CSP Darknet53

CSP Darknet53

代码复现:这里的代码参考CSDN@Bubbliiiing

在之前yolo v3的实战篇中，我们了解到yolo v3 使用的backbone是Darknet53，而今天要展现的是yolo v4的backbone CSP Darknet53。

他们有什么不同呢？

1. 激活函数的改变，之前Darknet53使用的是LeakyReLU，而CSP Darknet53使用的是Mish。

LeakyReLU的图像

LeakyReLU是分段函数，当x>0时， f(x) = x
				  当x<=0时， f(x) = ax

Mish的图像是

f(x) = x * tanh(softplus(x)) = x * tanh(ln(1+exp(x)))

两者并没有很大的差距，但是下面Mish的计算量相对来说会大一些,效果可能会好些

2. backbone从Darknet53变为CSP Darknet53，主要体现在

   

   从上图上看,CSP在保持原来的Bottleneck的基础上，利用卷积的方式将输入的特征数据分为2个部分(split_conv0, split_conv1),这两部分其中一部分做resnet的残差卷积，然后在做1x1卷积，最后与另一部分进行cat拼接操作。

import math
import torch
from torch import nn
import torch.nn.functional as F


class Mish(nn.Module):
    '''
    # 实现Mish函数
    Mish公式为 f(x) = x * tanh(softplus(x)) = x * tanh(ln(1+exp(x)))
    '''

    def __init__(self):
        super(Mish, self).__init__()

    def forward(self, x):
        return x * torch.tanh(F.softplus(x))


class BasicConv(nn.Module):
    '''
    实现基本的卷积层结构: conv -> BN -> activation
    python: // 表示结果向下取整   / 表示结果向上取整   6 // 4 == 1   6 / 4 == 1.5
    '''

    def __init__(self, in_channels, out_channels, kernel_size, stride=1):
        super(BasicConv, self).__init__()
        self.conv = nn.Conv2d(in_channels=in_channels, out_channels=out_channels,
                              kernel_size=kernel_size, stride=(stride, stride), padding=kernel_size // 2, bias=False)
        self.bn = nn.BatchNorm2d(out_channels)
        self.activation = Mish()

    def forward(self, x):
        x = self.conv(x)
        x = self.bn(x)
        x = self.activation(x)
        return x


class Resblock(nn.Module):
    '''
    和resnet 一样，做残差结构
    '''
    def __init__(self, channels, hidden_channels=None):
        super(Resblock, self).__init__()
        if hidden_channels is None:
            hidden_channels = channels

        self.block = nn.Sequential(
            BasicConv(channels, hidden_channels, 1),
            BasicConv(hidden_channels, channels, 3)
        )

    def forward(self, x):
        return x + self.block(x)


class Resblock_body(nn.Module):
    '''
    这里与darknet53最大的不同是:
        每一次Resblock_body 都会采用csp结构，将输入利用卷积的方式分为2个部分(split_conv0, split_conv1),这两部分其中一部分会和原始darknet一样，做resnet的残差卷积，然后在做1x1卷积，最后与另一部分进行cat拼接操作，很像残差结构，但并不是，残差是各个位置相加，而这里是cat，拼接操作
    '''
    def __init__(self, in_channels, out_channels, num_blocks, first):
        super(Resblock_body, self).__init__()
        self.downsample_conv = BasicConv(in_channels=in_channels, out_channels=out_channels,
                                         kernel_size=3, stride=2)
        if first:
            self.split_conv0 = BasicConv(in_channels=out_channels, out_channels=out_channels,
                                         kernel_size=1)
            self.split_conv1 = BasicConv(in_channels=out_channels, out_channels=out_channels,
                                         kernel_size=1)
            self.blocks_conv = nn.Sequential(
                Resblock(channels=out_channels, hidden_channels=out_channels // 2),
                BasicConv(in_channels=out_channels, out_channels=out_channels, kernel_size=1)
            )
            self.concat_conv = BasicConv(in_channels=out_channels * 2, out_channels=out_channels,
                                         kernel_size=1)
        else:
            self.split_conv0 = BasicConv(in_channels=out_channels,
                                         out_channels=out_channels // 2,
                                         kernel_size=1)
            self.split_conv1 = BasicConv(in_channels=out_channels,
                                         out_channels=out_channels // 2,
                                         kernel_size=1)
            self.blocks_conv = nn.Sequential(
                *[Resblock(channels=out_channels // 2) for _ in range(num_blocks)],
                BasicConv(in_channels=out_channels // 2, out_channels=out_channels // 2,
                          kernel_size=1)
            )
            self.concat_conv = BasicConv(in_channels=out_channels, out_channels=out_channels,
                                         kernel_size=1)

    def forward(self, x):
        x = self.downsample_conv(x)

        x0 = self.split_conv0(x)

        x1 = self.split_conv1(x)
        x1 = self.blocks_conv(x1)
        x = torch.cat([x1, x0], dim=1)
        x = self.concat_conv(x)

        return x


class CSPDarknet(nn.Module):
    def __init__(self, layers):
        super(CSPDarknet, self).__init__()
        self.inplanes = 32
        self.conv1 = BasicConv(in_channels=3, out_channels=self.inplanes, kernel_size=3,
                               stride=1)
        self.feature_channels = [64, 128, 256, 512, 1024]

        self.stages = nn.ModuleList([
            Resblock_body(self.inplanes, self.feature_channels[0], layers[0], first=True),
            Resblock_body(self.feature_channels[0], self.feature_channels[1], layers[1], first=False),
            Resblock_body(self.feature_channels[1], self.feature_channels[2], layers[2], first=False),
            Resblock_body(self.feature_channels[2], self.feature_channels[3], layers[3], first=False),
            Resblock_body(self.feature_channels[3], self.feature_channels[4], layers[4], first=False)
        ])

        self.num_features = 1
        # 进行权值初始化
        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
                m.weight.data.normal_(0, math.sqrt(2. / n))
            elif isinstance(m, nn.BatchNorm2d):
                m.weight.data.fill_(1)
                m.bias.data.zero_()

    def forward(self, x):
        x = self.conv1(x)

        x = self.stages[0](x)
        x = self.stages[1](x)
        out3 = self.stages[2](x)
        out4 = self.stages[3](out3)
        out5 = self.stages[4](out4)

        return out3, out4, out5


def darknet53(pretrained, activation=None):
    model = CSPDarknet([1, 2, 8, 8, 4])
    if pretrained:
        if isinstance(pretrained, str):
            model.load_state_dict(torch.load(pretrained))
        else:
            raise Exception("darknet request a pretrained path. got [{}]".format(pretrained))
    return model


if __name__ == '__main__':
    model = darknet53(pretrained=False)
    from torchsummary import summary

    print(summary(model, (3, 416, 416)))
.format(pretrained))
    return model


if __name__ == '__main__':
    model = darknet53(pretrained=False)
    from torchsummary import summary

    print(summary(model, (3, 416, 416)))