Darknet网络结构

关于Darknet训练模型，可以见我的其他文章Darknet训练数据集Linux，环境配置win/lin

Darknet网络结构

Darknet的最大创新之处在于将卷积和残差链接，网络结构如图所示，最后三层 Avgpool、Connected 和 Softmax 层没有标注，用来在 ImageNet 数据集上训练分类任务时使用
该表格与Darknet53模块相对应，一起看效果更佳
图片来源CSDN Tc.小浩

Darknet53的卷积单元CBL，在Yolov5网络结构及其他有提到，

CBL为标准卷积模块，包括普通卷积层Conv、批量归一化层BN和LeakyReLU激活函数层

CBL模块

import torch
import torch.nn as nn

def autopad(k, p=None):  # kernel,padding
    # pad to "same":
    # H_out=H_in;W_out=W_in
    # H_out=(H_in+2p-k)/s+1
    # W_out=(W_in+2p-k)/s+1
    if p is None:
        if isinstance(k,int):
            p = k//2 if k%2==1 else k//2 -1
        else:
            p = [x // 2 if x % 2 == 1 else x // 2 - 1 for x in k]
    return p


class Conv(nn.Module):
    def __init__(self,c1,c2,k=1,s=1,p=None,g=1,act=True):
        super(Conv,self).__init__()  # 初始化
        self.conv=nn.Conv2d(c1,c2,k,s,autopad(k, p),groups=g,bias=False)  # 分组与偏置项即b
        self.bn=nn.BatchNorm2d(c2)
        self.act=nn.LeakyReLU(0.01) if act else nn.Identity

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

残差模块

class ResidualLayer(nn.Module):
    def __init__(self, c1):		# converlution * 2 + residual
        c_ = c1 // 2
        super(ResidualLayer, self).__init__()
        self.conv = nn.Sequential(
            Conv(c1, c_, k=1, s=1,p=0),  # kernel_size = 1降通道
            Conv(c_, c1, k=3, s=1,p=1),  # 再用kernel_size = 3通道升
            )

    def forward(self, x):
        return x + self.conv(x)	 # 前向传播，增加残差，原始信息与特征连接

Darknet53模块

class Darknet53(nn.Module):
    def __init__(self):
        super(Darknet53, self).__init__()
        self.conv1 = Conv(3, 32, 3, 1, 1)  # 一个卷积块 = 1层卷积
        self.conv2 = Conv(32, 64, 3, 2, 1)
        self.conv3_4 = ResidualLayer(64)  # 一个残差块 = 2层卷积
        self.conv5 = Conv(64, 128, 3, 2, 1)
        self.conv6_9 = nn.Sequential(  # = 4层卷积
            ResidualLayer(128),
            ResidualLayer(128),
        )
        self.conv10 = Conv(128, 256, 3, 2, 1)
        self.conv11_26 = nn.Sequential(  # = 16层卷积
            ResidualLayer(256),
            ResidualLayer(256),
            ResidualLayer(256),
            ResidualLayer(256),
            ResidualLayer(256),
            ResidualLayer(256),
            ResidualLayer(256),
            ResidualLayer(256),
        )
        self.conv27 = Conv(256, 512, 3, 2, 1)
        self.conv28_43 = nn.Sequential(  # = 16层卷积
            ResidualLayer(512),
            ResidualLayer(512),
            ResidualLayer(512),
            ResidualLayer(512),
            ResidualLayer(512),
            ResidualLayer(512),
            ResidualLayer(512),
            ResidualLayer(512),
        )
        self.conv44 = Conv(512, 1024, 3, 2, 1)
        self.conv45_52 = nn.Sequential(  # = 8层卷积  Darknet53有52层卷积以及1层输出
            ResidualLayer(1024),
            ResidualLayer(1024),
            ResidualLayer(1024),
            ResidualLayer(1024),
        )

    def forward(self, entry):
        conv1 = self.conv1(entry)
        conv2 = self.conv2(conv1)
        conv3_4 = self.conv3_4(conv2)
        conv5 = self.conv5(conv3_4)
        conv6_9 = self.conv6_9(conv5)
        conv10 = self.conv10(conv6_9)
        conv11_26 = self.conv11_26(conv10)
        conv27 = self.conv27(conv11_26)
        conv28_43 = self.conv28_43(conv27)
        conv44 = self.conv44(conv28_43)
        conv45_52 = self.conv45_52(conv44)
        return conv45_52, conv28_43, conv11_26  # YOLOv3用，所以输出了3次特征

测试

Darknetmodel = Darknet53()
if __name__ == '__main__':
    # 输入一张416*416的图片，batch输入维度为（1，3，416，416）
    batch = torch.randn(1, 3, 416, 416)
    conv45_52, conv28_43, conv11_26 = Darknetmodel(batch)
    print(conv45_52.shape)
    print(conv28_43.shape)
    print(conv11_26.shape)
    """
    torch.Size([1, 1024, 13, 13])
    torch.Size([1, 512, 26, 26])
    torch.Size([1, 256, 52, 52])
    """