TeethSegmentation/model.py at main · RobertSmithers/TeethSegmentation · GitHub

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
import torch
import torch.nn as nn
import torch.nn.functional as F

# NOTE: On torch version 1.7 (which I was previously on), the "same" padding option was not valid for Conv2d. Requires torch v1.10 and up

class DoubleConvBlock(nn.Module):
  # (conv) => (dropout+ReLU) => (conv) => (bn+ReLU)

  def __init__(self, in_ch, out_ch, kernel_size, batch_norm=True):
    super(DoubleConvBlock, self).__init__()

    if batch_norm:
      self.conv = nn.Sequential(
          nn.Conv2d(in_ch, out_ch, 3, padding="same"),
          nn.Dropout2d(0.1),
          nn.BatchNorm2d(out_ch),
          # nn.LeakyReLU(inplace=True),
          nn.Conv2d(out_ch, out_ch, 3, padding="same"),
          nn.BatchNorm2d(out_ch),             # Heavily debated, but have decided to normalize before activation
          nn.LeakyReLU(inplace=True)
      )
    else:
      self.conv = nn.Sequential(
          nn.Conv2d(in_ch, out_ch, 3, padding="same"),
          nn.Dropout2d(0.1),
          nn.LeakyReLU(inplace=True),
          nn.Conv2d(out_ch, out_ch, 3, padding="same"),
          nn.LeakyReLU(inplace=True)
      )

  def forward(self, x):
      return self.conv(x)

# This is our encoder
class PoolConvBlock(nn.Module):
  # (Pool) => (DoubleConvBlock)

  def __init__(self, pool_kernel_size=(2,2), conv_in_ch=32, conv_out_ch=32, conv_kernel_size=(3,3), batch_norm=True):
    super(PoolConvBlock, self).__init__()
    self.pool = nn.MaxPool2d(pool_kernel_size)
    self.conv = DoubleConvBlock(conv_in_ch, conv_out_ch, conv_kernel_size, batch_norm=batch_norm)

  def forward(self, x):
    p = self.pool(x)
    x = self.conv(p)
    return x, p

class TConv(nn.Module):
  # (ConvTranspose) => (ReLU)

  def __init__(self, in_ch, out_ch, kernel_size, stride=None):
    super(TConv, self).__init__()
    self.conv = nn.Sequential(
          nn.ConvTranspose2d(in_ch, out_ch, kernel_size, padding=1, stride=stride),
          # nn.LeakyReLU(inplace=True)
      )

  def forward(self, x, skip):
    x = self.conv(x)
    x = torch.cat([x, skip], axis=1)
    return x

class UNet(nn.Module):
    def __init__(self, input_shape=(512,512,1)):
        super(UNet, self).__init__()

        # conv1 = Conv2D(32,(3,3), activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(inputs)
        # d1=Dropout(0.1)(conv1)
        # conv2 = Conv2D(32,(3,3), activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(d1)
        # b=BatchNormalization()(conv2)
        self.conv = DoubleConvBlock(1, 32, (3,3))

        # pool1 = MaxPooling2D(pool_size=(2, 2))(b)
        # conv3 = Conv2D(64,(3,3), activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(pool1)
        # d2=Dropout(0.2)(conv3)
        # conv4 = Conv2D(64,(3,3), activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(d2)
        # b1=BatchNormalization()(conv4)
        self.pconv1 = PoolConvBlock(conv_in_ch=32, conv_out_ch=64)

        # pool2 = MaxPooling2D(pool_size=(2, 2))(b1)
        # conv5 = Conv2D(128,(3,3), activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(pool2)
        # d3=Dropout(0.3)(conv5)
        # conv6 = Conv2D(128,(3,3), activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(d3)
        # b2=BatchNormalization()(conv6)
        self.pconv2 = PoolConvBlock(conv_in_ch=64, conv_out_ch=128)

        # pool3 = MaxPooling2D(pool_size=(2, 2))(b2)
        # conv7 = Conv2D(256,(3,3), activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(pool3)
        # d4=Dropout(0.4)(conv7)
        # conv8 = Conv2D(256,(3,3), activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(d4)
        # b3=BatchNormalization()(conv8)
        self.pconv3 = PoolConvBlock(conv_in_ch=128, conv_out_ch=256)

        # pool4 = MaxPooling2D(pool_size=(2, 2))(b3)
        # conv9 = Conv2D(512,(3,3),activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(pool4)
        # d5=Dropout(0.5)(conv9)
        # conv10 = Conv2D(512,(3,3), activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(d5)
        # b4=BatchNormalization()(conv10)
        self.pconv4 = PoolConvBlock(conv_in_ch=256, conv_out_ch=512)

        # conv11 = Conv2DTranspose(512,(4,4), activation = 'relu', padding = 'same', strides=(2,2),kernel_initializer = 'he_normal')(b4)
        self.tconv1 = TConv(512, 256, (4,4), stride=(2,2))
        # x= concatenate([conv11,conv8])

        # conv12 = Conv2D(256,(3,3), activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(x)
        # d6=Dropout(0.4)(conv12)
        # conv13 = Conv2D(256,(3,3), activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(d6)
        # b5=BatchNormalization()(conv13)
        self.conv1 = DoubleConvBlock(512, 256, (3,3))

        # conv14 = Conv2DTranspose(256,(4,4), activation = 'relu', padding = 'same', strides=(2,2),kernel_initializer = 'he_normal')(b5)
        self.tconv2 = TConv(256, 128, (4,4), stride=(2,2))
        # x1=concatenate([conv14,conv6])

        # conv15 = Conv2D(128,3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(x1)
        # d7=Dropout(0.3)(conv15)
        # conv16 = Conv2D(128,3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(d7)
        # b6=BatchNormalization()(conv16)
        self.conv2 = DoubleConvBlock(256, 128, 3)

        # conv17 = Conv2DTranspose(128,(4,4), activation = 'relu', padding = 'same',strides=(2,2), kernel_initializer = 'he_normal')(b6)
        self.tconv3 = TConv(128, 64, (4,4), stride=(2,2))

        # x2=concatenate([conv17,conv4])

        # conv18 = Conv2D(64,(3,3), activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(x2)
        # d8=Dropout(0.2)(conv18)
        # conv19 = Conv2D(64,(3,3) ,activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(d8)
        # b7=BatchNormalization()(conv19)
        self.conv3 = DoubleConvBlock(128, 64, 3)

        # conv20 = Conv2DTranspose(64,(4,4), activation = 'relu', padding = 'same',strides=(2,2), kernel_initializer = 'he_normal')(b7)
        self.tconv4 = TConv(64, 32, (4,4), stride=(2,2))

        # x3=concatenate([conv20,conv2])

        # conv21 = Conv2D(32,(3,3) ,activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(x3)
        # d9=Dropout(0.1)(conv21)
        # conv22 = Conv2D(32,(3,3), activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(d9)
        self.conv4 = DoubleConvBlock(64, 32, (3,3), batch_norm=False)

        self.conv5 = nn.Sequential(
            nn.Conv2d(32, 1, (1,1), padding="same"),
            nn.ReLU()
        )
        # outputs = Conv2D(1,(1,1), activation = last_activation, padding = 'same', kernel_initializer = 'he_normal')(conv22)

    def forward(self, x):

      ### Reduce ###

      # Outputs 512 x 512 x 32
      skip1 = self.conv(x)
      # print("level 1 layer+skip", skip1.shape)

      # Outputs 256 x 256 x 64
      skip2, layer2 = self.pconv1(skip1)
      # print("level 2 layer/skip", layer2.shape, skip2.shape)

      # Outputs 128 x 128 x 128
      skip3, layer3 = self.pconv2(skip2)
      # print("level 3 layer/skip", layer3.shape, skip3.shape)

      # Outputs 64 x 64 x 256
      skip4, layer4 = self.pconv3(skip3)
      # print("level 4 layer/skip", layer4.shape, skip4.shape)

      # Outputs 32 x 32 x 512
      skip5, x = self.pconv4(skip4)
      # print("level 5 channels", x.shape)

      ### Expand ###

      # Outputs 64 x 64 x 256
      expand4 = self.tconv1(skip5, skip4)
      expand4 = self.conv1(expand4)

      # Outputs 128 x 128 x 128
      expand3 = self.tconv2(expand4, skip3)
      expand3 = self.conv2(expand3)

      # Outputs 256 x 256 x 64
      expand2 = self.tconv3(expand3, skip2)
      expand2 = self.conv3(expand2)

      # Outputs 512 x 512 x 32
      expand1 = self.tconv4(expand2, skip1)
      expand1 = self.conv4(expand1)

      # Outputs 512 x 512 x 1
      out = self.conv5(expand1)

      # return torch.sigmoid(x)
      # No sigmoid because using BCEWithLogistLoss which is more numerically stable (using log-sum-exp trick)
      return out