理解yolov7网络结构_[-1, 1, conv, [32, 3, 1]

 以下是yolov7网络结构配置的yaml，对每一层的输出加了注释。

# parameters
nc: 80  # number of classes
depth_multiple: 1.0  # model depth multiple
width_multiple: 1.0  # layer channel multiple
 
# anchors
anchors:
  - [12,16, 19,36, 40,28]  # P3/8
  - [36,75, 76,55, 72,146]  # P4/16
  - [142,110, 192,243, 459,401]  # P5/32
 
# yolov7 backbone
backbone:
  # [from, number, module, args]           [N,3,640,640]
  [[-1, 1, Conv, [32, 3, 1]],  # 0         [N,32,640,640]
  
   [-1, 1, Conv, [64, 3, 2]],  # 1-P1/2     [N,64,320,320]
   [-1, 1, Conv, [64, 3, 1]],              #[N,64,320,320]
   
   [-1, 1, Conv, [128, 3, 2]],  # 3-P2/4    [N,128,160,160]
   [-1, 1, Conv, [64, 1, 1]],  # -6         [N,64,160,160]
   [-2, 1, Conv, [64, 1, 1]],  # -5         [N,64,160,160]
   [-1, 1, Conv, [64, 3, 1]],              #[N,64,160,160]
   [-1, 1, Conv, [64, 3, 1]],  # -3        #[N,64,160,160]
   [-1, 1, Conv, [64, 3, 1]],              #[N,64,160,160]
   [-1, 1, Conv, [64, 3, 1]],  # -1        #[N,64,160,160]
   [[-1, -3, -5, -6], 1, Concat, [1]],     #[N,256,160,160]
   [-1, 1, Conv, [256, 1, 1]],  # 11       #[N,256,160,160]
         
   [-1, 1, MP, []],                        #[N,256,80,80]
   [-1, 1, Conv, [128, 1, 1]],             #[N,128,80,80]
   [-3, 1, Conv, [128, 1, 1]],             #[N,128,160,160]
   [-1, 1, Conv, [128, 3, 2]],             #[N,128,80,80]
   [[-1, -3], 1, Concat, [1]],  # 16-P3/8  #[N,256,80,80]
   [-1, 1, Conv, [128, 1, 1]],             #[N,128,80,80]
   [-2, 1, Conv, [128, 1, 1]],             #[N,128,80,80]
   [-1, 1, Conv, [128, 3, 1]],             #[N,128,80,80]
   [-1, 1, Conv, [128, 3, 1]],             #[N,128,80,80]
   [-1, 1, Conv, [128, 3, 1]],             #[N,128,80,80]
   [-1, 1, Conv, [128, 3, 1]],             #[N,128,80,80]
   [[-1, -3, -5, -6], 1, Concat, [1]],     #[N,512,80,80]
   [-1, 1, Conv, [512, 1, 1]],  # 24       #[N,512,80,80]
         
   [-1, 1, MP, []],                        #[N,512,40,40]
   [-1, 1, Conv, [256, 1, 1]],             #[N,256,40,40]
   [-3, 1, Conv, [256, 1, 1]],             #[N,256,80,80]
   [-1, 1, Conv, [256, 3, 2]],             #[N,256,40,40]
   [[-1, -3], 1, Concat, [1]],  # 29-P4/16 #[N,512,40,40]
   [-1, 1, Conv, [256, 1, 1]],             #[N,256,40,40]
   [-2, 1, Conv, [256, 1, 1]],             #[N,256,40,40]
   [-1, 1, Conv, [256, 3, 1]],             #[N,256,40,40]
   [-1, 1, Conv, [256, 3, 1]],             #[N,256,40,40]
   [-1, 1, Conv, [256, 3, 1]],             #[N,256,40,40]
   [-1, 1, Conv, [256, 3, 1]],             #[N,256,40,40]
   [[-1, -3, -5, -6], 1, Concat, [1]],     #[N,1024,40,40]
   [-1, 1, Conv, [1024, 1, 1]],  # 37      #[N,1024,40,40]
         
   [-1, 1, MP, []],                        #[N,1024,20,20]
   [-1, 1, Conv, [512, 1, 1]],             #[N,512,20,20]
   [-3, 1, Conv, [512, 1, 1]],             #[N,512,40,40]
   [-1, 1, Conv, [512, 3, 2]],             #[N,512,20,20]
   [[-1, -3], 1, Concat, [1]],  # 42-P5/32 #[N,1024,20,20]
   [-1, 1, Conv, [256, 1, 1]],             #[N,256,20,20]
   [-2, 1, Conv, [256, 1, 1]],             #[N,256,20,20]
   [-1, 1, Conv, [256, 3, 1]],             #[N,256,20,20]
   [-1, 1, Conv, [256, 3, 1]],             #[N,256,20,20]
   [-1, 1, Conv, [256, 3, 1]],             #[N,256,20,20]
   [-1, 1, Conv, [256, 3, 1]],             #[N,256,20,20]
   [[-1, -3, -5, -6], 1, Concat, [1]],     #[N,1024,20,20]
   [-1, 1, Conv, [1024, 1, 1]],  # 50      #[N,1024,20,20]
  ]
 
# yolov7 head
head:
  [[-1, 1, SPPCSPC, [512]], # 51                   #[N,512,20,20]
  
   [-1, 1, Conv, [256, 1, 1]],                     #[N,256,20,20]
   [-1, 1, nn.Upsample, [None, 2, 'nearest']],     #[N,256,40,40]
   [37, 1, Conv, [256, 1, 1]], # route backbone P4 #[N,256,40,40]
   [[-1, -2], 1, Concat, [1]],                     #[N,512,40,40]
   
   [-1, 1, Conv, [256, 1, 1]],                     #[N,256,40,40]
   [-2, 1, Conv, [256, 1, 1]],                     #[N,256,40,40]
   [-1, 1, Conv, [128, 3, 1]],                     #[N,128,40,40]
   [-1, 1, Conv, [128, 3, 1]],                     #[N,128,40,40]
   [-1, 1, Conv, [128, 3, 1]],                     #[N,128,40,40]
   [-1, 1, Conv, [128, 3, 1]],                     #[N,128,40,40]
   [[-1, -2, -3, -4, -5, -6], 1, Concat, [1]],     #[N,1024,40,40]
   [-1, 1, Conv, [256, 1, 1]], # 63                #[N,256,40,40]
   
   [-1, 1, Conv, [128, 1, 1]],                     #[N,128,40,40]
   [-1, 1, nn.Upsample, [None, 2, 'nearest']],     #[N,128,80,80]
   [24, 1, Conv, [128, 1, 1]], # route backbone P3 #[N,128,80,80]
   [[-1, -2], 1, Concat, [1]],                     #[N,256,80,80]
   
   [-1, 1, Conv, [128, 1, 1]],                     #[N,128,80,80]
   [-2, 1, Conv, [128, 1, 1]],                     #[N,128,80,80]
   [-1, 1, Conv, [64, 3, 1]],                      #[N,64,80,80]
   [-1, 1, Conv, [64, 3, 1]],                      #[N,64,80,80]
   [-1, 1, Conv, [64, 3, 1]],                      #[N,64,80,80]
   [-1, 1, Conv, [64, 3, 1]],                      #[N,64,80,80]
   [[-1, -2, -3, -4, -5, -6], 1, Concat, [1]],     #[N,512,80,80]
   [-1, 1, Conv, [128, 1, 1]], # 75                #[N,128,80,80]
      
   [-1, 1, MP, []],                                #[N,128,40,40]
   [-1, 1, Conv, [128, 1, 1]],                     #[N,128,40,40]
   [-3, 1, Conv, [128, 1, 1]],                     #[N,128,80,80]
   [-1, 1, Conv, [128, 3, 2]],                     #[N,128,40,40]
   [[-1, -3, 63], 1, Concat, [1]],                 #[N,512,40,40]
   
   [-1, 1, Conv, [256, 1, 1]],                     #[N,256,40,40]
   [-2, 1, Conv, [256, 1, 1]],                     #[N,256,40,40]
   [-1, 1, Conv, [128, 3, 1]],                     #[N,128,40,40]
   [-1, 1, Conv, [128, 3, 1]],                     #[N,128,40,40]
   [-1, 1, Conv, [128, 3, 1]],                     #[N,128,40,40]
   [-1, 1, Conv, [128, 3, 1]],                     #[N,128,40,40]
   [[-1, -2, -3, -4, -5, -6], 1, Concat, [1]],     #[N,1024,40,40]
   [-1, 1, Conv, [256, 1, 1]], # 88                #[N,256,40,40]
      
   [-1, 1, MP, []],                                #[N,256,20,20]
   [-1, 1, Conv, [256, 1, 1]],                     #[N,256,20,20]
   [-3, 1, Conv, [256, 1, 1]],                     #[N,256,40,40]
   [-1, 1, Conv, [256, 3, 2]],                     #[N,256,20,20]
   [[-1, -3, 51], 1, Concat, [1]],                 #[N,1024,20,20]
   
   [-1, 1, Conv, [512, 1, 1]],                     #[N,512,20,20]
   [-2, 1, Conv, [512, 1, 1]],                     #[N,512,20,20]
   [-1, 1, Conv, [256, 3, 1]],                     #[N,256,20,20]
   [-1, 1, Conv, [256, 3, 1]],                     #[N,256,20,20]
   [-1, 1, Conv, [256, 3, 1]],                     #[N,256,20,20]
   [-1, 1, Conv, [256, 3, 1]],                     #[N,256,20,20]
   [[-1, -2, -3, -4, -5, -6], 1, Concat, [1]],     #[N,2048,20,20]
   [-1, 1, Conv, [512, 1, 1]], # 101               #[N,512,20,20]
   
   [75, 1, RepConv, [256, 3, 1]],                  #[N,256,80,80]
   [88, 1, RepConv, [512, 3, 1]],                  #[N,512,40,40]
   [101, 1, RepConv, [1024, 3, 1]],                #[N,1024,20,20]
 
   [[102,103,104], 1, IDetect, [nc, anchors]],   # Detect(P3, P4, P5)
  ]

网络结构图

网络结构分块讲解

1、基础模块

 主要是1x1以及3x3的卷积，3x3的卷积区别主要是stride不同，最后都是用SiLU进行激活。

 2、多路卷积模块

[-1, 1, Conv, [256, 1, 1]],                     #[4,256,40,40]
[-2, 1, Conv, [256, 1, 1]],                     #[4,256,40,40]
[-1, 1, Conv, [128, 3, 1]],                     #[4,128,40,40]
[-1, 1, Conv, [128, 3, 1]],                     #[4,128,40,40]
[-1, 1, Conv, [128, 3, 1]],                     #[4,128,40,40]
[-1, 1, Conv, [128, 3, 1]],                     #[4,128,40,40]
[[-1, -2, -3, -4, -5, -6], 1, Concat, [1]],     #[4,1024,40,40]
[-1, 1, Conv, [256, 1, 1]], # 63                #[4,256,40,40]

or

[-1, 1, Conv, [64, 1, 1]],  # -6         [4,64,160,160]
[-2, 1, Conv, [64, 1, 1]],  # -5         [4,64,160,160]
[-1, 1, Conv, [64, 3, 1]],              #[4,64,160,160]
[-1, 1, Conv, [64, 3, 1]],  # -3        #[4,64,160,160]
[-1, 1, Conv, [64, 3, 1]],              #[4,64,160,160]
[-1, 1, Conv, [64, 3, 1]],  # -1        #[4,64,160,160]
[[-1, -3, -5, -6], 1, Concat, [1]],     #[4,256,160,160]
[-1, 1, Conv, [256, 1, 1]],  # 11       #[4,256,160,160]

 对应的模块

or

         其中主要是用了较多的1x1和3x3的卷积，每个卷积的输出会作为下一个卷积的输入，同时还会和其他卷积的输出进行concat，这样的结构提升了网络的精度但是这也会增加一定的耗时。

3、MP模块

[-1, 1, MP, []],                        #[4,256,80,80]
[-1, 1, Conv, [128, 1, 1]],             #[4,128,80,80]
[-3, 1, Conv, [128, 1, 1]],             #[4,128,160,160]
[-1, 1, Conv, [128, 3, 2]],             #[4,128,80,80]

MP代码如下：

class MP(nn.Module):
    def __init__(self, k=2):
        super(MP, self).__init__()
        self.m = nn.MaxPool2d(kernel_size=k, stride=k)
 
    def forward(self, x):
        return self.m(x)

图解结构如下： 

          

        将kernel_size和stride设置为2的Maxpool2d，该模块会用到1x1卷积，3x3卷积，以及一个Maxpool2d，之后将卷积输出进行concat，从而组成了该模块。

4、SPPCSPC模块

        该模块是利用金字塔池化和CSP结构得到的模块，其中也是用了较多的支路，具体代码如下：

class SPPCSPC(nn.Module):
    # CSP https://github.com/WongKinYiu/CrossStagePartialNetworks
    def __init__(self, c1, c2, n=1, shortcut=False, g=1, e=0.5, k=(5, 9, 13)):
        super(SPPCSPC, self).__init__()
        c_ = int(2 * c2 * e)  # hidden channels
        self.cv1 = Conv(c1, c_, 1, 1)
        self.cv2 = Conv(c1, c_, 1, 1)
        self.cv3 = Conv(c_, c_, 3, 1)
        self.cv4 = Conv(c_, c_, 1, 1)
        self.m = nn.ModuleList([nn.MaxPool2d(kernel_size=x, stride=1, padding=x // 2) for x in k])
        self.cv5 = Conv(4 * c_, c_, 1, 1)
        self.cv6 = Conv(c_, c_, 3, 1)
        self.cv7 = Conv(2 * c_, c2, 1, 1)
 
    def forward(self, x):
        x1 = self.cv4(self.cv3(self.cv1(x)))
        y1 = self.cv6(self.cv5(torch.cat([x1] + [m(x1) for m in self.m], 1)))
        y2 = self.cv2(x)
        return self.cv7(torch.cat((y1, y2), dim=1))

        图解之后得到下图：

         其中的MP就是金字塔池化，最后将两条分支的Y1和Y2的输出进行concat。

5、RepConv模块

 代码如下：

class RepConv(nn.Module):
    # Represented convolution
    # https://arxiv.org/abs/2101.03697
 
    def __init__(self, c1, c2, k=3, s=1, p=None, g=1, act=True, deploy=False):
        super(RepConv, self).__init__()
 
        self.deploy = deploy
        self.groups = g
        self.in_channels = c1
        self.out_channels = c2
 
        assert k == 3
        assert autopad(k, p) == 1
 
        padding_11 = autopad(k, p) - k // 2
 
        self.act = nn.SiLU() if act is True else (act if isinstance(act, nn.Module) else nn.Identity())
 
        if deploy:
            self.rbr_reparam = nn.Conv2d(c1, c2, k, s, autopad(k, p), groups=g, bias=True)
 
        else:
            self.rbr_identity = (nn.BatchNorm2d(num_features=c1) if c2 == c1 and s == 1 else None)
 
            self.rbr_dense = nn.Sequential(
                nn.Conv2d(c1, c2, k, s, autopad(k, p), groups=g, bias=False),
                nn.BatchNorm2d(num_features=c2),
            )
 
            self.rbr_1x1 = nn.Sequential(
                nn.Conv2d( c1, c2, 1, s, padding_11, groups=g, bias=False),
                nn.BatchNorm2d(num_features=c2),
            )
 
    def forward(self, inputs):
        if hasattr(self, "rbr_reparam"):
            return self.act(self.rbr_reparam(inputs))
 
        if self.rbr_identity is None:
            id_out = 0
        else:
            id_out = self.rbr_identity(inputs)
 
        return self.act(self.rbr_dense(inputs) + self.rbr_1x1(inputs) + id_out)

图解之后得到下图：

该模块的具体讲解可以参看：BACKBONE部分。

最后，如有理解不当的地方，欢迎大佬来纠错。