File size: 5,539 Bytes
2a13495
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import torch.nn as nn
import torch.nn.functional as F

from segmentation_models_pytorch.base import modules as md


class DecoderBlock(nn.Module):
    def __init__(
        self,
        in_channels,
        skip_channels,
        out_channels,
        use_batchnorm=True,
        attention_type=None,
    ):
        super().__init__()
        self.conv1 = md.Conv2dReLU(
            in_channels + skip_channels,
            out_channels,
            kernel_size=3,
            padding=1,
            use_batchnorm=use_batchnorm,
        )
        self.attention1 = md.Attention(
            attention_type, in_channels=in_channels + skip_channels
        )
        self.conv2 = md.Conv2dReLU(
            out_channels,
            out_channels,
            kernel_size=3,
            padding=1,
            use_batchnorm=use_batchnorm,
        )
        self.attention2 = md.Attention(attention_type, in_channels=out_channels)

    def forward(self, x, skip=None):
        x = F.interpolate(x, scale_factor=2, mode="nearest")
        if skip is not None:
            x = torch.cat([x, skip], dim=1)
            x = self.attention1(x)
        x = self.conv1(x)
        x = self.conv2(x)
        x = self.attention2(x)
        return x


class CenterBlock(nn.Sequential):
    def __init__(self, in_channels, out_channels, use_batchnorm=True):
        conv1 = md.Conv2dReLU(
            in_channels,
            out_channels,
            kernel_size=3,
            padding=1,
            use_batchnorm=use_batchnorm,
        )
        conv2 = md.Conv2dReLU(
            out_channels,
            out_channels,
            kernel_size=3,
            padding=1,
            use_batchnorm=use_batchnorm,
        )
        super().__init__(conv1, conv2)


class UnetPlusPlusDecoder(nn.Module):
    def __init__(
        self,
        encoder_channels,
        decoder_channels,
        n_blocks=5,
        use_batchnorm=True,
        attention_type=None,
        center=False,
    ):
        super().__init__()

        if n_blocks != len(decoder_channels):
            raise ValueError(
                "Model depth is {}, but you provide `decoder_channels` for {} blocks.".format(
                    n_blocks, len(decoder_channels)
                )
            )

        # remove first skip with same spatial resolution
        encoder_channels = encoder_channels[1:]
        # reverse channels to start from head of encoder
        encoder_channels = encoder_channels[::-1]

        # computing blocks input and output channels
        head_channels = encoder_channels[0]
        self.in_channels = [head_channels] + list(decoder_channels[:-1])
        self.skip_channels = list(encoder_channels[1:]) + [0]
        self.out_channels = decoder_channels
        if center:
            self.center = CenterBlock(
                head_channels, head_channels, use_batchnorm=use_batchnorm
            )
        else:
            self.center = nn.Identity()

        # combine decoder keyword arguments
        kwargs = dict(use_batchnorm=use_batchnorm, attention_type=attention_type)

        blocks = {}
        for layer_idx in range(len(self.in_channels) - 1):
            for depth_idx in range(layer_idx + 1):
                if depth_idx == 0:
                    in_ch = self.in_channels[layer_idx]
                    skip_ch = self.skip_channels[layer_idx] * (layer_idx + 1)
                    out_ch = self.out_channels[layer_idx]
                else:
                    out_ch = self.skip_channels[layer_idx]
                    skip_ch = self.skip_channels[layer_idx] * (
                        layer_idx + 1 - depth_idx
                    )
                    in_ch = self.skip_channels[layer_idx - 1]
                blocks[f"x_{depth_idx}_{layer_idx}"] = DecoderBlock(
                    in_ch, skip_ch, out_ch, **kwargs
                )
        blocks[f"x_{0}_{len(self.in_channels)-1}"] = DecoderBlock(
            self.in_channels[-1], 0, self.out_channels[-1], **kwargs
        )
        self.blocks = nn.ModuleDict(blocks)
        self.depth = len(self.in_channels) - 1

    def forward(self, *features):

        features = features[1:]  # remove first skip with same spatial resolution
        features = features[::-1]  # reverse channels to start from head of encoder
        # start building dense connections
        dense_x = {}
        for layer_idx in range(len(self.in_channels) - 1):
            for depth_idx in range(self.depth - layer_idx):
                if layer_idx == 0:
                    output = self.blocks[f"x_{depth_idx}_{depth_idx}"](
                        features[depth_idx], features[depth_idx + 1]
                    )
                    dense_x[f"x_{depth_idx}_{depth_idx}"] = output
                else:
                    dense_l_i = depth_idx + layer_idx
                    cat_features = [
                        dense_x[f"x_{idx}_{dense_l_i}"]
                        for idx in range(depth_idx + 1, dense_l_i + 1)
                    ]
                    cat_features = torch.cat(
                        cat_features + [features[dense_l_i + 1]], dim=1
                    )
                    dense_x[f"x_{depth_idx}_{dense_l_i}"] = self.blocks[
                        f"x_{depth_idx}_{dense_l_i}"
                    ](dense_x[f"x_{depth_idx}_{dense_l_i-1}"], cat_features)
        dense_x[f"x_{0}_{self.depth}"] = self.blocks[f"x_{0}_{self.depth}"](
            dense_x[f"x_{0}_{self.depth-1}"]
        )
        return dense_x[f"x_{0}_{self.depth}"]