-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathbionet3d.py
164 lines (147 loc) · 7.85 KB
/
bionet3d.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
from torch.nn import Module, Sequential, Conv3d, BatchNorm3d, ConvTranspose3d, ReLU, MaxPool3d, Sigmoid, Parameter
from torch import tensor, cat
class BiONet(Module):
def __init__(self,
input_channels: int = 1,
num_classes: int = 1,
iterations: int = 2,
multiplier: float = 1.0,
num_layers: int = 4,
integrate: bool = False):
super(BiONet, self).__init__()
#
self.input_channels = input_channels
self.iterations = iterations
self.multiplier = multiplier
self.num_layers = num_layers
self.integrate = integrate
self.batch_norm_momentum = 0.01
self.filters_list = [int(32 * (2 ** i) * self.multiplier) for i in range(self.num_layers + 1)]
# First downsizing block
self.pre_transform_conv_block = Sequential(
Conv3d(self.input_channels, self.filters_list[0], kernel_size=(3, 3, 3),padding=(1, 1, 1), stride=1),
ReLU(),
BatchNorm3d(self.filters_list[0], momentum=self.batch_norm_momentum),
Conv3d(self.filters_list[0], self.filters_list[0], kernel_size=(3, 3, 3),padding=(1, 1, 1), stride=1),
ReLU(),
BatchNorm3d(self.filters_list[0], momentum=self.batch_norm_momentum),
Conv3d(self.filters_list[0], self.filters_list[0], kernel_size=(3, 3, 3),padding=(1, 1, 1), stride=1),
ReLU(),
BatchNorm3d(self.filters_list[0], momentum=self.batch_norm_momentum),
MaxPool3d(kernel_size=(2, 2, 2), stride=2, padding=(0, 0, 0))
)
self.reuse_convs = []
self.encoders = []
self.reuse_deconvs = []
self.decoders = []
for iteration in range(self.iterations):
for layer in range(self.num_layers):
in_channel = self.filters_list[layer] * 2
mid_channel = self.filters_list[layer]
out_channel = self.filters_list[layer + 1]
# Encoder
if iteration == 0:
conv1 = Conv3d(in_channel, mid_channel, kernel_size=(3, 3, 3),padding=(1, 1, 1), stride=1)
conv2 = Conv3d(mid_channel, mid_channel, kernel_size=(3, 3, 3),padding=(1, 1, 1), stride=1)
conv3 = Conv3d(mid_channel, out_channel, kernel_size=(3, 3, 3),padding=(1, 1, 1), stride=1)
self.reuse_convs.append((conv1, conv2, conv3))
convs = Sequential(
self.reuse_convs[layer][0],
ReLU(),
BatchNorm3d(mid_channel, momentum=self.batch_norm_momentum),
self.reuse_convs[layer][1],
ReLU(),
BatchNorm3d(mid_channel, momentum=self.batch_norm_momentum)
)
# DOWN
down = Sequential(
self.reuse_convs[layer][2],
ReLU(),
BatchNorm3d(out_channel, momentum=self.batch_norm_momentum),
MaxPool3d(kernel_size=(2, 2, 2), stride=2, padding=(0, 0, 0))
)
self.add_module("iteration{0}_layer{1}_encoder_convs".format(iteration, layer), convs)
self.add_module("iteration{0}_layer{1}_encoder_down".format(iteration, layer), down)
self.encoders.append((convs, down))
# Decoders
in_channel = self.filters_list[self.num_layers - layer] + self.filters_list[self.num_layers - 1 - layer]
out_channel = self.filters_list[self.num_layers - 1 - layer]
if iteration == 0:
conv1 = Conv3d(in_channel, out_channel, kernel_size=(3, 3, 3),padding=(1, 1, 1), stride=1)
conv2 = Conv3d(out_channel, out_channel, kernel_size=(3, 3, 3),padding=(1, 1, 1), stride=1)
conv3 = ConvTranspose3d(out_channel, out_channel, kernel_size=(3, 3, 3),padding=(1, 1, 1), stride=2, output_padding=(1, 1, 1))
self.reuse_deconvs.append((conv1, conv2, conv3))
convs = Sequential(
self.reuse_deconvs[layer][0],
ReLU(),
BatchNorm3d(out_channel, momentum=self.batch_norm_momentum),
self.reuse_deconvs[layer][1],
ReLU(),
BatchNorm3d(out_channel, momentum=self.batch_norm_momentum)
)
# UP
up = Sequential(
self.reuse_deconvs[layer][2],
ReLU(),
BatchNorm3d(out_channel, momentum=self.batch_norm_momentum)
)
self.add_module("iteration{0}_layer{1}_decoder_convs".format(iteration, layer), convs)
self.add_module("iteration{0}_layer{1}_decoder_up".format(iteration, layer), up)
self.decoders.append((convs, up))
# Bottleneck
self.middles = Sequential(
Conv3d(self.filters_list[-1], self.filters_list[-1], kernel_size=(3, 3, 3),padding=(1, 1, 1), stride=1),
ReLU(),
BatchNorm3d(self.filters_list[-1], momentum=self.batch_norm_momentum),
Conv3d(self.filters_list[-1], self.filters_list[-1], kernel_size=(3, 3, 3),padding=(1, 1, 1), stride=1),
ReLU(),
BatchNorm3d(self.filters_list[-1], momentum=self.batch_norm_momentum),
ConvTranspose3d(self.filters_list[-1], self.filters_list[-1], kernel_size=(3, 3, 3),padding=(1, 1, 1),
stride=2, output_padding=(1, 1, 1)),
ReLU(),
BatchNorm3d(self.filters_list[-1], momentum=self.batch_norm_momentum)
)
self.post_transform_conv_block = Sequential(
Conv3d(self.filters_list[0] * self.iterations, self.filters_list[0], kernel_size=(3, 3, 3),padding=(1, 1, 1),
stride=1) if self.integrate else Conv3d(self.filters_list[0],
self.filters_list[0], kernel_size=(3, 3, 3),
padding=(1, 1, 1), stride=1),
ReLU(),
BatchNorm3d(self.filters_list[0], momentum=self.batch_norm_momentum),
Conv3d(self.filters_list[0], self.filters_list[0], kernel_size=(3, 3, 3),padding=(1, 1, 1), stride=1),
ReLU(),
BatchNorm3d(self.filters_list[0], momentum=self.batch_norm_momentum),
Conv3d(self.filters_list[0], 1, kernel_size=(1, 1, 1), stride=1),
Sigmoid(),
)
def forward(self, x: tensor) -> tensor:
enc = [None for i in range(self.num_layers)]
dec = [None for i in range(self.num_layers)]
all_output = [None for i in range(self.iterations)]
x = self.pre_transform_conv_block(x)
e_i = 0
d_i = 0
for iteration in range(self.iterations):
# Through encoder
for layer in range(self.num_layers):
if layer == 0:
x_in = x
# print(x_in.shape)
x_in = self.encoders[e_i][0](cat([x_in, x_in if dec[-1 - layer] is None else dec[-1 - layer]], dim=1))
enc[layer] = x_in
x_in = self.encoders[e_i][1](x_in)
e_i = e_i + 1
# Bottleneck
x_in = self.middles(x_in)
# print(x_in.shape)
# Decoder
for layer in range(self.num_layers):
x_in = self.decoders[d_i][0](cat([x_in, enc[-1 - layer]], dim=1))
dec[layer] = x_in
x_in = self.decoders[d_i][1](x_in)
d_i = d_i + 1
all_output[iteration] = x_in
if self.integrate:
x_in = cat(all_output, dim=1)
x_in = self.post_transform_conv_block(x_in)
return x_in