-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathreconstruction_architectures.py
201 lines (142 loc) · 10.4 KB
/
reconstruction_architectures.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
## Deep Active Lesion Segmention (DALS), Code by Ali Hatamizadeh ( http://web.cs.ucla.edu/~ahatamiz/ )
import tensorflow as tf
import tf_slim as slim
_BATCH_NORM_EPSILON = 1e-5
_BATCH_NORM_DECAY = 0.01
regularizer = 0.0008
dropout_rate = 0.3
regularizer = tf.keras.regularizers.L2(l2=regularizer)
w_init_xavi = tf.keras.initializers.GlorotUniform()
w_init_he = tf.keras.initializers.VarianceScaling()
w_init_zero = tf.keras.initializers.VarianceScaling()
def Concatenation(layers) :
return tf.concat(layers, axis=3)
def bottleneck_layer(x, reg, dilation, kernel_init,is_training):
x = Batch_Normalization(x, is_training)
x = tf.compat.v1.layers.conv2d(x,filters=4*growth_k,kernel_size=3,strides=1,padding="same",activation=tf.nn.relu, kernel_regularizer=reg, kernel_initializer= kernel_init, dilation_rate=dilation)
x = tf.compat.v1.layers.dropout(x, rate=dropout_rate, training=is_training,)
x = Batch_Normalization(x, is_training)
x = tf.compat.v1.layers.conv2d(x,filters=4*growth_k,kernel_size=3,strides=1,padding="same",activation=tf.nn.relu, kernel_regularizer=reg, kernel_initializer= kernel_init, dilation_rate=dilation)
x = tf.compat.v1.layers.dropout(x, rate=dropout_rate, training=is_training,)
return x
def transition_layer(x, reg, kernel_init,is_training):
x = Batch_Normalization(x, is_training)
x = tf.compat.v1.layers.conv2d(x,filters=growth_k,kernel_size=3,strides=1,padding="same",activation=tf.nn.relu, kernel_regularizer=reg, kernel_initializer= kernel_init)
x = tf.compat.v1.layers.dropout(x, rate=dropout_rate, training=is_training)
x = tf.compat.v1.layers.average_pooling2d(inputs=x, pool_size=2, strides=2, padding="VALID")
return x
def dense_block(input_x, nb_layers, reg, dilation, kernel_init,is_training):
layers_concat = list()
layers_concat.append(input_x)
x = bottleneck_layer(input_x, reg, dilation, kernel_init,is_training)
layers_concat.append(x)
for i in range(nb_layers - 1):
x = Concatenation(layers_concat)
x = bottleneck_layer(x, reg, dilation, kernel_init,is_training)
layers_concat.append(x)
x = Concatenation(layers_concat)
return x
def Batch_Normalization(inputs, training):
return tf.compat.v1.layers.batch_normalization(inputs=inputs,axis=-1, momentum=_BATCH_NORM_DECAY, epsilon=_BATCH_NORM_EPSILON, training=training)
def conv_Norm_Relu(x,f,k,s,is_training,padding_mode="same"):
x=tf.compat.v1.layers.conv2d(x, kernel_initializer= w_init_zero, filters=f, kernel_size=k, strides=s, padding=padding_mode, activation = None, kernel_regularizer=regularizer)
return tf.nn.relu(Batch_Normalization(x, is_training))
@slim.add_arg_scope
def atrous_spatial_pyramid_pooling(net, depth=512):
image_level_features = tf.reduce_mean(net, [1, 2], name='image_level_global_pool', keep_dims=True)
image_level_features = slim.conv2d(image_level_features, depth, [1, 1], scope="image_level_conv_1x1", activation_fn=None)
at_pool1x1 = slim.conv2d(net, depth, [1, 1], scope="conv_1x1_0", activation_fn=None)
at_pool3x3_1 = slim.conv2d(net, depth, [3, 3], scope="conv_3x3_1", rate=6, activation_fn=None)
at_pool3x3_2 = slim.conv2d(net, depth, [3, 3], scope="conv_3x3_2", rate=12, activation_fn=None)
at_pool3x3_3 = slim.conv2d(net, depth, [3, 3], scope="conv_3x3_3", rate=18, activation_fn=None)
net = tf.concat((image_level_features, at_pool1x1, at_pool3x3_1, at_pool3x3_2, at_pool3x3_3), axis=3,name="concat")
net = slim.conv2d(net, depth, [1, 1], scope="conv_1x1_output", activation_fn=None)
return net
def transpose_conv_block(x,y,f,k,s,is_training):
conv1 = tf.concat([tf.compat.v1.layers.conv2d_transpose(x,kernel_initializer= w_init_zero, filters=f, kernel_size=2, strides=2, padding="same", activation = tf.nn.relu, kernel_regularizer=regularizer), y], 3)
conv1 = tf.compat.v1.layers.conv2d(conv1, kernel_initializer= w_init_zero, filters=f, kernel_size=k, strides=s, padding="same", activation = tf.nn.relu, kernel_regularizer=regularizer)
conv1 = Batch_Normalization(conv1,is_training)
conv1 = tf.compat.v1.layers.conv2d(conv1, kernel_initializer= w_init_zero, filters=f, kernel_size=k, strides=s, padding="same", activation = tf.nn.relu, kernel_regularizer=regularizer)
conv1 = Batch_Normalization(conv1,is_training)
return conv1
def res_block(x,f,is_training):
conv1=conv_Norm_Relu(x,f,3,1,is_training,padding_mode="same")
conv1 = Batch_Normalization(tf.compat.v1.layers.conv2d(conv1,filters=f,kernel_size=3,strides=1,padding="same",activation=None, kernel_regularizer=regularizer, kernel_initializer= w_init_zero))
return tf.nn.relu(tf.add(conv1,x))
def max_pool(x, p, s):
pool1 = tf.compat.v1.layers.max_pooling2d(x, pool_size=p, strides=s, padding="valid")
return pool1
def conv_block(x,f,k,s):
conv1=conv_Norm_Relu(conv_Norm_Relu(x,f,k,s,padding_mode="same"),f,k,s,padding_mode="same",)
return tf.compat.v1.layers.max_pooling2d(conv1, pool_size=2, strides=2, padding="valid")
def conv_start_ddunet(x,f,k,s,is_training):
conv1 = Batch_Normalization(x, is_training)
conv1 = tf.compat.v1.layers.conv2d(conv1, kernel_initializer= w_init_zero, filters=f, kernel_size=k, strides=s, padding="same", activation = tf.nn.relu, kernel_regularizer=regularizer)
return conv1
def upsample_block(x,y):
size = tf.shape(y)
up = tf.concat([tf.image.resize_bilinear(x, (size[1], size[2]),align_corners=True), y], 3)
return up
def dilation_block(x, d1, d2, d3, d4, filter_1,is_training):
dense51 = tf.compat.v1.layers.conv2d(inputs=x, filters=filter_1,kernel_size=3,strides=1,padding="same",activation=tf.nn.relu, dilation_rate=d1, kernel_regularizer=regularizer, kernel_initializer= w_init_zero)
dense51 = tf.compat.v1.layers.batch_normalization(inputs=dense51,axis=-1,momentum=0.9,epsilon=0.001,center=True,scale=True,training=is_training,)
dense52 = tf.compat.v1.layers.conv2d(inputs=x, filters=filter_1,kernel_size=3,strides=1,padding="same",activation=tf.nn.relu, dilation_rate=d2, kernel_regularizer=regularizer, kernel_initializer= w_init_zero)
dense52 = tf.compat.v1.layers.batch_normalization(inputs=dense52,axis=-1,momentum=0.9,epsilon=0.001,center=True,scale=True,training=is_training,)
dense53 = tf.compat.v1.layers.conv2d(inputs=x, filters=filter_1,kernel_size=3,strides=1,padding="same",activation=tf.nn.relu, dilation_rate=d3, kernel_regularizer=regularizer, kernel_initializer= w_init_zero)
dense53 = tf.compat.v1.layers.batch_normalization(inputs=dense53,axis=-1,momentum=0.9,epsilon=0.001,center=True,scale=True,training=is_training,)
dense54 = tf.compat.v1.layers.conv2d(inputs=x, filters=filter_1,kernel_size=3,strides=1,padding="same",activation=tf.nn.relu, dilation_rate=d4, kernel_regularizer=regularizer, kernel_initializer= w_init_zero)
dense54 = tf.compat.v1.layers.batch_normalization(inputs=dense54,axis=-1,momentum=0.9,epsilon=0.001,center=True,scale=True,training=is_training,)
dilation = tf.concat([dense51, dense52, dense53, dense54],3)
return dilation
def unet(x):
conv1=conv_block(x,64,3,1)
conv2=conv_block(conv1,128,3,1)
conv3=conv_block(conv2,256,3,1)
conv4=conv_block(conv3,512,3,1)
conv5=conv_block(conv4,1024,3,1)
conv6=upsample_block(conv5,conv4)
conv7=upsample_block(conv6,conv3)
conv8=upsample_block(conv7,conv2)
conv9=upsample_block(conv8,conv1)
conv10=upsample_block(conv9,conv_Norm_Relu(conv_Norm_Relu(x,256,3,1),256,3,1))
out_seg=tf.nn.sigmoid(conv_Norm_Relu(conv_Norm_Relu(conv10,256,3,1,padding_mode="same"),1,1,1))
return out_seg
growth_k = 6
def ddunet(x,is_training,lab):
conv1 = conv_start_ddunet(x, 64, 7, 2,is_training)
pool = max_pool(conv1, 2, 2)
dense1 = dense_block(pool, 3, regularizer,1, w_init_zero,is_training)
transition1 = transition_layer(dense1, regularizer, w_init_zero,is_training)
dense2 = dense_block(transition1, 6, regularizer,1, w_init_zero,is_training)
transition2 = transition_layer(dense2, regularizer, w_init_zero,is_training)
dense3 = dense_block(transition2, 18, regularizer,1, w_init_zero,is_training)
transition3 = transition_layer(dense3, regularizer, w_init_zero,is_training)
dense4 = dense_block(transition3, 12, regularizer,1, w_init_zero,is_training)
dilation = dilation_block(dense4, 6, 12, 18, 24, 1024,is_training)
# labeling architecture
if lab == 1:
up6 = transpose_conv_block(dilation,dense3,512,3,1,is_training)
up7 = transpose_conv_block(up6,dense2,256,3,1,is_training)
up8 = transpose_conv_block(up7,dense1,128,3,1,is_training)
up9 = transpose_conv_block(up8,conv1,64,3,1,is_training)
up10 = tf.compat.v1.layers.conv2d_transpose(up9, filters=64,kernel_size=2,strides=2,padding="same", activation = tf.nn.relu, kernel_regularizer=regularizer, kernel_initializer= w_init_zero)
conv2 = Batch_Normalization(up10, is_training)
conv11 = tf.compat.v1.layers.conv2d(conv2,filters=128,kernel_size=3,strides=1,padding="same",activation=tf.nn.relu, kernel_regularizer=regularizer, kernel_initializer= w_init_zero)
conv11 = Batch_Normalization(conv11, is_training)
#conv11 gives 256*256*128
out_seg = tf.compat.v1.layers.conv2d(conv11,filters=1,kernel_size=1,strides=1,activation=tf.nn.sigmoid)
map_lambda1 = tf.exp(tf.divide(tf.subtract(2.0,out_seg),tf.add(1.0,out_seg)))
map_lambda2 = tf.exp(tf.divide(tf.add(1.0, out_seg), tf.subtract(2.0, out_seg)))
return out_seg,map_lambda1,map_lambda2
#reconstruction architecture
if lab ==0:
up6_t = transpose_conv_block(dilation,dense3,512,3,1,is_training)
up7_t = transpose_conv_block(up6_t,dense2,256,3,1,is_training)
up8_t = transpose_conv_block(up7_t,dense1,128,3,1,is_training)
up9_t = transpose_conv_block(up8_t,conv1,64,3,1,is_training)
up10_t = tf.compat.v1.layers.conv2d_transpose(up9_t, filters=64,kernel_size=2,strides=2,padding="same", activation = tf.nn.relu, kernel_regularizer=regularizer, kernel_initializer= w_init_zero)
conv2_t = Batch_Normalization(up10_t, is_training)
conv11_t = tf.compat.v1.layers.conv2d(conv2_t,filters=128,kernel_size=3,strides=1,padding="same",activation=tf.nn.relu, kernel_regularizer=regularizer, kernel_initializer= w_init_zero)
conv11_t = Batch_Normalization(conv11_t, is_training)
rec_img = tf.compat.v1.layers.conv2d(conv11_t,filters=3,kernel_size=1,strides=1,padding="same",activation=tf.nn.relu)
return rec_img