CBNetV2：复合骨干架构

前言

今天我们基于CBNetV2: A Composite Backbone Network Architecture for Object Detection这篇文章来复现其中的复合骨干部分。类似于CBNetV2类的网络相当于爸妈模型变宽了，重复利用特征，对目标检测有一定的优势。

模型结构

模型结构如图所示，方框中的我们就先不进行复现了，主要复现这种复合结构。
这种复合结构有两种一种是Backbone 1，另一种Backbone K

代码实现

在Backbone 1中无输入递归结构，输出分别是x2到x5的特征，所以结构可以表示为：

class First_Backbone(nn.Module): #定义上采样函数
    def __init__(self, in_channels, in_channel1, out_channels, bilinear=False):
        super().__init__()
        self.layer1 = Conv(in_channels, out_channels) 
        self.layer2 = Down(out_channels, out_channels) 
        self.layer3 = Down(out_channels, out_channels) 
        self.layer4 = Down(out_channels, out_channels) 

    def forward(self, x1): #x是设置的参数，代表输入数据
        x2 =  self.layer1(x1)
        x3 =  self.layer2(x2)
        x4 =  self.layer3(x3)
        x5 =  self.layer4(x4)
        return x2, x3, x4, x5

N Backbone比较复杂，他需要前面的x2到x5的特征，并且通过反卷积复原：

class N_Backbone(nn.Module): #定义上采样函数
    def __init__(self, in_channels, in_channel1, out_channels, bilinear=False):
        super().__init__()
        self.layer1 = Conv(in_channels, out_channels) 
        self.layer2 = Down(out_channels, out_channels) 
        self.layer3 = Down(out_channels, out_channels) 
        self.layer4 = Down(out_channels, out_channels) 
        self.layer1_1 = Up(in_channels, out_channels) 
        self.layer2_1 = Up(out_channels, out_channels) 
        self.layer3_1 = Up(out_channels, out_channels) 
        

    def forward(self, x1, x2, x3, x4, x5): #x是设置的参数，代表输入数据
        x2_1 =  self.layer1(x1) + x2 + self.layer1_1(x3) + self.layer2_1(self.layer1_1(x4)) + self.layer3_1(self.layer2_1(self.layer1_1(x5)))
        x3_1 =  self.layer1(x2_1) + x3 +self.layer2_1(x4) + self.layer3_1(self.layer2_1(x5))
        x4_1 =  self.layer1(x3_1) + x4 + self.layer3_1(x5)
        x5_1 =  self.layer1(x4_1) + x5
        return x1, x2, x3, x4

其中Down，Up，和Conv函数如下：

mport torch  
import torch.nn as nn   



class Conv(nn.Module):  #定义函数DoubleConv
    # 卷积+ReLU函数+卷积+ReLU函数
    def __init__(self, in_channels, out_channels):  ##in_channels, out_channels是设置的两个参数
        super().__init__()
        self.conv = nn.Sequential(  ##nn.Sequential:封装函数
            nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1),  ####卷积（输入channel，输出channel，卷积核大小3（一般取3/5，根据GPU计算力），卷积前扩张维度1）
            #nn.BatchNorm2d(out_channels),
            nn.ReLU(inplace=True),     #####激活函数，增强模型表达能力
        )

    def forward(self, x):  #x是设置的参数，代表输入数据
        x = self.conv(x)
        return x


class Down(nn.Module):   #定义下采样函数，给数据降维

    def __init__(self, in_channels, out_channels):
        super().__init__()
        
        self.maxpool_conv = nn.Sequential(
            nn.MaxPool2d(2), 
            Conv(in_channels, out_channels) 
        )

    def forward(self, x):
        return self.maxpool_conv(x)  ####池化操作


class Up(nn.Module): #定义上采样函数
    def __init__(self, in_channels, in_channel1, out_channels, bilinear=False):
        super().__init__()
        self.up = nn.ConvTranspose2d(in_channels, in_channels, kernel_size=3, stride=2, padding=1, output_padding=1)  

    def forward(self, x1): #x是设置的参数，代表输入数据F
        # 四维batch, channel, height, width
        x1 = self.up(x1)   ######反卷积
        return x