Image/MobileNetv3/code/model.py

'''
MobileNetV3 in PyTorch.

论文: "Searching for MobileNetV3"
参考: https://arxiv.org/abs/1905.02244

主要特点：
1. 引入基于NAS的网络架构搜索
2. 使用改进的SE注意力机块
3. 使用h-swish激活函数
4. 重新设计了网络的最后几层
5. 提供了Large和Small两个版本
'''

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


def get_activation(name):
    '''获取激活函数
    
    Args:
        name: 激活函数名称 ('relu' 或 'hardswish')
    '''
    if name == 'relu':
        return nn.ReLU(inplace=True)
    elif name == 'hardswish':
        return nn.Hardswish(inplace=True)
    else:
        raise NotImplementedError


class SEModule(nn.Module):
    '''Squeeze-and-Excitation模块
    
    通过全局平均池化和两层全连接网络学习通道注意力权重
    
    Args:
        channel: 输入通道数
        reduction: 降维比例
    '''
    def __init__(self, channel, reduction=4):
        super(SEModule, self).__init__()
        self.avg_pool = nn.AdaptiveAvgPool2d(1)
        self.fc = nn.Sequential(
            nn.Linear(channel, channel // reduction, bias=False),
            nn.ReLU(inplace=True),
            nn.Linear(channel // reduction, channel, bias=False),
            nn.Hardsigmoid(inplace=True)
        )

    def forward(self, x):
        b, c, _, _ = x.size()
        y = self.avg_pool(x).view(b, c)  # squeeze
        y = self.fc(y).view(b, c, 1, 1)  # excitation
        return x * y.expand_as(x)  # scale


class Bottleneck(nn.Module):
    '''MobileNetV3 Bottleneck
    
    包含:
    1. Expansion layer (1x1 conv)
    2. Depthwise layer (3x3 or 5x5 depthwise conv)
    3. SE module (optional)
    4. Projection layer (1x1 conv)
    
    Args:
        in_channels: 输入通道数
        exp_channels: 扩展层通道数
        out_channels: 输出通道数
        kernel_size: 深度卷积核大小
        stride: 步长
        use_SE: 是否使用SE模块
        activation: 激活函数类型
        use_residual: 是否使用残差连接
    '''
    def __init__(self, in_channels, exp_channels, out_channels, kernel_size, 
                 stride, use_SE, activation, use_residual=True):
        super(Bottleneck, self).__init__()
        self.use_residual = use_residual and stride == 1 and in_channels == out_channels
        padding = (kernel_size - 1) // 2

        layers = []
        # Expansion layer
        if exp_channels != in_channels:
            layers.extend([
                nn.Conv2d(in_channels, exp_channels, 1, bias=False),
                nn.BatchNorm2d(exp_channels),
                get_activation(activation)
            ])
        
        # Depthwise conv
        layers.extend([
            nn.Conv2d(
                exp_channels, exp_channels, kernel_size, 
                stride, padding, groups=exp_channels, bias=False
            ),
            nn.BatchNorm2d(exp_channels),
            get_activation(activation)
        ])

        # SE module
        if use_SE:
            layers.append(SEModule(exp_channels))

        # Projection layer
        layers.extend([
            nn.Conv2d(exp_channels, out_channels, 1, bias=False),
            nn.BatchNorm2d(out_channels)
        ])

        self.conv = nn.Sequential(*layers)

    def forward(self, x):
        if self.use_residual:
            return x + self.conv(x)
        else:
            return self.conv(x)


class MobileNetV3(nn.Module):
    '''MobileNetV3网络
    
    Args:
        num_classes: 分类数量
        mode: 'large' 或 'small'，选择网络版本
    '''
    def __init__(self, num_classes=10, mode='small'):
        super(MobileNetV3, self).__init__()
        
        if mode == 'large':
            # MobileNetV3-Large架构
            self.config = [
                # k, exp, out, SE, activation, stride
                [3, 16, 16, False, 'relu', 1],
                [3, 64, 24, False, 'relu', 2],
                [3, 72, 24, False, 'relu', 1],
                [5, 72, 40, True, 'relu', 2],
                [5, 120, 40, True, 'relu', 1],
                [5, 120, 40, True, 'relu', 1],
                [3, 240, 80, False, 'hardswish', 2],
                [3, 200, 80, False, 'hardswish', 1],
                [3, 184, 80, False, 'hardswish', 1],
                [3, 184, 80, False, 'hardswish', 1],
                [3, 480, 112, True, 'hardswish', 1],
                [3, 672, 112, True, 'hardswish', 1],
                [5, 672, 160, True, 'hardswish', 2],
                [5, 960, 160, True, 'hardswish', 1],
                [5, 960, 160, True, 'hardswish', 1],
            ]
            init_conv_out = 16
            final_conv_out = 960
        else:
            # MobileNetV3-Small架构
            self.config = [
                # k, exp, out, SE, activation, stride
                [3, 16, 16, True, 'relu', 2],
                [3, 72, 24, False, 'relu', 2],
                [3, 88, 24, False, 'relu', 1],
                [5, 96, 40, True, 'hardswish', 2],
                [5, 240, 40, True, 'hardswish', 1],
                [5, 240, 40, True, 'hardswish', 1],
                [5, 120, 48, True, 'hardswish', 1],
                [5, 144, 48, True, 'hardswish', 1],
                [5, 288, 96, True, 'hardswish', 2],
                [5, 576, 96, True, 'hardswish', 1],
                [5, 576, 96, True, 'hardswish', 1],
            ]
            init_conv_out = 16
            final_conv_out = 576

        # 第一层卷积
        self.conv_stem = nn.Sequential(
            nn.Conv2d(3, init_conv_out, 3, 2, 1, bias=False),
            nn.BatchNorm2d(init_conv_out),
            get_activation('hardswish')
        )

        # 构建Bottleneck层
        features = []
        in_channels = init_conv_out
        for k, exp, out, se, activation, stride in self.config:
            features.append(
                Bottleneck(in_channels, exp, out, k, stride, se, activation)
            )
            in_channels = out
        self.features = nn.Sequential(*features)

        # 最后的卷积层
        self.conv_head = nn.Sequential(
            nn.Conv2d(in_channels, final_conv_out, 1, bias=False),
            nn.BatchNorm2d(final_conv_out),
            get_activation('hardswish')
        )

        # 分类器
        self.avgpool = nn.AdaptiveAvgPool2d(1)
        self.classifier = nn.Sequential(
            nn.Linear(final_conv_out, num_classes)
        )

        # 初始化权重
        self._initialize_weights()

    def _initialize_weights(self):
        '''初始化模型权重'''
        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                nn.init.kaiming_normal_(m.weight, mode='fan_out')
                if m.bias is not None:
                    nn.init.zeros_(m.bias)
            elif isinstance(m, nn.BatchNorm2d):
                nn.init.ones_(m.weight)
                nn.init.zeros_(m.bias)
            elif isinstance(m, nn.Linear):
                nn.init.normal_(m.weight, 0, 0.01)
                if m.bias is not None:
                    nn.init.zeros_(m.bias)

    def forward(self, x):
        x = self.conv_stem(x)
        x = self.features(x)
        x = self.conv_head(x)
        x = self.avgpool(x)
        x = x.view(x.size(0), -1)
        x = self.classifier(x)
        return x


def test():
    """测试函数"""
    # 测试Large版本
    net_large = MobileNetV3(mode='large')
    x = torch.randn(2, 3, 32, 32)
    y = net_large(x)
    print('Large output size:', y.size())
    
    # 测试Small版本
    net_small = MobileNetV3(mode='small')
    y = net_small(x)
    print('Small output size:', y.size())
    
    # 打印模型结构
    from torchinfo import summary
    device = 'cuda' if torch.cuda.is_available() else 'cpu'
    net_small = net_small.to(device)
    summary(net_small, (2, 3, 32, 32))

if __name__ == '__main__':
    test()