-
Notifications
You must be signed in to change notification settings - Fork 1
/
model_old.py
156 lines (106 loc) · 6.03 KB
/
model_old.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
import tensorflow as tf
import numpy as np
# FLAG: use Instance normalization instead of batch normalization
def instance_norm(x):
epsilon = 1e-9
mean, var = tf.nn.moments(x, [1, 2], keep_dims=True)
return tf.div(x - mean, tf.sqrt(tf.add(var, epsilon)))
initializer = tf.contrib.layers.xavier_initializer()
def weight_variable(shape):
shape = [int(element) for element in shape]
return tf.Variable(initializer(shape))
def bias_variable(shape):
shape = [int(element) for element in shape]
initial = tf.constant(0.0, shape=shape)
return tf.Variable(initial)
class ResidualBlock():
"""Residual Block."""
def __init__(self, dim_in=256, dim_out=256):
self.residualBlock1_W = weight_variable([3, 3, dim_in, dim_out])
self.residualBlock2_W = weight_variable([3, 3, dim_out, dim_out])
def forward(self, x):
Y1 = tf.nn.conv2d(x, self.residualBlock1_W, strides=[1, 1, 1, 1], padding='SAME')
Y1_norm = instance_norm(Y1)
Y1_relu = tf.nn.relu(Y1_norm)
Y2 = tf.nn.conv2d(Y1_relu, self.residualBlock2_W, strides=[1, 1, 1, 1], padding='SAME')
Y2_norm = instance_norm(Y2)
return x + Y2_norm
class Generator():
def __init__(self, imgDim=(128,128), numClass=5):
with tf.name_scope("Generator") as self.Gscope:
self.imgDim = imgDim
# Weights initialised
self.downSampling1_W = weight_variable([7, 7, 3 + numClass, 64])
self.downSampling2_W = weight_variable([4, 4, 64, 128])
self.downSampling3_W = weight_variable([4, 4, 128, 256])
self.residualBlock1 = ResidualBlock()
self.residualBlock2 = ResidualBlock()
self.residualBlock3 = ResidualBlock()
self.residualBlock4 = ResidualBlock()
self.residualBlock5 = ResidualBlock()
self.residualBlock6 = ResidualBlock()
self.upSampling1_W = weight_variable([4, 4, 128, 256])
self.upSampling2_W = weight_variable([4, 4, 64, 128])
self.upSampling3_W = weight_variable([7, 7, 64, 3])
def forward(self, X_G):
with tf.name_scope(self.Gscope):
batch_size = tf.shape(X_G)[0]
Y_downSampling1 = tf.nn.relu(tf.nn.conv2d(X_G, self.downSampling1_W, strides=[1, 1, 1, 1], padding='SAME'))
Y_downSampling1_norm = instance_norm(Y_downSampling1)
Y_downSampling2 = tf.nn.relu(tf.nn.conv2d(Y_downSampling1_norm, self.downSampling2_W, strides=[1, 2, 2, 1], padding='SAME'))
Y_downSampling2_norm = instance_norm(Y_downSampling2)
Y_downSampling3 = tf.nn.relu(tf.nn.conv2d(Y_downSampling2_norm, self.downSampling3_W, strides=[1, 2, 2, 1], padding='SAME'))
Y_downSampling3_norm = instance_norm(Y_downSampling3)
Y_residual1 = self.residualBlock1.forward(Y_downSampling3_norm)
Y_residual2 = self.residualBlock2.forward(Y_residual1)
Y_residual3 = self.residualBlock3.forward(Y_residual2)
Y_residual4 = self.residualBlock4.forward(Y_residual3)
Y_residual5 = self.residualBlock5.forward(Y_residual4)
Y_residual6 = self.residualBlock6.forward(Y_residual5)
Y_upSampling1 = tf.nn.relu(tf.nn.conv2d_transpose(Y_residual6, self.upSampling1_W, [batch_size, int(self.imgDim[0]/2), int(self.imgDim[1]/2),128], strides=[1, 2, 2, 1], padding='SAME'))
Y_upSampling1_norm = instance_norm(Y_upSampling1)
Y_upSampling2 = tf.nn.relu(tf.nn.conv2d_transpose(Y_upSampling1_norm, self.upSampling2_W, [batch_size, self.imgDim[0], self.imgDim[1],64], strides=[1, 2, 2, 1], padding='SAME'))
Y_upSampling2_norm = instance_norm(Y_upSampling2)
fakeGeneration = tf.nn.tanh(tf.nn.conv2d(Y_upSampling2_norm, self.upSampling3_W, strides=[1, 1, 1, 1], padding='SAME'))
return fakeGeneration
def recForward(self, fake, trueLabels):
with tf.name_scope(self.Gscope):
fakeWithRealLabels = tf.concat([fake, trueLabels], 3)
return self.forward(fakeWithRealLabels)
class Discriminator():
def __init__(self, imageSize=128, convDim=64, numClass=5):
with tf.name_scope("Discriminator") as self.Dscope:
# Weights initialised
self.imageSize = imageSize
currDim = convDim
self.InputLayer = weight_variable([4, 4, 3, currDim])
self.InputLayer_b = bias_variable([currDim])
self.HiddenLayer1 = weight_variable([4, 4, currDim, currDim*2])
self.HiddenLayer1_b = bias_variable([currDim*2])
currDim = currDim*2
self.HiddenLayer2 = weight_variable([4, 4, currDim, currDim*2])
self.HiddenLayer2_b = bias_variable([currDim*2])
currDim = currDim*2
self.HiddenLayer3 = weight_variable([4, 4, currDim, currDim*2])
self.HiddenLayer3_b = bias_variable([currDim*2])
currDim = currDim*2
self.HiddenLayer4 = weight_variable([4, 4, currDim, currDim*2])
self.HiddenLayer4_b = bias_variable([currDim*2])
currDim = currDim*2
self.HiddenLayer5 = weight_variable([4, 4, currDim, currDim*2])
self.HiddenLayer5_b = bias_variable([currDim*2])
currDim = currDim*2
self.OutputLayerSrc = weight_variable([3, 3, currDim, 1])
self.OutputLayerCls = weight_variable([imageSize/64, imageSize/64, currDim, numClass])
def forward(self, x):
with tf.name_scope(self.Dscope):
# x has to be a placeHolder or conexion with generator output
Y_input = tf.nn.leaky_relu(tf.nn.conv2d(x, self.InputLayer, strides=[1, 2, 2, 1], padding='SAME') + self.InputLayer_b)
Y_hiddenLayer1 = tf.nn.leaky_relu(tf.nn.conv2d(Y_input, self.HiddenLayer1, strides=[1, 2, 2, 1], padding='SAME') + self.HiddenLayer1_b)
Y_hiddenLayer2 = tf.nn.leaky_relu(tf.nn.conv2d(Y_hiddenLayer1, self.HiddenLayer2, strides=[1, 2, 2, 1], padding='SAME') + self.HiddenLayer2_b)
Y_hiddenLayer3 = tf.nn.leaky_relu(tf.nn.conv2d(Y_hiddenLayer2, self.HiddenLayer3, strides=[1, 2, 2, 1], padding='SAME') + self.HiddenLayer3_b)
Y_hiddenLayer4 = tf.nn.leaky_relu(tf.nn.conv2d(Y_hiddenLayer3, self.HiddenLayer4, strides=[1, 2, 2, 1], padding='SAME') + self.HiddenLayer4_b)
Y_hiddenLayer5 = tf.nn.leaky_relu(tf.nn.conv2d(Y_hiddenLayer4, self.HiddenLayer5, strides=[1, 2, 2, 1], padding='SAME') + self.HiddenLayer5_b)
Y_outputLayerSrc = tf.nn.conv2d(Y_hiddenLayer5, self.OutputLayerSrc, strides=[1, 1, 1, 1],padding='SAME')
Y_outputLayerCls = tf.nn.conv2d(Y_hiddenLayer5, self.OutputLayerCls, strides=[1, 1, 1, 1], padding='VALID') # TODO padding in YCls should BE "TYPE1"
return Y_outputLayerSrc, Y_outputLayerCls