-
Notifications
You must be signed in to change notification settings - Fork 56
/
Copy pathmodels.py
212 lines (164 loc) · 10.5 KB
/
models.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
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
import tensorflow as tf
import os
import sys
import tensorlayer as tl
import numpy as np
from tensorlayer.layers import *
import random
from glob import glob
def Deconv(inputs, f_dim_in, dim, net, batch_size, f_dim_out = None, stride = 2 ):
if f_dim_out is None:
f_dim_out = f_dim_in/2
return tl.layers.DeConv3dLayer(inputs,
shape = [4, 4, 4, f_dim_out, f_dim_in],
output_shape = [batch_size, dim, dim, dim, f_dim_out],
strides=[1, stride, stride, stride, 1],
W_init = tf.random_normal_initializer(stddev=0.02),
act=tf.identity, name='g/net_' + net + '/deconv')
def Conv3D(inputs, f_dim_out, net, f_dim_in = None, batch_norm = False, is_train = True):
if f_dim_in is None:
f_dim_in = f_dim_out/2
layer = tl.layers.Conv3dLayer(inputs,
shape=[4, 4, 4, f_dim_in, f_dim_out],
W_init = tf.random_normal_initializer(stddev=0.02),
strides=[1, 2, 2, 2, 1], name= 'd/net_' + net + '/conv')
if batch_norm:
return tl.layers.BatchNormLayer(layer, is_train=is_train, name='d/net_' + net + '/batch_norm')
else:
return layer
def generator_20(inputs, is_train=True, reuse=False, batch_size = 128):
output_size, half, forth = 20, 10, 5
gf_dim = 128 # Dimension of gen filters in first conv layer
with tf.variable_scope("gen", reuse=reuse) as vs:
tl.layers.set_name_reuse(reuse)
net_0 = tl.layers.InputLayer(inputs, name='g/net_0/in')
net_1 = tl.layers.DenseLayer(net_0, n_units = gf_dim*forth*forth*forth, W_init = tf.random_normal_initializer(stddev=0.02), act = tf.identity, name='g/net_1/dense')
net_1 = tl.layers.ReshapeLayer(net_1, shape = [-1, forth, forth, forth, gf_dim], name='g/net_1/reshape')
net_1 = tl.layers.BatchNormLayer(net_1, is_train=is_train, gamma_init=tf.random_normal_initializer(1., 0.02), name='g/net_1/batch_norm')
net_1.outputs = tf.nn.relu(net_1.outputs, name='g/net_1/relu')
net_2 = Deconv(net_1, gf_dim, half,'2', batch_size)
net_2 = tl.layers.BatchNormLayer(net_2, is_train=is_train, gamma_init=tf.random_normal_initializer(1., 0.02), name='g/net_2/batch_norm')
net_2.outputs = tf.nn.relu(net_2.outputs, name='g/net_2/relu')
net_3 = Deconv(net_2, gf_dim/2, output_size, '3', batch_size)
net_3 = tl.layers.BatchNormLayer(net_3, is_train=is_train, gamma_init=tf.random_normal_initializer(1., 0.02), name='g/net_3/batch_norm')
net_3.outputs = tf.nn.relu(net_3.outputs, name='g/net_3/relu')
net_4 = Deconv(net_3,gf_dim/4, output_size, '4', batch_size, f_dim_out = 1, stride = 1)
net_4.outputs = tf.reshape(net_4.outputs,[batch_size,output_size,output_size,output_size], name='g/net_4/reshape')
net_4.outputs = tf.nn.tanh(net_4.outputs, name='g/net_4/tanh')
return net_4, net_4.outputs
def generator_32(inputs, is_train=True, reuse=False, batch_size = 128, sig = False):
output_size, half, forth, eighth, sixteenth = 32, 16, 8, 4, 2
gf_dim = 256 # Dimension of gen filters in first conv layer
with tf.variable_scope("gen", reuse=reuse) as vs:
tl.layers.set_name_reuse(reuse)
net_0 = tl.layers.InputLayer(inputs, name='g/net_0/in')
net_1 = tl.layers.DenseLayer(net_0, n_units = gf_dim*sixteenth*sixteenth*sixteenth, W_init = tf.random_normal_initializer(stddev=0.02), act = tf.identity, name='g/net_1/dense')
net_1 = tl.layers.ReshapeLayer(net_1, shape = [-1, sixteenth, sixteenth, sixteenth, gf_dim], name='g/net_1/reshape')
net_1 = tl.layers.BatchNormLayer(net_1, is_train=is_train, gamma_init=tf.random_normal_initializer(1., 0.02), name='g/net_1/batch_norm')
net_1.outputs = tf.nn.relu(net_1.outputs, name='g/net_1/relu')
net_2 = Deconv(net_1, gf_dim, eighth, '2', batch_size)
net_2 = tl.layers.BatchNormLayer(net_2, is_train=is_train, gamma_init=tf.random_normal_initializer(1., 0.02), name='g/net_2/batch_norm')
net_2.outputs = tf.nn.relu(net_2.outputs, name='g/net_2/relu')
net_3 = Deconv(net_2, gf_dim/2, forth, '3', batch_size)
net_3 = tl.layers.BatchNormLayer(net_3, is_train=is_train, gamma_init=tf.random_normal_initializer(1., 0.02), name='g/net_3/batch_norm')
net_3.outputs = tf.nn.relu(net_3.outputs, name='g/net_3/relu')
net_4 = Deconv(net_3, gf_dim/4, half, '4', batch_size)
net_4 = tl.layers.BatchNormLayer(net_4, is_train=is_train, gamma_init=tf.random_normal_initializer(1., 0.02), name='g/net_4/batch_norm')
net_4.outputs = tf.nn.relu(net_4.outputs, name='g/net_4/relu')
net_5 = Deconv(net_4, gf_dim/8, output_size, '5', batch_size, f_dim_out = 1)
net_5.outputs = tf.reshape(net_5.outputs,[batch_size,output_size,output_size,output_size])
if sig:
net_5.outputs = tf.nn.sigmoid(net_5.outputs)
else:
net_5.outputs = tf.nn.tanh(net_5.outputs)
return net_5, net_5.outputs
def discriminator(inputs ,output_size, improved = False, VAE_loss = False, sig = False, is_train=True, reuse=False, batch_size=128, output_units= 1):
inputs = tf.reshape(inputs,[batch_size,output_size,output_size,output_size,1])
df_dim = output_size # Dimension of discrim filters in first conv layer
with tf.variable_scope("dis", reuse=reuse) as vs:
tl.layers.set_name_reuse(reuse)
net_0 = tl.layers.InputLayer(inputs, name='d/net_0/in')
net_1 = Conv3D(net_0, df_dim, '1', f_dim_in = 1 , batch_norm = False )
net_1.outputs = tl.activation.leaky_relu(net_1.outputs, alpha=0.2, name='d/net_1/lrelu')
net_2 = Conv3D(net_1, df_dim*2, '2', batch_norm = not improved, is_train = is_train,)
net_2.outputs = tl.activation.leaky_relu(net_2.outputs, alpha=0.2, name='d/net_2/lrelu')
net_3 = Conv3D(net_2, df_dim*4, '3', batch_norm = not improved, is_train = is_train)
net_3.outputs = tl.activation.leaky_relu(net_3.outputs, alpha=0.2, name='d/net_3/lrelu')
net_4 = Conv3D(net_3, df_dim*8, '4', batch_norm = not improved, is_train = is_train)
net_4.outputs = tl.activation.leaky_relu(net_4.outputs, alpha=0.2, name='d/net_4/lrelu')
net_5 = FlattenLayer(net_4, name='d/net_5/flatten')
net_5 = tl.layers.DenseLayer(net_5, n_units=output_units, act=tf.identity,
W_init = tf.random_normal_initializer(stddev=0.02),
name='d/net_5/dense')
if sig:
return net_5, tf.nn.sigmoid(net_5.outputs)
else:
return net_5, net_5.outputs
def VAE(images, is_train = True):
sizes = [64,128,256,512,400]
with tf.variable_scope("vae") as vs:
net_0 = tl.layers.InputLayer(images, name='v/net_0/in')
net_1 = tl.layers.Conv2dLayer(net_0, shape=[11, 11, 3, sizes[0]],
W_init = tf.random_normal_initializer(stddev=0.02),
strides=[1, 4, 4, 1], name='v/net_1/conv2d')
net_1.outputs = tl.activation.leaky_relu(net_1.outputs, alpha=0.2, name='v/net_1/lrelu')
net_2 = tl.layers.Conv2dLayer(net_1, shape=[5, 5, sizes[0], sizes[1]],
W_init = tf.random_normal_initializer(stddev=0.02),
strides=[1, 4, 4, 1], name='v/net_2/conv2d')
net_2 = tl.layers.BatchNormLayer(net_2, is_train=is_train, name='v/net_2/batch_norm')
net_2.outputs = tl.activation.leaky_relu(net_2.outputs, alpha=0.2, name='v/net_2/lrelu')
net_3 = tl.layers.Conv2dLayer(net_2, shape=[5, 5, sizes[1], sizes[2]],
W_init = tf.random_normal_initializer(stddev=0.02),
strides=[1, 2, 2, 1], name='v/net_3/conv2d')
net_3 = tl.layers.BatchNormLayer(net_3, is_train=is_train, name='v/net_3/batch_norm')
net_3.outputs = tl.activation.leaky_relu(net_3.outputs, alpha=0.2, name='v/net_3/lrelu')
net_4 = tl.layers.Conv2dLayer(net_3, shape=[5, 5, sizes[2], sizes[3]],
W_init = tf.random_normal_initializer(stddev=0.02),
strides=[1, 2, 2, 1], name='v/net_4/conv2d')
net_4 = tl.layers.BatchNormLayer(net_4, is_train=is_train, name='v/net_4/batch_norm')
net_4.outputs = tl.activation.leaky_relu(net_4.outputs, alpha=0.2, name='v/net_4/lrelu')
net_5 = tl.layers.Conv2dLayer(net_4, shape=[8, 8, sizes[3], sizes[4]],
W_init = tf.random_normal_initializer(stddev=0.02),
strides=[1, 1, 1, 1], name='v/net_5/conv2d')
net_5 = tl.layers.BatchNormLayer(net_5, is_train=is_train, name='v/net_5/batch_norm')
net_5.outputs = tl.activation.leaky_relu(net_5.outputs, alpha=0.2, name='v/net_5/lrelu')
net_6 = FlattenLayer(net_5, name='v/net_6/flatten')
means = tl.layers.DenseLayer(net_6, n_units= 200, act=tf.identity,
W_init = tf.random_normal_initializer(stddev=0.02),
name='v/means')
sigmas = tl.layers.DenseLayer(net_6, n_units= 200, act=tf.tanh,
W_init = tf.random_normal_initializer(stddev=0.02),
name='v/sigmas')
return means,sigmas,means.outputs,sigmas.outputs
def surface_VAE(inputs, is_train = True, batch_size= 128, output_size = 20):
with tf.variable_scope("vae") as vs:
inputs = tf.reshape(inputs,[batch_size,output_size,output_size,output_size,1])
df_dim = output_size # Dimension of discrim filters in first conv layer. [64]
net_0 = tl.layers.InputLayer(inputs, name='v/net_0/in')
net_1 = tl.layers.Conv3dLayer(net_0, shape=[4, 4, 4, 1, df_dim],
W_init = tf.random_normal_initializer(stddev=0.02),
strides=[1, 2, 2, 2, 1], name='v/net_1/conv2d')
net_1.outputs = tl.activation.leaky_relu(net_1.outputs, alpha=0.2, name='v/net_1/lrelu')
net_2 = tl.layers.Conv3dLayer(net_1, shape=[4, 4, 4, df_dim, df_dim*2],
W_init = tf.random_normal_initializer(stddev=0.02),
strides=[1, 2, 2, 2, 1], name='v/net_2/conv2d')
net_2 = tl.layers.BatchNormLayer(net_2, is_train=is_train, name='v/net_2/batch_norm')
net_2.outputs = tl.activation.leaky_relu(net_2.outputs, alpha=0.2, name='v/net_2/lrelu')
net_3 = tl.layers.Conv3dLayer(net_2, shape=[4, 4, 4, df_dim*2, df_dim*4],
W_init = tf.random_normal_initializer(stddev=0.02),
strides=[1, 2, 2, 2, 1], name='v/net_3/conv2d')
net_3 = tl.layers.BatchNormLayer(net_3, is_train=is_train, name='v/net_3/batch_norm')
net_3.outputs = tl.activation.leaky_relu(net_3.outputs, alpha=0.2, name='v/net_3/lrelu')
net_4 = tl.layers.Conv3dLayer(net_3, shape=[4, 4, 4, df_dim*4, df_dim*8],
W_init = tf.random_normal_initializer(stddev=0.02),
strides=[1, 2, 2, 2, 1], name='v/net_4/conv2d')
net_4.outputs = tl.activation.leaky_relu(net_4.outputs, alpha=0.2, name='v/net_4/lrelu')
net_4 = tl.layers.BatchNormLayer(net_4, is_train=is_train, name='v/net_4/batch_norm')
net_5 = FlattenLayer(net_4, name='v/net_5/flatten')
means = tl.layers.DenseLayer(net_5, n_units= 200, act=tf.identity,
W_init = tf.random_normal_initializer(stddev=0.02),
name='v/means/id')
sigmas = tl.layers.DenseLayer(net_5, n_units= 200, act=tf.tanh,
W_init = tf.random_normal_initializer(stddev=0.02),
name='v/sigmas/id')
return means,sigmas,means.outputs,sigmas.outputs