SlowFast/model.py at master · leaderj1001/SlowFast · 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
197
198
199
200
201
202
203
204
import torch
import torch.nn as nn
import torch.nn.functional as F


# Reference
# https://github.com/kenshohara/3D-ResNets-PyTorch/blob/master/models/resnet.py
class Bottleneck(nn.Module):
    expansion = 4

    def __init__(self, in_planes, planes, stride=1, downsample=None, block_type='spatial', bias=False):
        super().__init__()
        if block_type == 'spatial':
            self.conv1 = nn.Conv3d(in_planes, planes, kernel_size=1, bias=bias)
        else:
            self.conv1 = nn.Conv3d(in_planes, planes, kernel_size=(3, 1, 1), padding=(1, 0, 0), bias=bias)
        self.bn1 = nn.BatchNorm3d(planes)
        self.conv2 = nn.Conv3d(planes, planes, kernel_size=(1, 3, 3), stride=stride, padding=(0, 1, 1), bias=bias)
        self.bn2 = nn.BatchNorm3d(planes)
        self.conv3 = nn.Conv3d(planes, planes * self.expansion, kernel_size=1, bias=bias)
        self.bn3 = nn.BatchNorm3d(planes * self.expansion)
        self.relu = nn.ReLU(inplace=True)
        self.downsample = downsample
        self.stride = stride

    def forward(self, x):
        residual = x

        out = self.conv1(x)
        out = self.bn1(out)
        out = self.relu(out)

        out = self.conv2(out)
        out = self.bn2(out)
        out = self.relu(out)

        out = self.conv3(out)
        out = self.bn3(out)

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

        out += residual
        out = self.relu(out)

        return out


class SlowFast(nn.Module):
    def __init__(self, block=Bottleneck, layers=[3, 4, 6, 8], n_classes=101, lateral_type='time_strided_conv', alpha=8, beta=8):
        super(SlowFast, self).__init__()
        self.alpha, self.beta = alpha, beta
        self.lateral_type = lateral_type

        self.slow_stem = nn.Sequential(
            nn.Conv3d(3, 64, kernel_size=(1, 7, 7), stride=(1, 2, 2), padding=(0, 3, 3)),
            nn.BatchNorm3d(64),
            nn.ReLU(True)
        )

        if lateral_type == 'time_strided_conv':
            self.slow_in_planes = 64 + 64 // self.beta * 2
            self.lateral_connection1 = nn.Conv3d(8, 8 * 2, kernel_size=(5, 1, 1), stride=(self.alpha, 1, 1), padding=(2, 0, 0))
            self.lateral_connection2 = nn.Conv3d(32, 32 * 2, kernel_size=(5, 1, 1), stride=(self.alpha, 1, 1), padding=(2, 0, 0))
            self.lateral_connection3 = nn.Conv3d(64, 64 * 2, kernel_size=(5, 1, 1), stride=(self.alpha, 1, 1), padding=(2, 0, 0))
            self.lateral_connection4 = nn.Conv3d(128, 128 * 2, kernel_size=(5, 1, 1), stride=(self.alpha, 1, 1), padding=(2, 0, 0))
        elif lateral_type == 'time2channel':
            self.slow_in_planes = 64 + 64 // self.beta * self.alpha
        elif lateral_type == 'time_strided_sampling':
            self.slow_in_planes = 64 + 64 // self.beta

        self.slow_pool = nn.MaxPool3d(kernel_size=(1, 3, 3), stride=(1, 2, 2), padding=(0, 1, 1))
        self.slow_layer1 = self._make_layer(block, 64, layers[0], stride=1, block_type='spatial', path='slow')
        self.slow_layer2 = self._make_layer(block, 128, layers[1], stride=(1, 2, 2), block_type='spatial', path='slow')
        self.slow_layer3 = self._make_layer(block, 256, layers[2], stride=(1, 2, 2), block_type='temporal', path='slow')
        self.slow_layer4 = self._make_layer(block, 512, layers[3], stride=(1, 2, 2), block_type='temporal', path='slow')

        self.fast_in_planes = 8
        self.fast_stem = nn.Sequential(
            nn.Conv3d(3, 8, kernel_size=(5, 7, 7), stride=(1, 2, 2), padding=(2, 3, 3)),
            nn.BatchNorm3d(8),
            nn.ReLU(True)
        )
        self.fast_pool = nn.MaxPool3d(kernel_size=(1, 3, 3), stride=(1, 2, 2), padding=(0, 1, 1))
        self.fast_layer1 = self._make_layer(block, 8, layers[0], stride=1, block_type='temporal', path='fast')
        self.fast_layer2 = self._make_layer(block, 16, layers[1], stride=(1, 2, 2), block_type='temporal', path='fast')
        self.fast_layer3 = self._make_layer(block, 32, layers[2], stride=(1, 2, 2), block_type='temporal', path='fast')
        self.fast_layer4 = self._make_layer(block, 64, layers[3], stride=(1, 2, 2), block_type='temporal', path='fast')

        self.fast_gap = nn.AdaptiveAvgPool3d(1)
        self.slow_gap = nn.AdaptiveAvgPool3d(1)

        self.fc = nn.Sequential(
            nn.Linear(2048 + self.fast_in_planes, n_classes)
        )

    def forward(self, x):
        slow_x, fast_x = self._data_layer(x)

        slow_out, fast_out = self.slow_stem(slow_x), self.fast_stem(fast_x)
        slow_out, fast_out = self.slow_pool(slow_out), self.fast_pool(fast_out)

        fast_out1 = self.fast_layer1(fast_out)
        fast_out2 = self.fast_layer2(fast_out1)
        fast_out3 = self.fast_layer3(fast_out2)
        fast_out4 = self.fast_layer4(fast_out3)

        if self.lateral_type == 'time_strided_conv':
            lateral_out1 = self.lateral_connection1(fast_out)
            lateral_out2 = self.lateral_connection2(fast_out1)
            lateral_out3 = self.lateral_connection3(fast_out2)
            lateral_out4 = self.lateral_connection4(fast_out3)
        elif self.lateral_type == 'time2channel':
            batch, C, temporal, width, height = fast_out.size()
            lateral_out1 = fast_out.view(batch, -1, temporal // self.alpha, width, height)
            batch, C, temporal, width, height = fast_out1.size()
            lateral_out2 = fast_out1.view(batch, -1, temporal // self.alpha, width, height)
            batch, C, temporal, width, height = fast_out2.size()
            lateral_out3 = fast_out2.view(batch, -1, temporal // self.alpha, width, height)
            batch, C, temporal, width, height = fast_out3.size()
            lateral_out4 = fast_out3.view(batch, -1, temporal // self.alpha, width, height)
        elif self.lateral_type == 'time_strided_sampling':
            lateral_out1 = fast_out[:, :, ::self.alpha, :, :]
            lateral_out2 = fast_out1[:, :, ::self.alpha, :, :]
            lateral_out3 = fast_out2[:, :, ::self.alpha, :, :]
            lateral_out4 = fast_out3[:, :, ::self.alpha, :, :]

        slow_out = self.slow_layer1(torch.cat([slow_out, lateral_out1], dim=1))
        slow_out = self.slow_layer2(torch.cat([slow_out, lateral_out2], dim=1))
        slow_out = self.slow_layer3(torch.cat([slow_out, lateral_out3], dim=1))
        slow_out = self.slow_layer4(torch.cat([slow_out, lateral_out4], dim=1))

        fast_out = self.fast_gap(fast_out4)
        slow_out = self.slow_gap(slow_out)

        out = torch.cat([slow_out, fast_out], dim=1).view(x.size(0), -1)
        out = self.fc(out)

        return out

    def _make_layer(self, block, planes, blocks, stride=(1, 1, 1), block_type='spatial', path='slow'):
        downsample = None
        layers = []

        if path == 'slow':
            if stride != 1 or self.slow_in_planes != planes * block.expansion:
                downsample = nn.Sequential(
                    nn.Conv3d(self.slow_in_planes, planes * block.expansion, kernel_size=1, stride=stride),
                    nn.BatchNorm3d(planes * block.expansion))

            layers.append(
                block(in_planes=self.slow_in_planes,
                      planes=planes,
                      stride=stride,
                      downsample=downsample,
                      block_type=block_type))
            self.slow_in_planes = planes * block.expansion
            for i in range(1, blocks):
                layers.append(block(self.slow_in_planes, planes))
            if self.lateral_type == 'time_strided_conv':
                self.slow_in_planes += 2 * planes * block.expansion // self.beta
            elif self.lateral_type == 'time2channel':
                self.slow_in_planes += planes * block.expansion // self.beta * self.alpha
            elif self.lateral_type == 'time_strided_sampling':
                self.slow_in_planes += planes * block.expansion // self.beta
        else:
            if stride != 1 or self.fast_in_planes != planes * block.expansion:
                downsample = nn.Sequential(
                    nn.Conv3d(self.fast_in_planes, planes * block.expansion, kernel_size=1, stride=stride),
                    nn.BatchNorm3d(planes * block.expansion))

            layers.append(
                block(in_planes=self.fast_in_planes,
                      planes=planes,
                      stride=stride,
                      downsample=downsample,
                      block_type=block_type))
            self.fast_in_planes = planes * block.expansion
            for i in range(1, blocks):
                layers.append(block(self.fast_in_planes, planes))

        return nn.Sequential(*layers)

    def _data_layer(self, x):
        # We opt to not perform temporal downsampling in this instantiation, as doing so would be detrimental when the input stride is large.
        slow_data = x[:, :, ::16, :, :]
        fast_data = x[:, :, ::2, :, :]

        return slow_data, fast_data


def ResNet50(block=Bottleneck, layers=[3, 4, 6, 8], n_classes=101):
    return SlowFast(block=block, layers=layers, n_classes=n_classes)


def main():
    sf = SlowFast()
    x = torch.randn([2, 3, 64, 224, 224])

    sf(x)


if __name__ == '__main__':
    main()