YOLOv7的yolov7.yaml详解

ILoveNanjing

已于 2023-10-25 15:27:51 修改

阅读量1.1k

点赞数

分类专栏： YOLO 文章标签： YOLO 1024程序员节

于 2023-10-20 20:29:12 首次发布

本文链接：https://blog.csdn.net/m0_70654337/article/details/133953551

版权

YOLO 专栏收录该内容

7 篇文章

订阅专栏

将YOLOv7流程图与yolov7.yaml文件进行一一对应，相互匹配，便于理解整个网络过程。

具体注释如下

# 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]    输入为 640*640*3
  [[-1, 1, Conv, [32, 3, 1]],  # 0    第零层
  
   [-1, 1, Conv, [64, 3, 2]],  # 1-P1/2   320*320*64
   [-1, 1, Conv, [64, 3, 1]],
   
   [-1, 1, Conv, [128, 3, 2]],  # 3-P2/4    160*160*128

#  该部分整体为ELAN，H和W不发生变化，倒数第二行channel数由C变为2C，最终输出channel由倒数第一行的Conv决定，该部分是2C
   [-1, 1, Conv, [64, 1, 1]],
   [-2, 1, Conv, [64, 1, 1]],
   [-1, 1, Conv, [64, 3, 1]],
   [-1, 1, Conv, [64, 3, 1]],
   [-1, 1, Conv, [64, 3, 1]],
   [-1, 1, Conv, [64, 3, 1]],
   [[-1, -3, -5, -6], 1, Concat, [1]],    # 160*160*256   Concat是在channel维度上进行拼接
   [-1, 1, Conv, [256, 1, 1]],  # 11    160*160*256

#  该部分为MP1，H和W变为原来的0.5倍,C不变
   [-1, 1, MP, []],   # 80*80*256   MP为最大池化MaxPooling，H和W变为原来的0.5倍，channel不变
   [-1, 1, Conv, [128, 1, 1]],
   [-3, 1, Conv, [128, 1, 1]],
   [-1, 1, Conv, [128, 3, 2]],
   [[-1, -3], 1, Concat, [1]],  # 16-P3/8   80*80*256

#   ELAN，H和W不发生变化，倒数第二行channel数由C变为2C，最终输出channel由倒数第一行的Conv决定，该部分是2C
   [-1, 1, Conv, [128, 1, 1]],
   [-2, 1, Conv, [128, 1, 1]],
   [-1, 1, Conv, [128, 3, 1]],
   [-1, 1, Conv, [128, 3, 1]],
   [-1, 1, Conv, [128, 3, 1]],
   [-1, 1, Conv, [128, 3, 1]],
   [[-1, -3, -5, -6], 1, Concat, [1]],    # 80*80*512
   [-1, 1, Conv, [512, 1, 1]],  # 24    80*80*512

#   MP1 + ELAN
   [-1, 1, MP, []],
   [-1, 1, Conv, [256, 1, 1]],
   [-3, 1, Conv, [256, 1, 1]],
   [-1, 1, Conv, [256, 3, 2]],
   [[-1, -3], 1, Concat, [1]],  # 29-P4/16    40*40*512
   [-1, 1, Conv, [256, 1, 1]],
   [-2, 1, Conv, [256, 1, 1]],
   [-1, 1, Conv, [256, 3, 1]],
   [-1, 1, Conv, [256, 3, 1]],
   [-1, 1, Conv, [256, 3, 1]],
   [-1, 1, Conv, [256, 3, 1]],
   [[-1, -3, -5, -6], 1, Concat, [1]],    # 40*40*1024
   [-1, 1, Conv, [1024, 1, 1]],  # 37   40*40*1024

#   MP1 + ELAN  注意该部分ELAN与前面的不同，输出的channel与之前相同
   [-1, 1, MP, []],
   [-1, 1, Conv, [512, 1, 1]],
   [-3, 1, Conv, [512, 1, 1]],
   [-1, 1, Conv, [512, 3, 2]],
   [[-1, -3], 1, Concat, [1]],  # 42-P5/32    20*20*1024
   [-1, 1, Conv, [256, 1, 1]],
   [-2, 1, Conv, [256, 1, 1]],
   [-1, 1, Conv, [256, 3, 1]],
   [-1, 1, Conv, [256, 3, 1]],
   [-1, 1, Conv, [256, 3, 1]],
   [-1, 1, Conv, [256, 3, 1]],
   [[-1, -3, -5, -6], 1, Concat, [1]],    # 20*20*1024
   [-1, 1, Conv, [1024, 1, 1]],  # 50   20*20*1024
  ]

# yolov7 head
head:
  [[-1, 1, SPPCSPC, [512]], # 51    20*20*512   最终输出只有channel发生变化

#  UP+CONCAT 前两行构成UP模块，第三行对第二组MP1+ELAN后的输出（# 37）进行卷积，最后统一concat
   [-1, 1, Conv, [256, 1, 1]],
   [-1, 1, nn.Upsample, [None, 2, 'nearest']],    # 40*40*256   上采样，H和W变为原来的2倍，channel不变
   [37, 1, Conv, [256, 1, 1]], # route backbone P4
   [[-1, -2], 1, Concat, [1]],    # 40*40*512

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

#  UP+CONCAT
   [-1, 1, Conv, [128, 1, 1]],
   [-1, 1, nn.Upsample, [None, 2, 'nearest']],    # 80*80*128
   [24, 1, Conv, [128, 1, 1]], # route backbone P3
   [[-1, -2], 1, Concat, [1]],    # 80*80*256

#   ELAN-H
   [-1, 1, Conv, [128, 1, 1]],
   [-2, 1, Conv, [128, 1, 1]],
   [-1, 1, Conv, [64, 3, 1]],
   [-1, 1, Conv, [64, 3, 1]],
   [-1, 1, Conv, [64, 3, 1]],
   [-1, 1, Conv, [64, 3, 1]],
   [[-1, -2, -3, -4, -5, -6], 1, Concat, [1]],    # 80*80*512
   [-1, 1, Conv, [128, 1, 1]], # 75   80*80*128

#   MP2
   [-1, 1, MP, []],   # 40*40*128
   [-1, 1, Conv, [128, 1, 1]],
   [-3, 1, Conv, [128, 1, 1]],
   [-1, 1, Conv, [128, 3, 2]],    # 40*40*128
   [[-1, -3, 63], 1, Concat, [1]],    # 40*40*512

#   ELAN-H
   [-1, 1, Conv, [256, 1, 1]],
   [-2, 1, Conv, [256, 1, 1]],
   [-1, 1, Conv, [128, 3, 1]],
   [-1, 1, Conv, [128, 3, 1]],
   [-1, 1, Conv, [128, 3, 1]],
   [-1, 1, Conv, [128, 3, 1]],
   [[-1, -2, -3, -4, -5, -6], 1, Concat, [1]],    # 40*40*1024
   [-1, 1, Conv, [256, 1, 1]], # 88   40*40*256

#  MP2
   [-1, 1, MP, []],   # 20*20*256
   [-1, 1, Conv, [256, 1, 1]],
   [-3, 1, Conv, [256, 1, 1]],
   [-1, 1, Conv, [256, 3, 2]],    # 20*20*256
   [[-1, -3, 51], 1, Concat, [1]],    # 20*20*1024

#   ELAN-H
   [-1, 1, Conv, [512, 1, 1]],
   [-2, 1, Conv, [512, 1, 1]],
   [-1, 1, Conv, [256, 3, 1]],
   [-1, 1, Conv, [256, 3, 1]],
   [-1, 1, Conv, [256, 3, 1]],
   [-1, 1, Conv, [256, 3, 1]],
   [[-1, -2, -3, -4, -5, -6], 1, Concat, [1]],    # 20*20*2048
   [-1, 1, Conv, [512, 1, 1]], # 101    20*20*512

#   RepConv是重参数化卷积，用3*3卷积重参数化，加速推理，不改变H和W，具体内容参考论文
   [75, 1, RepConv, [256, 3, 1]],
   [88, 1, RepConv, [512, 3, 1]],
   [101, 1, RepConv, [1024, 3, 1]],

   [[102,103,104], 1, IDetect, [nc, anchors]],   # Detect(P3, P4, P5)
  ]

参考链接：深入浅出 Yolo 系列之 Yolov7 基础网络结构详解_yolov7网络结构_计算机视觉linke的博客-CSDN博客

上图为参照的V7结构图，其中，右下方的MP2的输出是80*80*256，个人计算结果是40*40*256。

MP1、MP2、UP模块看结构图不易理解，容易误解，建议自己梳理yaml文件，思路很清晰。

图里的ELAN模块比论文的好理解很多，

SPPCSPC在yaml文件里没有解释，具体代码在common.py里，附上代码如下，可结合流程图梳理，很清晰

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))

最右面的是检测头部分，自下到上为P3 P4 P5

RepConv为重参数化卷积，简单来说就是在训练时的多分支结构等效为推理时的单路径结构，精度提升一点点，速度提升很多，个人看完RepVGG论文后感觉很有意义，数学推导很有意思也很有道理。附上论文中的示意图以及参考博主的流程图