File size: 5,354 Bytes
8ed2f16
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# PD-FGC motion encoder, modified from https://github.com/Dorniwang/PD-FGC-inference
import torch
import torch.nn as nn
import torch.nn.functional as F


def conv3x3(in_planes, out_planes, strd=1, padding=1, bias=False):
    "3x3 convolution with padding"
    return nn.Conv2d(in_planes, out_planes, kernel_size=3,
                     stride=strd, padding=padding, bias=bias)


class ConvBlock(nn.Module):
    def __init__(self, in_planes, out_planes):
        super(ConvBlock, self).__init__()
        self.bn1 = nn.BatchNorm2d(in_planes)
        self.conv1 = conv3x3(in_planes, int(out_planes / 2))
        self.bn2 = nn.BatchNorm2d(int(out_planes / 2))
        self.conv2 = conv3x3(int(out_planes / 2), int(out_planes / 4))
        self.bn3 = nn.BatchNorm2d(int(out_planes / 4))
        self.conv3 = conv3x3(int(out_planes / 4), int(out_planes / 4))

        if in_planes != out_planes:
            self.downsample = nn.Sequential(
                nn.BatchNorm2d(in_planes),
                nn.ReLU(True),
                nn.Conv2d(in_planes, out_planes,
                          kernel_size=1, stride=1, bias=False),
            )
        else:
            self.downsample = None

    def forward(self, x):
        residual = x

        out1 = self.bn1(x)
        out1 = F.relu(out1, True)
        out1 = self.conv1(out1)

        out2 = self.bn2(out1)
        out2 = F.relu(out2, True)
        out2 = self.conv2(out2)

        out3 = self.bn3(out2)
        out3 = F.relu(out3, True)
        out3 = self.conv3(out3)

        out3 = torch.cat((out1, out2, out3), 1)

        if self.downsample is not None:
            residual = self.downsample(residual)

        out3 += residual

        return out3


class HourGlass(nn.Module):
    def __init__(self, num_modules, depth, num_features):
        super(HourGlass, self).__init__()
        self.num_modules = num_modules
        self.depth = depth
        self.features = num_features
        self.dropout = nn.Dropout(0.5)

        self._generate_network(self.depth)

    def _generate_network(self, level):
        self.add_module('b1_' + str(level), ConvBlock(256, 256))

        self.add_module('b2_' + str(level), ConvBlock(256, 256))

        if level > 1:
            self._generate_network(level - 1)
        else:
            self.add_module('b2_plus_' + str(level), ConvBlock(256, 256))

        self.add_module('b3_' + str(level), ConvBlock(256, 256))

    def _forward(self, level, inp):
        # Upper branch
        up1 = inp
        up1 = self._modules['b1_' + str(level)](up1)
        up1 = self.dropout(up1)
        # Lower branch
        low1 = F.max_pool2d(inp, 2, stride=2)
        low1 = self._modules['b2_' + str(level)](low1)

        if level > 1:
            low2 = self._forward(level - 1, low1)
        else:
            low2 = low1
            low2 = self._modules['b2_plus_' + str(level)](low2)

        low3 = low2
        low3 = self._modules['b3_' + str(level)](low3)
        up1size = up1.size()
        rescale_size = (up1size[2], up1size[3])
        up2 = F.upsample(low3, size=rescale_size, mode='bilinear')

        return up1 + up2

    def forward(self, x):
        return self._forward(self.depth, x)


class FAN_use(nn.Module):
    def __init__(self):
        super(FAN_use, self).__init__()
        self.num_modules = 1

        # Base part
        self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3)
        self.bn1 = nn.BatchNorm2d(64)
        self.conv2 = ConvBlock(64, 128)
        self.conv3 = ConvBlock(128, 128)
        self.conv4 = ConvBlock(128, 256)

        # Stacking part
        hg_module = 0
        self.add_module('m' + str(hg_module), HourGlass(1, 4, 256))
        self.add_module('top_m_' + str(hg_module), ConvBlock(256, 256))
        self.add_module('conv_last' + str(hg_module),
                        nn.Conv2d(256, 256, kernel_size=1, stride=1, padding=0))
        self.add_module('l' + str(hg_module), nn.Conv2d(256,
                                                        68, kernel_size=1, stride=1, padding=0))
        self.add_module('bn_end' + str(hg_module), nn.BatchNorm2d(256))

        if hg_module < self.num_modules - 1:
            self.add_module(
                'bl' + str(hg_module), nn.Conv2d(256, 256, kernel_size=1, stride=1, padding=0))
            self.add_module('al' + str(hg_module), nn.Conv2d(68,
                                                             256, kernel_size=1, stride=1, padding=0))

        self.avgpool = nn.MaxPool2d((2, 2), 2)
        self.conv6 = nn.Conv2d(68, 1, 3, 2, 1)
        self.fc = nn.Linear(28 * 28, 512)
        self.bn5 = nn.BatchNorm2d(68)
        self.relu = nn.ReLU(True)

    def forward(self, x):
        x = F.relu(self.bn1(self.conv1(x)), True)
        x = F.max_pool2d(self.conv2(x), 2)
        x = self.conv3(x)
        x = self.conv4(x)

        previous = x

        i = 0
        hg = self._modules['m' + str(i)](previous)

        ll = hg
        ll = self._modules['top_m_' + str(i)](ll)

        ll = self._modules['bn_end' + str(i)](self._modules['conv_last' + str(i)](ll))
        tmp_out = self._modules['l' + str(i)](F.relu(ll))

        net = self.relu(self.bn5(tmp_out))
        net = self.conv6(net)
        net = net.view(-1, net.shape[-2] * net.shape[-1])
        net = self.relu(net)
        net = self.fc(net)
        return net