devbox
菜鸟追梦旅行
这个屌丝很懒，什么也没留下！
关注作者
热门标签
热门文章
当前位置:   article > 正文

yolo增加RFEM_c3rfem

作者：菜鸟追梦旅行 | 2024-02-16 13:52:33

赞

踩

c3rfem

论文地址：https://arxiv.org/pdf/2208.02019.pdf

代码地址：GitHub - Krasjet-Yu/YOLO-FaceV2: YOLO-FaceV2: A Scale and Occlusion Aware Face Detector

总的来说就是RFEM利用了感受野在特征图中的优势，通过使用不同膨胀卷积率的分支来捕捉多尺度信息和不同范围的依赖关系。这种设计有助于减少参数数量，降低过拟合风险，并充分利用每个样本。

1、yolov7-tiny

 创建配置文件yolov7-tiny-RFEM.yaml

  1. # parameters
  2. nc: 80  # number of classes
  3. depth_multiple: 1.0  # model depth multiple
  4. width_multiple: 1.0  # layer channel multiple
  5.  
  6. activation: nn.LeakyReLU(0.1)
  7. # anchors
  8. anchors:
  9.   - [10,13, 16,30, 33,23]  # P3/8
  10.   - [30,61, 62,45, 59,119]  # P4/16
  11.   - [116,90, 156,198, 373,326]  # P5/32
  12.  
  13. # yolov7-tiny backbone
  14. backbone:
  15.   # [from, number, module, args] c2, k=1, s=1, p=None, g=1, act=True
  16.   [[-1, 1, Conv, [32, 3, 2, None, 1]],  # 0-P1/2
  17.   
  18.    [-1, 1, Conv, [64, 3, 2, None, 1]],  # 1-P2/4
  19.    
  20.    [-1, 1, Conv, [32, 1, 1, None, 1]],
  21.    [-2, 1, Conv, [32, 1, 1, None, 1]],
  22.    [-1, 1, Conv, [32, 3, 1, None, 1]],
  23.    [-1, 1, Conv, [32, 3, 1, None, 1]],
  24.    [[-1, -2, -3, -4], 1, Concat, [1]],
  25.    [-1, 1, Conv, [64, 1, 1, None, 1]],  # 7
  26.    
  27.    [-1, 1, MP, []],  # 8-P3/8
  28.    [-1, 1, Conv, [64, 1, 1, None, 1]],
  29.    [-2, 1, Conv, [64, 1, 1, None, 1]],
  30.    [-1, 1, Conv, [64, 3, 1, None, 1]],
  31.    [-1, 1, Conv, [64, 3, 1, None, 1]],
  32.    [[-1, -2, -3, -4], 1, Concat, [1]],
  33.    [-1, 1, Conv, [128, 1, 1, None, 1]],  # 14
  34.    
  35.    [-1, 1, MP, []],  # 15-P4/16
  36.    [-1, 1, Conv, [128, 1, 1, None, 1]],
  37.    [-2, 1, Conv, [128, 1, 1, None, 1]],
  38.    [-1, 1, Conv, [128, 3, 1, None, 1]],
  39.    [-1, 1, Conv, [128, 3, 1, None, 1]],
  40.    [[-1, -2, -3, -4], 1, Concat, [1]],
  41.    [-1, 1, Conv, [256, 1, 1, None, 1]],  # 21
  42.    
  43.    [-1, 1, MP, []],  # 22-P5/32
  44.    [-1, 1, Conv, [256, 1, 1, None, 1]],
  45.    [-2, 1, Conv, [256, 1, 1, None, 1]],
  46.    [-1, 1, Conv, [256, 3, 1, None, 1]],
  47.    [-1, 1, Conv, [256, 3, 1, None, 1]],
  48.    [[-1, -2, -3, -4], 1, Concat, [1]],
  49.    [-1, 1, Conv, [512, 1, 1, None, 1]],  # 28
  50.   ]
  51.  
  52. # yolov7-tiny head
  53. head:
  54.   [[-1, 1, Conv, [256, 1, 1, None, 1]],
  55.    [-2, 1, Conv, [256, 1, 1, None, 1]],
  56.    [-1, 1, SP, [5]],
  57.    [-2, 1, SP, [9]],
  58.    [-3, 1, SP, [13]],
  59.    [[-1, -2, -3, -4], 1, Concat, [1]],
  60.    [-1, 1, Conv, [256, 1, 1, None, 1]],
  61.    [[-1, -7], 1, Concat, [1]],
  62.    [-1, 1, Conv, [256, 1, 1, None, 1]],  # 37
  63.    [-1, 1, RFEM, [256]],
  64.   
  65.    [-1, 1, Conv, [128, 1, 1, None, 1]],
  66.    [-1, 1, nn.Upsample, [None, 2, 'nearest']],
  67.    [21, 1, Conv, [128, 1, 1, None, 1]], # route backbone P4
  68.    [[-1, -2], 1, Concat, [1]],
  69.    
  70.    [-1, 1, Conv, [64, 1, 1, None, 1]],
  71.    [-2, 1, Conv, [64, 1, 1, None, 1]],
  72.    [-1, 1, Conv, [64, 3, 1, None, 1]],
  73.    [-1, 1, Conv, [64, 3, 1, None, 1]],
  74.    [[-1, -2, -3, -4], 1, Concat, [1]],
  75.    [-1, 1, Conv, [128, 1, 1, None, 1]],  # 48
  76.   
  77.    [-1, 1, Conv, [64, 1, 1, None, 1]],
  78.    [-1, 1, nn.Upsample, [None, 2, 'nearest']],
  79.    [14, 1, Conv, [64, 1, 1, None, 1]], # route backbone P3
  80.    [[-1, -2], 1, Concat, [1]],
  81.    
  82.    [-1, 1, Conv, [32, 1, 1, None, 1]],
  83.    [-2, 1, Conv, [32, 1, 1, None, 1]],
  84.    [-1, 1, Conv, [32, 3, 1, None, 1]],
  85.    [-1, 1, Conv, [32, 3, 1, None, 1]],
  86.    [[-1, -2, -3, -4], 1, Concat, [1]],
  87.    [-1, 1, Conv, [64, 1, 1, None, 1]],  # 58
  88.    
  89.    [-1, 1, Conv, [128, 3, 2, None, 1]],
  90.    [[-1, 48], 1, Concat, [1]],
  91.    
  92.    [-1, 1, Conv, [64, 1, 1, None, 1]],
  93.    [-2, 1, Conv, [64, 1, 1, None, 1]],
  94.    [-1, 1, Conv, [64, 3, 1, None, 1]],
  95.    [-1, 1, Conv, [64, 3, 1, None, 1]],
  96.    [[-1, -2, -3, -4], 1, Concat, [1]],
  97.    [-1, 1, Conv, [128, 1, 1, None, 1]],  # 66
  98.    
  99.    [-1, 1, Conv, [256, 3, 2, None, 1]],
  100.    [[-1, 37], 1, Concat, [1]],
  101.    
  102.    [-1, 1, Conv, [128, 1, 1, None, 1]],
  103.    [-2, 1, Conv, [128, 1, 1, None, 1]],
  104.    [-1, 1, Conv, [128, 3, 1, None, 1]],
  105.    [-1, 1, Conv, [128, 3, 1, None, 1]],
  106.    [[-1, -2, -3, -4], 1, Concat, [1]],
  107.    [-1, 1, Conv, [256, 1, 1, None, 1]],  # 74
  108.       
  109.    [58, 1, Conv, [128, 3, 1, None, 1]],
  110.    [66, 1, Conv, [256, 3, 1, None, 1]],
  111.    [74, 1, Conv, [512, 3, 1, None, 1]],
  112.  
  113.    [[75,76,77], 1, Detect, [nc, anchors]],   # Detect(P3, P4, P5)
  114.   ]

在common.py中增加

  1. # RFEM
  2. class TridentBlock(nn.Module):
  3.     def __init__(self, c1, c2, stride=1, c=False, e=0.5, padding=[1, 2, 3], dilate=[1, 2, 3], bias=False):
  4.         super(TridentBlock, self).__init__()
  5.         self.stride = stride
  6.         self.c = c
  7.         c_ = int(c2 * e)
  8.         self.padding = padding
  9.         self.dilate = dilate
  10.         self.share_weightconv1 = nn.Parameter(torch.Tensor(c_, c1, 1, 1))
  11.         self.share_weightconv2 = nn.Parameter(torch.Tensor(c2, c_, 3, 3))
  12.  
  13.         self.bn1 = nn.BatchNorm2d(c_)
  14.         self.bn2 = nn.BatchNorm2d(c2)
  15.  
  16.         # self.act = nn.SiLU()
  17.         self.act = Conv.default_act
  18.  
  19.         nn.init.kaiming_uniform_(self.share_weightconv1, nonlinearity="relu")
  20.         nn.init.kaiming_uniform_(self.share_weightconv2, nonlinearity="relu")
  21.  
  22.         if bias:
  23.             self.bias = nn.Parameter(torch.Tensor(c2))
  24.         else:
  25.             self.bias = None
  26.  
  27.         if self.bias is not None:
  28.             nn.init.constant_(self.bias, 0)
  29.  
  30.     def forward_for_small(self, x):
  31.         residual = x
  32.         out = nn.functional.conv2d(x, self.share_weightconv1, bias=self.bias)
  33.         out = self.bn1(out)
  34.         out = self.act(out)
  35.  
  36.         out = nn.functional.conv2d(out, self.share_weightconv2, bias=self.bias, stride=self.stride,
  37.                                    padding=self.padding[0],
  38.                                    dilation=self.dilate[0])
  39.         out = self.bn2(out)
  40.         out += residual
  41.         out = self.act(out)
  42.  
  43.         return out
  44.  
  45.     def forward_for_middle(self, x):
  46.         residual = x
  47.         out = nn.functional.conv2d(x, self.share_weightconv1, bias=self.bias)
  48.         out = self.bn1(out)
  49.         out = self.act(out)
  50.  
  51.         out = nn.functional.conv2d(out, self.share_weightconv2, bias=self.bias, stride=self.stride,
  52.                                    padding=self.padding[1],
  53.                                    dilation=self.dilate[1])
  54.         out = self.bn2(out)
  55.         out += residual
  56.         out = self.act(out)
  57.  
  58.         return out
  59.  
  60.     def forward_for_big(self, x):
  61.         residual = x
  62.         out = nn.functional.conv2d(x, self.share_weightconv1, bias=self.bias)
  63.         out = self.bn1(out)
  64.         out = self.act(out)
  65.  
  66.         out = nn.functional.conv2d(out, self.share_weightconv2, bias=self.bias, stride=self.stride,
  67.                                    padding=self.padding[2],
  68.                                    dilation=self.dilate[2])
  69.         out = self.bn2(out)
  70.         out += residual
  71.         out = self.act(out)
  72.  
  73.         return out
  74.  
  75.     def forward(self, x):
  76.         xm = x
  77.         base_feat = []
  78.         if self.c is not False:
  79.             x1 = self.forward_for_small(x)
  80.             x2 = self.forward_for_middle(x)
  81.             x3 = self.forward_for_big(x)
  82.         else:
  83.             x1 = self.forward_for_small(xm[0])
  84.             x2 = self.forward_for_middle(xm[1])
  85.             x3 = self.forward_for_big(xm[2])
  86.  
  87.         base_feat.append(x1)
  88.         base_feat.append(x2)
  89.         base_feat.append(x3)
  90.  
  91.         return base_feat
  92.  
  93.  
  94. class RFEM(nn.Module):
  95.     def __init__(self, c1, c2, n=1, e=0.5, stride=1):
  96.         super(RFEM, self).__init__()
  97.         c = True
  98.         layers = []
  99.         layers.append(TridentBlock(c1, c2, stride=stride, c=c, e=e))
  100.         c1 = c2
  101.         for i in range(1, n):
  102.             layers.append(TridentBlock(c1, c2))
  103.         self.layer = nn.Sequential(*layers)
  104.         # self.cv = Conv(c2, c2)
  105.         self.bn = nn.BatchNorm2d(c2)
  106.         # self.act = nn.SiLU()
  107.         self.act = Conv.default_act
  108.  
  109.     def forward(self, x):
  110.         out = self.layer(x)
  111.         out = out[0] + out[1] + out[2] + x
  112.         out = self.act(self.bn(out))
  113.         return out
  114.  
  115.  
  116. class C3RFEM(nn.Module):
  117.     def __init__(self, c1, c2, n=1, shortcut=True, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
  118.         super().__init__()
  119.         c_ = int(c2 * e)  # hidden channels
  120.         self.cv1 = Conv(c1, c_, 1, 1)
  121.         self.cv2 = Conv(c1, c_, 1, 1)
  122.         self.cv3 = Conv(2 * c_, c2, 1)  # act=FReLU(c2)
  123.         # self.m = nn.Sequential(*(Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)))
  124.         # self.rfem = RFEM(c_, c_, n)
  125.         self.m = nn.Sequential(*[RFEM(c_, c_, n=1, e=e) for _ in range(n)])
  126.         # self.m = nn.Sequential(*[CrossConv(c_, c_, 3, 1, g, 1.0, shortcut) for _ in range(n)])
  127.  
  128.     def forward(self, x):
  129.         return self.cv3(torch.cat((self.m(self.cv1(x)), self.cv2(x)), dim=1))

其中C3RFEM对应原作者的实现

在yolo.py中修改：

  1.         n = n_ = max(round(n * gd), 1) if n > 1 else n  # depth gain
  2.         if m in {
  3.             Conv, GhostConv, Bottleneck, GhostBottleneck, SPP, SPPF, DWConv, MixConv2d, Focus, CrossConv,
  4.             BottleneckCSP, C3, C3TR, C3SPP, C3Ghost, nn.ConvTranspose2d, DWConvTranspose2d, C3x, StemBlock,
  5.             BlazeBlock, DoubleBlazeBlock, ShuffleV2Block, MobileBottleneck, InvertedResidual, ConvBNReLU,
  6.             RepVGGBlock, SEBlock, RepBlock, SimCSPSPPF, C3_P, SPPCSPC, RepConv, RFEM, C3RFEM}:
  7.             c1, c2 = ch[f], args[0]
  8.             if c2 != no:  # if not output
  9.                 c2 = make_divisible(c2 * gw, 8)
  10.             if m == InvertedResidual:
  11.                 c2 = make_divisible(c2 * gw, 4 if gw == 0.1 else 8)
  12.                 
  13.             args = [c1, c2, *args[1:]]
  14.             if m in {BottleneckCSP, C3, C3TR, C3Ghost, C3x, C3_P, C3RFEM}:
  15.                 args.insert(2, n)  # number of repeats
  16.                 n = 1

运行yolo.py

2、yolov5

yolov5s-RFEM.yaml

  1. # YOLOv5 
    声明：本文内容由网友自发贡献，不代表【wpsshop博客】立场，版权归原作者所有，本站不承担相应法律责任。如您发现有侵权的内容，请联系我们。转载请注明出处：https://www.wpsshop.cn/w/菜鸟追梦旅行/article/detail/93958
    推荐阅读
    相关标签

Copyright © 2003-2013 www.wpsshop.cn 版权所有，并保留所有权利。