-
Notifications
You must be signed in to change notification settings - Fork 2
/
RaSGAN_GP.py
183 lines (150 loc) · 8.2 KB
/
RaSGAN_GP.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
import tensorflow as tf
from models.generative.ops import *
from models.generative.utils import *
from models.generative.loss import *
from models.generative.optimizer import *
from models.generative.gans.GAN import GAN
class RaSGAN_GP(GAN):
def __init__(self,
data, # Dataset class, training and test data.
z_dim, # Latent space dimensions.
use_bn, # Batch Normalization flag to control usage in discriminator.
alpha, # Alpha value for LeakyReLU.
beta_1, # Beta 1 value for Adam Optimizer.
beta_2, # Beta 2 value for Adam Optimizer.
n_critic, # Number of batch gradient iterations in Discriminator per Generator.
gp_coeff, # Gradient Penalty coefficient for the Wasserstein Gradient Penalty loss.
learning_rate_g, # Learning rate of the Generator.
learning_rate_d, # Learning rate of the Discriminator.
loss_type='relativistic gradient penalty', # Loss function type: Standard, Least Square, Wasserstein, Wasserstein Gradient Penalty.
model_name='RaSGAN_GP' # Name to give to the model.
):
# Training parameters
self.gp_coeff = gp_coeff
self.beta_2 = beta_2
super().__init__(data=data, z_dim=z_dim, use_bn=use_bn, alpha=alpha, beta_1=beta_1, learning_rate_g=learning_rate_g, learning_rate_d=learning_rate_d,
n_critic=n_critic, loss_type=loss_type, model_name=model_name)
def discriminator(self, images, reuse):
with tf.variable_scope('discriminator', reuse=reuse):
# Padding = 'Same' -> H_new = H_old // Stride
# Input Shape = (None, 448, 448, 3)
# Conv.
# net = convolutional(inputs=images, output_channels=16, filter_size=5, stride=2, padding='SAME', conv_type='convolutional')
# net = tf.layers.conv2d(inputs=images, filters=16, kernel_size=(5,5), strides=(2, 2), padding='same', kernel_initializer=tf.contrib.layers.xavier_initializer())
# net = leakyReLU(net, self.alpha)
# Shape = (None, 224, 224, 16)
# Conv.
net = convolutional(inputs=images, output_channels=32, filter_size=5, stride=2, padding='SAME', conv_type='convolutional', scope=1)
if self.use_bn: net = tf.layers.batch_normalization(inputs=net, training=True)
net = leakyReLU(net, self.alpha)
# Shape = (None, 112, 112, 32)
# Conv.
net = convolutional(inputs=net, output_channels=64, filter_size=5, stride=2, padding='SAME', conv_type='convolutional', scope=2)
if self.use_bn: net = tf.layers.batch_normalization(inputs=net, training=True)
net = leakyReLU(net, self.alpha)
# Shape = (None, 56, 56, 64)
# Conv.
net = convolutional(inputs=net, output_channels=128, filter_size=5, stride=2, padding='SAME', conv_type='convolutional', scope=3)
if self.use_bn: net = tf.layers.batch_normalization(inputs=net, training=True)
net = leakyReLU(net, self.alpha)
# Shape = (None, 28, 28, 128)
# Conv.
net = convolutional(inputs=net, output_channels=256, filter_size=5, stride=2, padding='SAME', conv_type='convolutional', scope=4)
if self.use_bn: net = tf.layers.batch_normalization(inputs=net, training=True)
net = leakyReLU(net, self.alpha)
# Shape = (None, 14, 14, 256)
# Conv.
net = convolutional(inputs=net, output_channels=512, filter_size=5, stride=2, padding='SAME', conv_type='convolutional', scope=5)
if self.use_bn: net = tf.layers.batch_normalization(inputs=net, training=True)
net = leakyReLU(net, self.alpha)
# Shape = (None, 7, 7, 512)
# Flatten.
net = tf.layers.flatten(inputs=net)
# Shape = (None, 7*7*512)
# Dense.
net = dense(inputs=net, out_dim=1024, scope=1)
if self.use_bn: net = tf.layers.batch_normalization(inputs=net, training=True)
net = leakyReLU(net, self.alpha)
# Shape = (None, 1024)
# Dense
logits = dense(inputs=net, out_dim=1, scope=2)
# Shape = (None, 1)
output = tf.nn.sigmoid(x=logits)
# Padding = 'Same' -> H_new = H_old // Stride
return output, logits
def generator(self, z_input, reuse, is_train):
with tf.variable_scope('generator', reuse=reuse):
# Doesn't work ReLU, tried.
# Input Shape = (None, 100)
# Dense.
net = dense(inputs=z_input, out_dim=1024, scope=1)
net = tf.layers.batch_normalization(inputs=net, training=is_train)
net = leakyReLU(net, self.alpha)
# Shape = (None, 1024)
# Dense.
net = dense(inputs=net, out_dim=256*7*7, scope=2)
net = tf.layers.batch_normalization(inputs=net, training=is_train)
net = leakyReLU(net, self.alpha)
# Shape = (None, 256*7*7)
# Reshape
net = tf.reshape(tensor=net, shape=(-1, 7, 7, 256), name='reshape')
# Shape = (None, 7, 7, 256)
# Conv.
net = convolutional(inputs=net, output_channels=256, filter_size=2, stride=2, padding='SAME', conv_type='transpose', scope=1)
net = tf.layers.batch_normalization(inputs=net, training=is_train)
net = leakyReLU(net, self.alpha)
# Shape = (None, 14, 14, 256)
# Conv.
net = convolutional(inputs=net, output_channels=128, filter_size=5, stride=1, padding='SAME', conv_type='convolutional', scope=2)
net = tf.layers.batch_normalization(inputs=net, training=is_train)
net = leakyReLU(net, self.alpha)
# Shape = (None, 14, 14, 128)
# Conv.
net = convolutional(inputs=net, output_channels=128, filter_size=2, stride=2, padding='SAME', conv_type='transpose', scope=3)
net = tf.layers.batch_normalization(inputs=net, training=is_train)
net = leakyReLU(net, self.alpha)
# Shape = (None, 28, 28, 128)
# Conv.
net = convolutional(inputs=net, output_channels=64, filter_size=5, stride=1, padding='SAME', conv_type='convolutional', scope=4)
net = tf.layers.batch_normalization(inputs=net, training=is_train)
net = leakyReLU(net, self.alpha)
# Shape = (None, 28, 28, 64)
# Conv.
net = convolutional(inputs=net, output_channels=64, filter_size=2, stride=2, padding='SAME', conv_type='transpose', scope=5)
net = tf.layers.batch_normalization(inputs=net, training=is_train)
net = leakyReLU(net, self.alpha)
# Shape = (None, 56, 56, 64)
# Conv.
net = convolutional(inputs=net, output_channels=32, filter_size=5, stride=1, padding='SAME', conv_type='convolutional', scope=6)
net = tf.layers.batch_normalization(inputs=net, training=is_train)
net = leakyReLU(net, self.alpha)
# Shape = (None, 56, 56, 32)
# Conv.
net = convolutional(inputs=net, output_channels=32, filter_size=2, stride=2, padding='SAME', conv_type='transpose', scope=7)
net = tf.layers.batch_normalization(inputs=net, training=is_train)
net = leakyReLU(net, self.alpha)
# Shape = (None, 112, 112, 32)
# Conv.
# net = convolutional(inputs=net, output_channels=16, filter_size=5, stride=1, padding='SAME', conv_type='transpose')
# net = tf.layers.batch_normalization(inputs=net, training=is_train)
# net = leakyReLU(net, self.alpha)
# Shape = (None, 112, 112, 16)
# Conv.
# net = convolutional(inputs=net, output_channels=16, filter_size=2, stride=2, padding='SAME', conv_type='upscale')
# net = tf.layers.conv2d_transpose(inputs=net, filters=16, kernel_size=(2,2), strides=(2,2), padding='same', kernel_initializer=tf.contrib.layers.xavier_initializer())
# net = tf.layers.batch_normalization(inputs=net, training=is_train)
# net = leakyReLU(net, self.alpha)
# Shape = (None, 224, 224, 16)
# Conv.
logits = convolutional(inputs=net, output_channels=self.image_channels, filter_size=2, stride=2, padding='SAME', conv_type='transpose', scope=8)
# Shape = (None, 448, 448, 3)
output = tf.nn.sigmoid(x=logits, name='output')
return output
def loss(self):
loss_dis, loss_gen = losses(self.loss_type, self.output_fake, self.output_real, self.logits_fake, self.logits_real, real_images=self.real_images,
fake_images=self.fake_images, discriminator=self.discriminator, gp_coeff=self.gp_coeff)
return loss_dis, loss_gen
def optimization(self):
train_discriminator, train_generator = optimizer(self.beta_1, self.loss_gen, self.loss_dis, self.loss_type, self.learning_rate_input_g, self.learning_rate_input_d,
beta_2=self.beta_2)
return train_discriminator, train_generator