-
Notifications
You must be signed in to change notification settings - Fork 6
/
unet3plus_deep_supervision_cgm.py
167 lines (122 loc) · 6.98 KB
/
unet3plus_deep_supervision_cgm.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
"""
UNet_3Plus with Deep Supervision and Classification Guided Module
"""
import tensorflow as tf
import tensorflow.keras as k
from .unet3plus_utils import conv_block, dot_product
def unet3plus_deepsup_cgm(input_shape, output_channels, training=False):
""" UNet_3Plus with Deep Supervision and Classification Guided Module """
filters = [64, 128, 256, 512, 1024]
input_layer = k.layers.Input(shape=input_shape, name="input_layer") # 320*320*3
""" Encoder"""
# block 1
e1 = conv_block(input_layer, filters[0]) # 320*320*64
# block 2
e2 = k.layers.MaxPool2D(pool_size=(2, 2))(e1) # 160*160*64
e2 = conv_block(e2, filters[1]) # 160*160*128
# block 3
e3 = k.layers.MaxPool2D(pool_size=(2, 2))(e2) # 80*80*128
e3 = conv_block(e3, filters[2]) # 80*80*256
# block 4
e4 = k.layers.MaxPool2D(pool_size=(2, 2))(e3) # 40*40*256
e4 = conv_block(e4, filters[3]) # 40*40*512
# block 5, bottleneck layer
e5 = k.layers.MaxPool2D(pool_size=(2, 2))(e4) # 20*20*512
e5 = conv_block(e5, filters[4]) # 20*20*1024
""" Classification Guided Module. Part 1"""
cls = k.layers.Dropout(rate=0.5)(e5)
cls = k.layers.Conv2D(2, kernel_size=(1, 1), padding="same", strides=(1, 1))(cls)
cls = k.layers.GlobalMaxPooling2D()(cls)
cls = k.activations.sigmoid(cls)
cls = tf.argmax(cls, axis=-1)
cls = cls[..., tf.newaxis]
cls = tf.cast(cls, dtype=tf.float32, )
""" Decoder """
cat_channels = filters[0]
cat_blocks = len(filters)
upsample_channels = cat_blocks * cat_channels
""" d4 """
e1_d4 = k.layers.MaxPool2D(pool_size=(8, 8))(e1) # 320*320*64 --> 40*40*64
e1_d4 = conv_block(e1_d4, cat_channels, n=1) # 320*320*64 --> 40*40*64
e2_d4 = k.layers.MaxPool2D(pool_size=(4, 4))(e2) # 160*160*128 --> 40*40*128
e2_d4 = conv_block(e2_d4, cat_channels, n=1) # 160*160*128 --> 40*40*64
e3_d4 = k.layers.MaxPool2D(pool_size=(2, 2))(e3) # 80*80*256 --> 40*40*256
e3_d4 = conv_block(e3_d4, cat_channels, n=1) # 80*80*256 --> 40*40*64
e4_d4 = conv_block(e4, cat_channels, n=1) # 40*40*512 --> 40*40*64
e5_d4 = k.layers.UpSampling2D(size=(2, 2), interpolation='bilinear')(e5) # 80*80*256 --> 40*40*256
e5_d4 = conv_block(e5_d4, cat_channels, n=1) # 20*20*1024 --> 20*20*64
d4 = k.layers.concatenate([e1_d4, e2_d4, e3_d4, e4_d4, e5_d4])
d4 = conv_block(d4, upsample_channels, n=1) # 40*40*320 --> 40*40*320
""" d3 """
e1_d3 = k.layers.MaxPool2D(pool_size=(4, 4))(e1) # 320*320*64 --> 80*80*64
e1_d3 = conv_block(e1_d3, cat_channels, n=1) # 80*80*64 --> 80*80*64
e2_d3 = k.layers.MaxPool2D(pool_size=(2, 2))(e2) # 160*160*256 --> 80*80*256
e2_d3 = conv_block(e2_d3, cat_channels, n=1) # 80*80*256 --> 80*80*64
e3_d3 = conv_block(e3, cat_channels, n=1) # 80*80*512 --> 80*80*64
e4_d3 = k.layers.UpSampling2D(size=(2, 2), interpolation='bilinear')(d4) # 40*40*320 --> 80*80*320
e4_d3 = conv_block(e4_d3, cat_channels, n=1) # 80*80*320 --> 80*80*64
e5_d3 = k.layers.UpSampling2D(size=(4, 4), interpolation='bilinear')(e5) # 20*20*320 --> 80*80*320
e5_d3 = conv_block(e5_d3, cat_channels, n=1) # 80*80*320 --> 80*80*64
d3 = k.layers.concatenate([e1_d3, e2_d3, e3_d3, e4_d3, e5_d3])
d3 = conv_block(d3, upsample_channels, n=1) # 80*80*320 --> 80*80*320
""" d2 """
e1_d2 = k.layers.MaxPool2D(pool_size=(2, 2))(e1) # 320*320*64 --> 160*160*64
e1_d2 = conv_block(e1_d2, cat_channels, n=1) # 160*160*64 --> 160*160*64
e2_d2 = conv_block(e2, cat_channels, n=1) # 160*160*256 --> 160*160*64
d3_d2 = k.layers.UpSampling2D(size=(2, 2), interpolation='bilinear')(d3) # 80*80*320 --> 160*160*320
d3_d2 = conv_block(d3_d2, cat_channels, n=1) # 160*160*320 --> 160*160*64
d4_d2 = k.layers.UpSampling2D(size=(4, 4), interpolation='bilinear')(d4) # 40*40*320 --> 160*160*320
d4_d2 = conv_block(d4_d2, cat_channels, n=1) # 160*160*320 --> 160*160*64
e5_d2 = k.layers.UpSampling2D(size=(8, 8), interpolation='bilinear')(e5) # 20*20*320 --> 160*160*320
e5_d2 = conv_block(e5_d2, cat_channels, n=1) # 160*160*320 --> 160*160*64
d2 = k.layers.concatenate([e1_d2, e2_d2, d3_d2, d4_d2, e5_d2])
d2 = conv_block(d2, upsample_channels, n=1) # 160*160*320 --> 160*160*320
""" d1 """
e1_d1 = conv_block(e1, cat_channels, n=1) # 320*320*64 --> 320*320*64
d2_d1 = k.layers.UpSampling2D(size=(2, 2), interpolation='bilinear')(d2) # 160*160*320 --> 320*320*320
d2_d1 = conv_block(d2_d1, cat_channels, n=1) # 160*160*320 --> 160*160*64
d3_d1 = k.layers.UpSampling2D(size=(4, 4), interpolation='bilinear')(d3) # 80*80*320 --> 320*320*320
d3_d1 = conv_block(d3_d1, cat_channels, n=1) # 320*320*320 --> 320*320*64
d4_d1 = k.layers.UpSampling2D(size=(8, 8), interpolation='bilinear')(d4) # 40*40*320 --> 320*320*320
d4_d1 = conv_block(d4_d1, cat_channels, n=1) # 320*320*320 --> 320*320*64
e5_d1 = k.layers.UpSampling2D(size=(16, 16), interpolation='bilinear')(e5) # 20*20*320 --> 320*320*320
e5_d1 = conv_block(e5_d1, cat_channels, n=1) # 320*320*320 --> 320*320*64
d1 = k.layers.concatenate([e1_d1, d2_d1, d3_d1, d4_d1, e5_d1, ])
d1 = conv_block(d1, upsample_channels, n=1) # 320*320*320 --> 320*320*320
""" Deep Supervision Part"""
# last layer does not have batch norm and relu
d1 = conv_block(d1, output_channels, n=1, is_bn=False, is_relu=False)
if training:
d2 = conv_block(d2, output_channels, n=1, is_bn=False, is_relu=False)
d3 = conv_block(d3, output_channels, n=1, is_bn=False, is_relu=False)
d4 = conv_block(d4, output_channels, n=1, is_bn=False, is_relu=False)
e5 = conv_block(e5, output_channels, n=1, is_bn=False, is_relu=False)
# d1 = no need for up sampling
d2 = k.layers.UpSampling2D(size=(2, 2), interpolation='bilinear')(d2)
d3 = k.layers.UpSampling2D(size=(4, 4), interpolation='bilinear')(d3)
d4 = k.layers.UpSampling2D(size=(8, 8), interpolation='bilinear')(d4)
e5 = k.layers.UpSampling2D(size=(16, 16), interpolation='bilinear')(e5)
""" Classification Guided Module. Part 2"""
d1 = dot_product(d1, cls)
d1 = k.activations.sigmoid(d1)
if training:
d2 = dot_product(d2, cls)
d3 = dot_product(d3, cls)
d4 = dot_product(d4, cls)
e5 = dot_product(e5, cls)
d2 = k.activations.sigmoid(d2)
d3 = k.activations.sigmoid(d3)
d4 = k.activations.sigmoid(d4)
e5 = k.activations.sigmoid(e5)
if training:
return tf.keras.Model(inputs=input_layer, outputs=[d1, d2, d3, d4, e5, cls], name='UNet3Plus_DeepSup_CGM')
else:
return tf.keras.Model(inputs=input_layer, outputs=[d1, ], name='UNet3Plus_DeepSup_CGM')
if __name__ == "__main__":
"""## Model Compilation"""
INPUT_SHAPE = [320, 320, 1]
OUTPUT_CHANNELS = 1
unet_3P = unet3plus_deepsup_cgm(INPUT_SHAPE, OUTPUT_CHANNELS)
unet_3P.summary()
# tf.keras.utils.plot_model(unet_3P, show_layer_names=True, show_shapes=True)
# unet_3P.save("unet_3P.hdf5")