File size: 6,530 Bytes
1b34a12
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
Copyright (c) Facebook, Inc. and its affiliates.

This source code is licensed under the MIT license found in the
LICENSE file in the root directory of this source tree.
"""

from typing import List, Tuple

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


class Unet(nn.Module):
    """
    PyTorch implementation of a U-Net model.

    O. Ronneberger, P. Fischer, and Thomas Brox. U-net: Convolutional networks
    for biomedical image segmentation. In International Conference on Medical
    image computing and computer-assisted intervention, pages 234–241.
    Springer, 2015.
    """

    def __init__(
        self,
        in_chans: int,
        out_chans: int,
        chans: int = 32,
        num_pool_layers: int = 4,
        drop_prob: float = 0.0,
    ):
        """
        Parameters
        ----------
        in_chans : int
            Number of channels in the input to the U-Net model.
        out_chans : int
            Number of channels in the output to the U-Net model.
        chans : int, optional
            Number of output channels of the first convolution layer. Default is 32.
        num_pool_layers : int, optional
            Number of down-sampling and up-sampling layers. Default is 4.
        drop_prob : float, optional
            Dropout probability. Default is 0.0.
        """
        super().__init__()

        self.in_chans = in_chans
        self.out_chans = out_chans
        self.chans = chans
        self.num_pool_layers = num_pool_layers
        self.drop_prob = drop_prob

        self.down_sample_layers = nn.ModuleList([ConvBlock(in_chans, chans, drop_prob)])
        ch = chans
        for _ in range(num_pool_layers - 1):
            self.down_sample_layers.append(ConvBlock(ch, ch * 2, drop_prob))
            ch *= 2
        self.conv = ConvBlock(ch, ch * 2, drop_prob)

        self.up_conv = nn.ModuleList()
        self.up_transpose_conv = nn.ModuleList()
        for _ in range(num_pool_layers - 1):
            self.up_transpose_conv.append(TransposeConvBlock(ch * 2, ch))
            self.up_conv.append(ConvBlock(ch * 2, ch, drop_prob))
            ch //= 2

        self.up_transpose_conv.append(TransposeConvBlock(ch * 2, ch))
        self.up_conv.append(
            nn.Sequential(
                ConvBlock(ch * 2, ch, drop_prob),
                nn.Conv2d(ch, self.out_chans, kernel_size=1, stride=1),
            )
        )

    def forward(self, image: torch.Tensor) -> torch.Tensor:
        """
        Parameters
        ----------
        image : torch.Tensor
            Input 4D tensor of shape `(N, in_chans, H, W)`.

        Returns
        -------
        torch.Tensor
            Output tensor of shape `(N, out_chans, H, W)`.
        """
        stack = []
        output = image

        # apply down-sampling layers
        for layer in self.down_sample_layers:
            output = layer(output)
            stack.append(output)
            output = F.avg_pool2d(output, kernel_size=2, stride=2, padding=0)

        output = self.conv(output)

        # apply up-sampling layers
        for transpose_conv, conv in zip(self.up_transpose_conv, self.up_conv):
            downsample_layer = stack.pop()
            output = transpose_conv(output)

            # reflect pad on the right/botton if needed to handle odd input dimensions
            padding = [0, 0, 0, 0]
            if output.shape[-1] != downsample_layer.shape[-1]:
                padding[1] = 1  # padding right
            if output.shape[-2] != downsample_layer.shape[-2]:
                padding[3] = 1  # padding bottom
            if torch.sum(torch.tensor(padding)) != 0:
                output = F.pad(output, padding, "reflect")

            output = torch.cat([output, downsample_layer], dim=1)
            output = conv(output)

        return output


class ConvBlock(nn.Module):
    """
    A Convolutional Block that consists of two convolution layers each followed by
    instance normalization, LeakyReLU activation and dropout.
    """

    def __init__(self, in_chans: int, out_chans: int, drop_prob: float):
        """
        Parameters
        ----------
        in_chans : int
            Number of channels in the input.
        out_chans : int
            Number of channels in the output.
        drop_prob : float
            Dropout probability.
        """
        super().__init__()

        self.in_chans = in_chans
        self.out_chans = out_chans
        self.drop_prob = drop_prob

        self.layers = nn.Sequential(
            nn.Conv2d(in_chans, out_chans, kernel_size=3, padding=1, bias=False),
            nn.InstanceNorm2d(out_chans),
            nn.LeakyReLU(negative_slope=0.2, inplace=True),
            nn.Dropout2d(drop_prob),
            nn.Conv2d(out_chans, out_chans, kernel_size=3, padding=1, bias=False),
            nn.InstanceNorm2d(out_chans),
            nn.LeakyReLU(negative_slope=0.2, inplace=True),
            nn.Dropout2d(drop_prob),
        )

    def forward(self, image: torch.Tensor) -> torch.Tensor:
        """
        Parameters
        ----------
        image : ndarray
            Input 4D tensor of shape `(N, in_chans, H, W)`.

        Returns
        -------
        ndarray
            Output tensor of shape `(N, out_chans, H, W)`.
        """
        return self.layers(image)


class TransposeConvBlock(nn.Module):
    """
    A Transpose Convolutional Block that consists of one convolution transpose
    layers followed by instance normalization and LeakyReLU activation.
    """

    def __init__(self, in_chans: int, out_chans: int):
        """
        Parameters
        ----------
        in_chans : int
            Number of channels in the input.
        out_chans : int
            Number of channels in the output.
        """
        super().__init__()

        self.in_chans = in_chans
        self.out_chans = out_chans

        self.layers = nn.Sequential(
            nn.ConvTranspose2d(
                in_chans, out_chans, kernel_size=2, stride=2, bias=False
            ),
            nn.InstanceNorm2d(out_chans),
            nn.LeakyReLU(negative_slope=0.2, inplace=True),
        )

    def forward(self, image: torch.Tensor) -> torch.Tensor:
        """
        Parameters
        ----------
        image : torch.Tensor
            Input 4D tensor of shape `(N, in_chans, H, W)`.

        Returns
        -------
        torch.Tensor
            Output tensor of shape `(N, out_chans, H*2, W*2)`.
        """
        return self.layers(image)