fkl100/Behavior_and_Emotion_Recognition

1# File: models.py (continued)
2
3import torch
4import torch.nn as nn
5import torch.nn.functional as F
6import math
7
8class Bottleneck(nn.Module):
9    expansion = 4
10
11    def __init__(self, in_channels, out_channels, i_downsample=None, stride=1):
12        super(Bottleneck, self).__init__()
13        self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=stride, padding=0, bias=False)
14        self.batch_norm1 = nn.BatchNorm2d(out_channels, eps=0.001, momentum=0.99)
15        self.conv2 = nn.Conv2d(out_channels, out_channels, kernel_size=3, padding='same', bias=False)
16        self.batch_norm2 = nn.BatchNorm2d(out_channels, eps=0.001, momentum=0.99)
17        self.conv3 = nn.Conv2d(out_channels, out_channels * self.expansion, kernel_size=1, stride=1, padding=0, bias=False)
18        self.batch_norm3 = nn.BatchNorm2d(out_channels * self.expansion, eps=0.001, momentum=0.99)
19        self.i_downsample = i_downsample
20        self.stride = stride
21        self.relu = nn.ReLU()
22
23    def forward(self, x):
24        identity = x.clone()
25        x = self.relu(self.batch_norm1(self.conv1(x)))
26        x = self.relu(self.batch_norm2(self.conv2(x)))
27        x = self.conv3(x)
28        x = self.batch_norm3(x)
29
30        if self.i_downsample is not None:
31            identity = self.i_downsample(identity)
32        x += identity
33        x = self.relu(x)
34        return x
35
36class Conv2dSame(torch.nn.Conv2d):
37    def calc_same_pad(self, i: int, k: int, s: int, d: int) -> int:
38        return max((math.ceil(i / s) - 1) * s + (k - 1) * d + 1 - i, 0)
39
40    def forward(self, x: torch.Tensor) -> torch.Tensor:
41        ih, iw = x.size()[-2:]
42
43        pad_h = self.calc_same_pad(i=ih, k=self.kernel_size[0], s=self.stride[0], d=self.dilation[0])
44        pad_w = self.calc_same_pad(i=iw, k=self.kernel_size[1], s=self.stride[1], d=self.dilation[1])
45
46        if pad_h > 0 or pad_w > 0:
47            x = F.pad(x, [pad_w // 2, pad_w - pad_w // 2, pad_h // 2, pad_h - pad_h // 2])
48        return F.conv2d(x, self.weight, self.bias, self.stride, self.padding, self.dilation, self.groups)
49
50class ResNet(nn.Module):
51    def __init__(self, ResBlock, layer_list, num_classes, num_channels=3):
52        super(ResNet, self).__init__()
53        self.in_channels = 64
54
55        self.conv_layer_s2_same = Conv2dSame(num_channels, 64, 7, stride=2, groups=1, bias=False)
56        self.batch_norm1 = nn.BatchNorm2d(64, eps=0.001, momentum=0.99)
57        self.relu = nn.ReLU()
58        self.max_pool = nn.MaxPool2d(kernel_size=3, stride=2)
59
60        self.layer1 = self._make_layer(ResBlock, layer_list[0], planes=64, stride=1)
61        self.layer2 = self._make_layer(ResBlock, layer_list[1], planes=128, stride=2)
62        self.layer3 = self._make_layer(ResBlock, layer_list[2], planes=256, stride=2)
63        self.layer4 = self._make_layer(ResBlock, layer_list[3], planes=512, stride=2)
64
65        self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
66        self.fc1 = nn.Linear(512 * ResBlock.expansion, 512)
67        self.relu1 = nn.ReLU()
68        self.fc2 = nn.Linear(512, num_classes)
69
70    def extract_features(self, x):
71        x = self.relu(self.batch_norm1(self.conv_layer_s2_same(x)))
72        x = self.max_pool(x)
73        x = self.layer1(x)
74        x = self.layer2(x)
75        x = self.layer3(x)
76        x = self.layer4(x)
77        x = self.avgpool(x)
78        x = x.reshape(x.shape[0], -1)
79        x = self.fc1(x)
80        return x
81
82    def forward(self, x):
83        x = self.extract_features(x)
84        x = self.relu1(x)
85        x = self.fc2(x)
86        return x
87
88    def _make_layer(self, ResBlock, blocks, planes, stride=1):
89        ii_downsample = None
90        layers = []
91
92        if stride != 1 or self.in_channels != planes * ResBlock.expansion:
93            ii_downsample = nn.Sequential(
94                nn.Conv2d(self.in_channels, planes * ResBlock.expansion, kernel_size=1, stride=stride, bias=False, padding=0),
95                nn.BatchNorm2d(planes * ResBlock.expansion, eps=0.001, momentum=0.99)
96            )
97
98        layers.append(ResBlock(self.in_channels, planes, i_downsample=ii_downsample, stride=stride))
99        self.in_channels = planes * ResBlock.expansion
100
101        for i in range(blocks - 1):
102            layers.append(ResBlock(self.in_channels, planes))
103
104        return nn.Sequential(*layers)
105
106def ResNet50(num_classes, channels=3):
107    return ResNet(Bottleneck, [3, 4, 6, 3], num_classes, channels)
108
109class LSTMPyTorch(nn.Module):
110    def __init__(self):
111        super(LSTMPyTorch, self).__init__()
112        self.lstm1 = nn.LSTM(input_size=512, hidden_size=512, batch_first=True, bidirectional=False)
113        self.lstm2 = nn.LSTM(input_size=512, hidden_size=256, batch_first=True, bidirectional=False)
114        self.fc = nn.Linear(256, 7)
115        self.softmax = nn.Softmax(dim=1)
116
117    def forward(self, x):
118        x, _ = self.lstm1(x)
119        x, _ = self.lstm2(x)
120        x = self.fc(x[:, -1, :])
121        x = self.softmax(x)
122        return x