G=BatchNorm2d(act=lrelu)(Conv2d(256,(4,4),(2,2),padding='SAME',name='G_conv_3')(G))
G=BatchNorm2d(act=lrelu)(Conv2d(512,(4,4),(2,2),padding='SAME',name='G_conv_4')(G))
G=BatchNorm2d(act=lrelu)(Conv2d(512,(4,4),(2,2),padding='SAME',name='G_conv_5')(G))
G=BatchNorm2d(act=lrelu)(Conv2d(512,(4,4),(2,2),padding='SAME',name='G_conv_6')(G))
G=BatchNorm2d(act=lrelu)(Conv2d(512,(4,4),(2,2),padding='SAME',name='G_conv_7')(G))
G=BatchNorm2d(act=lrelu)(Conv2d(512,(4,4),(2,2),padding='SAME',name='G_conv_8')(G))
G=BatchNorm2d(act=lrelu)(DeConv2d(512,(4,4),(2,2),padding='SAME',name='G_deconv_1')(G))
G=BatchNorm2d(act=lrelu)(DeConv2d(512,(4,4),(2,2),padding='SAME',name='G_deconv_2')(G))
G=BatchNorm2d(act=lrelu)(DeConv2d(512,(4,4),(2,2),padding='SAME',name='G_deconv_3')(G))
G=BatchNorm2d(act=lrelu)(DeConv2d(512,(4,4),(2,2),padding='SAME',name='G_deconv_4')(G))
G=BatchNorm2d(act=lrelu)(DeConv2d(256,(4,4),(2,2),padding='SAME',name='G_deconv_5')(G))
G=BatchNorm2d(act=lrelu)(DeConv2d(128,(4,4),(2,2),padding='SAME',name='G_deconv_6')(G))
G=BatchNorm2d(act=lrelu)(DeConv2d(64,(4,4),(2,2),padding='SAME',name='G_deconv_7')(G))
G=BatchNorm2d(act=tanh)(DeConv2d(3,(4,4),(2,2),padding='SAME',name='G_deconv_8')(G))
return D
def encoder(self,E):
E=Conv2d(64,(4,4),(2,2),act=lrelu,name='E_conv_1')(E)
E=MeanPool2d((2,2),(2,2),'SAME')(self.residual(E,128,3))
E=MeanPool2d((2,2),(2,2),'SAME')(self.residual(E,256,3))
E=MeanPool2d((2,2),(2,2),'SAME')(self.residual(E,512,3))
E=MeanPool2d((2,2),(2,2),'SAME')(self.residual(E,512,3))
E=MeanPool2d((2,2),(2,2),'SAME')(self.residual(E,512,3))
E=MeanPool2d((8,8),(8,8),'SAME')(E)
mu=Dense(self.Z_dim)(E)
log_sigma=Dense(self.Z_dim)(E)
z=mu+tf.random_normal(shape=tf.shape(self.Z_dim))*tf.exp(log_sigma)
return z,mu,log_sigma
def build_model(self):
image_dims=[self.input_width,self.input_height,self.channels]
self.image_A=
self.image_B=
self.z=
self.encoded_true_img,self.encoded_mu,self.encoded_log_sigma=self.Encoder(self.image_B)
self.desired_gen_img=self.Generator(self.image_A,self.encoded_true_img)
self.LR_desired_img=self.Generator(self.image_A,self.z)
self.reconst_z,self.reconst_mu,self.reconst_log_sigma=self.Encoder(self.LR_desired_img)
self.P_real=self.Discriminator(self.image_B)
self.P_fake=self.Discriminator(self.LR_desired_img)
self.P_fake_encoded=self.Discriminator(self.desired_gen_img)
self.loss_vae_D=(tf.reduce_mean(tf.squared_difference(self.P_real,0.9))+tf.reduce_mean(tf.square(self.P_fake_encoded)))
self.loss_lr_D=(tf.reduce_mean(tf.squared_difference(self.P_real,0.9))+tf.reduce_mean(tf.square(self.P_fake)))
self.loss_vae_GE=tf.reduce_mean(tf.squared_difference(self.P_fake_encoded,0.9))
self.loss_G=tf.reduce_mean(tf.squared_difference(self.P_fake,0.9))
self.loss_vae_GE=tf.reduce_mean(tf.abs(self.image_B-self.desired_gen_img))
self.loss_latent_GE=tf.reduce_mean(tf.abs(self.z-self.reconst_z))
self.loss_kl_E=0.5*tf.reduce_mean(-1-self.encoded_log_sigma+self.encoded_mu**2+tf.exp(self.encoded_log_sigma))
self.loss_D=self.loss_vae_gan_D+self.loss_lr_gan_D-tf.reduce_mean(tf.squared_difference(self.P_real, 0.9))
self.loss_G=self.loss_vae_gan_GE+self.loss_gan_G+self.reconst_C*self.loss_vae_GE+self.latent_C*self.loss_latent_GE
self.loss_E=self.loss_vae_gan_GE+self.reconst_C*self.loss_vae_GE+self.latent_C*self.loss_latent_GE+self.kl_C*self.loss_kl_E
def train(self):
def test(self):
def model_dir(self):
def save(self,checkpoint_dir,step):
def load(self,checkpoint_dir):