summaryrefslogtreecommitdiff
path: root/modules/layers.py
blob: a0116e2fc6a66a97a788d0dc8ed3f4159bbbdb26 (plain)
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
211
212
213
214
215
216
217
218
219
220
import copy
from typing import Union, Tuple

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


class Encoder(nn.Module):
    def __init__(self, in_channels: int, opt):
        super(Encoder, self).__init__()
        self.opt = opt
        self.em_dim = opt.em_dim
        nf = 64

        # Cx[HxW]
        # Conv1 3x64x32 -> 64x64x32
        self.conv1 = nn.Conv2d(in_channels, nf, kernel_size=3, padding=1)
        self.batch_norm1 = nn.BatchNorm2d(nf)
        # MaxPool1 64x64x32 -> 64x32x16
        self.max_pool1 = nn.AdaptiveMaxPool2d((32, 16))
        # Conv2 64x32x16 -> 256x32x16
        self.conv2 = nn.Conv2d(nf, nf * 4, kernel_size=3, padding=1)
        self.batch_norm2 = nn.BatchNorm2d(nf * 4)
        # MaxPool2 256x32x16 -> 256x16x8
        self.max_pool2 = nn.AdaptiveMaxPool2d((16, 8))
        # Conv3 256x16x8 -> 512x16x8
        self.conv3 = nn.Conv2d(nf * 4, nf * 8, kernel_size=3, padding=1)
        self.batch_norm3 = nn.BatchNorm2d(nf * 8)
        # Conv4 512x16x8 -> 512x16x8 (for large dataset)
        self.conv4 = nn.Conv2d(nf * 8, nf * 8, kernel_size=3, padding=1)
        self.batch_norm4 = nn.BatchNorm2d(nf * 8)
        # MaxPool3 512x16x8 -> 512x4x2
        self.max_pool3 = nn.AdaptiveMaxPool2d((4, 2))
        # FC 512*4*2 -> 320
        self.fc = nn.Linear(nf * 8 * 2 * 4, self.em_dim)
        self.batch_norm_fc = nn.BatchNorm1d(self.em_dim)

    def forward(self, x):
        x = F.leaky_relu(self.batch_norm1(self.conv1(x)), 0.2)
        x = self.max_pool1(x)
        x = F.leaky_relu(self.batch_norm2(self.conv2(x)), 0.2)
        x = self.max_pool2(x)
        x = F.leaky_relu(self.batch_norm3(self.conv3(x)), 0.2)
        x = F.leaky_relu(self.batch_norm4(self.conv4(x)), 0.2)
        x = self.max_pool3(x)
        x = x.view(-1, (64 * 8) * 2 * 4)
        embedding = self.batch_norm_fc(self.fc(x))

        fa, fgs, fgd = embedding.split(
            (self.opt.fa_dim, self.opt.fg_dim / 2, self.opt.fg_dim / 2), dim=1
        )
        return fa, fgs, fgd


class Decoder(nn.Module):
    def __init__(self, out_channels: int, opt):
        super(Decoder, self).__init__()
        self.em_dim = opt.em_dim
        nf = 64

        # Cx[HxW]
        # FC 320 -> 512*4*2
        self.fc = nn.Linear(self.em_dim, nf * 8 * 2 * 4)
        self.batch_norm_fc = nn.BatchNorm1d(nf * 8 * 2 * 4)
        # TransConv1 512x4x2 -> 256x8x4
        self.trans_conv1 = nn.ConvTranspose2d(nf * 8, nf * 4, kernel_size=4,
                                              stride=2, padding=1)
        self.batch_norm1 = nn.BatchNorm2d(nf * 4)
        # TransConv2 256x8x4 -> 128x16x8
        self.trans_conv2 = nn.ConvTranspose2d(nf * 4, nf * 2, kernel_size=4,
                                              stride=2, padding=1)
        self.batch_norm2 = nn.BatchNorm2d(nf * 2)
        # TransConv3 128x16x8 -> 64x32x16
        self.trans_conv3 = nn.ConvTranspose2d(nf * 2, nf, kernel_size=4,
                                              stride=2, padding=1)
        self.batch_norm3 = nn.BatchNorm2d(nf)
        # TransConv4 3x32x16
        self.trans_conv4 = nn.ConvTranspose2d(nf, out_channels, kernel_size=4,
                                              stride=2, padding=1)

    def forward(self, fa, fgs, fgd):
        x = torch.cat((fa, fgs, fgd), dim=1).view(-1, self.em_dim)
        x = F.leaky_relu(self.batch_norm_fc(self.fc(x)), 0.2)
        x = F.leaky_relu(self.batch_norm1(self.trans_conv1(x)), 0.2)
        x = F.leaky_relu(self.batch_norm2(self.trans_conv2(x)), 0.2)
        x = F.leaky_relu(self.batch_norm3(self.trans_conv3(x)), 0.2)
        x = F.sigmoid(self.trans_conv4(x))

        return x


class FocalConv2d(nn.Module):
    def __init__(
            self,
            in_channels: int,
            out_channels: int,
            kernel_size: Union[int, Tuple[int, int]],
            halving: int,
            **kwargs
    ):
        super(FocalConv2d, self).__init__()
        self.halving = halving
        self.conv = nn.Conv2d(in_channels, out_channels, kernel_size,
                              bias=False, **kwargs)

    def forward(self, x):
        h = x.size(2)
        split_size = h // 2 ** self.halving
        z = x.split(split_size, dim=2)
        z = torch.cat([self.conv(_) for _ in z], dim=2)
        return F.leaky_relu(z, inplace=True)


class FrameLevelPartFeatureExtractor(nn.Module):

    def __init__(self, in_channels: int):
        super(FrameLevelPartFeatureExtractor, self).__init__()
        nf = 32

        self.focal_conv1 = FocalConv2d(in_channels, nf, kernel_size=5,
                                       padding=2, halving=1)
        self.focal_conv2 = FocalConv2d(nf, nf, kernel_size=3,
                                       padding=1, halving=1)
        self.focal_conv3 = FocalConv2d(nf, nf * 2, kernel_size=3,
                                       padding=1, halving=4)
        self.focal_conv4 = FocalConv2d(nf * 2, nf * 2, kernel_size=3,
                                       padding=1, halving=4)
        self.focal_conv5 = FocalConv2d(nf * 2, nf * 4, kernel_size=3,
                                       padding=1, halving=8)
        self.focal_conv6 = FocalConv2d(nf * 4, nf * 4, kernel_size=3,
                                       padding=1, halving=8)
        self.max_pool = nn.MaxPool2d(kernel_size=2, stride=2)

    def forward(self, x):
        x = F.leaky_relu(self.focal_conv1(x))
        x = F.leaky_relu(self.focal_conv2(x))
        x = self.max_pool(x)
        x = F.leaky_relu(self.focal_conv3(x))
        x = F.leaky_relu(self.focal_conv4(x))
        x = self.max_pool(x)
        x = F.leaky_relu(self.focal_conv5(x))
        x = F.leaky_relu(self.focal_conv6(x))

        return x


class TemporalFeatureAggregator(nn.Module):
    def __init__(
            self,
            in_channels: int,
            squeeze: int = 4,
            num_part: int = 16
    ):
        super(TemporalFeatureAggregator, self).__init__()
        hidden_dim = in_channels // squeeze
        self.num_part = num_part

        # MTB1
        conv3x1 = nn.Sequential(
            nn.Conv1d(in_channels, hidden_dim,
                      kernel_size=3, padding=1, bias=False),
            nn.LeakyReLU(inplace=True),
            nn.Conv1d(hidden_dim, in_channels,
                      kernel_size=1, padding=0, bias=False)
        )
        self.conv1d3x1 = self._parted(conv3x1)
        self.avg_pool3x1 = nn.AvgPool1d(kernel_size=3, stride=1, padding=1)
        self.max_pool3x1 = nn.MaxPool1d(kernel_size=3, stride=1, padding=1)

        # MTB2
        conv3x3 = nn.Sequential(
            nn.Conv1d(in_channels, hidden_dim,
                      kernel_size=3, padding=1, bias=False),
            nn.LeakyReLU(inplace=True),
            nn.Conv1d(hidden_dim, in_channels,
                      kernel_size=3, padding=1, bias=False)
        )
        self.conv1d3x3 = self._parted(conv3x3)
        self.avg_pool3x3 = nn.AvgPool1d(kernel_size=5, stride=1, padding=2)
        self.max_pool3x3 = nn.MaxPool1d(kernel_size=5, stride=1, padding=2)

    def _parted(self, module: nn.Module):
        """Duplicate module `part_num` times."""
        return nn.ModuleList([copy.deepcopy(module)
                              for _ in range(self.num_part)])

    def forward(self, x):
        """
          Input: x, [p, n, c, s]
        """
        p, n, c, s = x.size()
        feature = x.split(1, 0)
        x = x.view(-1, c, s)

        # MTB1: ConvNet1d & Sigmoid
        logits3x1 = torch.cat(
            [conv(_.squeeze(0)).unsqueeze(0)
             for conv, _ in zip(self.conv1d3x1, feature)], dim=0
        )
        scores3x1 = torch.sigmoid(logits3x1)
        # MTB1: Template Function
        feature3x1 = self.avg_pool3x1(x) + self.max_pool3x1(x)
        feature3x1 = feature3x1.view(p, n, c, s)
        feature3x1 = feature3x1 * scores3x1

        # MTB2: ConvNet1d & Sigmoid
        logits3x3 = torch.cat(
            [conv(_.squeeze(0)).unsqueeze(0)
             for conv, _ in zip(self.conv1d3x3, feature)], dim=0
        )
        scores3x3 = torch.sigmoid(logits3x3)
        # MTB2: Template Function
        feature3x3 = self.avg_pool3x3(x) + self.max_pool3x3(x)
        feature3x3 = feature3x3.view(p, n, c, s)
        feature3x3 = feature3x3 * scores3x3

        # Temporal Pooling
        ret = (feature3x1 + feature3x3).max(-1)[0]
        return ret