devbox
IT小白
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

涨点技巧：注意力机制---Yolov5/Yolov7引入CBAM、GAM、Resnet_CBAM_yolo注意力机制

作者：IT小白 | 2024-07-24 08:53:16

赞

踩

yolo注意力机制

1.计算机视觉中的注意力机制

一般来说，注意力机制通常被分为以下基本四大类：

通道注意力 Channel Attention

空间注意力机制 Spatial Attention

时间注意力机制 Temporal Attention

分支注意力机制 Branch Attention

1.1.CBAM：通道注意力和空间注意力的集成者

轻量级的卷积注意力模块，它结合了通道和空间的注意力机制模块

论文题目：《CBAM: Convolutional Block Attention Module》
论文地址：  https://arxiv.org/pdf/1807.06521.pdf

上图可以看到，CBAM包含CAM（Channel Attention Module）和SAM（Spartial Attention Module）两个子模块，分别进行通道和空间上的Attention。这样不只能够节约参数和计算力，并且保证了其能够做为即插即用的模块集成到现有的网络架构中去。

1.2 GAM：Global Attention Mechanism

超越CBAM，全新注意力GAM：不计成本提高精度！
论文题目：Global Attention Mechanism: Retain Information to Enhance Channel-Spatial Interactions
论文地址：https://paperswithcode.com/paper/global-attention-mechanism-retain-information

从整体上可以看出，GAM和CBAM注意力机制还是比较相似的，同样是使用了通道注意力机制和空间注意力机制。但是不同的是对通道注意力和空间注意力的处理。
图片

1.3 ResBlock_CBAM

CBAM结构其实就是将通道注意力信息核空间注意力信息在一个block结构中进行运用。

在resnet中实现cbam：即在原始block和残差结构连接前，依次通过channel attention和spatial attention即可。

1.4性能评价

 2.Yolov5加入CBAM、GAM

2.1 CBAM加入common.py中

  1. class ChannelAttentionModule(nn.Module):  
  2.     def __init__(self, c1, reduction=16,light=False):
  3.         super(ChannelAttentionModule, self).__init__()
  4.         mid_channel = c1 // reduction
  5.         self.light=light
  6.         self.avg_pool = nn.AdaptiveAvgPool2d(1)
  7.         if self.light:
  8.             self.max_pool = nn.AdaptiveMaxPool2d(1) 
  9.             self.shared_MLP = nn.Sequential(
  10.                 nn.Linear(in_features=c1, out_features=mid_channel),
  11.                 nn.LeakyReLU(0.1, inplace=True),
  12.                 nn.Linear(in_features=mid_channel, out_features=c1)
  13.             )
  14.         else:
  15.  
  16.             self.shared_MLP = nn.Conv2d(c1, c1, 1, 1, 0, bias=True)    
  17.         self.act = nn.Sigmoid()
  18.        
  19.     def forward(self, x):
  20.         if self.light: 
  21.             avgout = self.shared_MLP(self.avg_pool(x).view(x.size(0),-1)).unsqueeze(2).unsqueeze(3)
  22.             maxout = self.shared_MLP(self.max_pool(x).view(x.size(0),-1)).unsqueeze(2).unsqueeze(3)
  23.             fc_out=(avgout + maxout)
  24.         else:
  25.             fc_out=(self.shared_MLP(self.avg_pool(x)))
  26.         return x * self.act(fc_out)
  27.         
  28. class SpatialAttentionModule(nn.Module): ##update:coding-style FOR LIGHTING
  29.     def __init__(self, kernel_size=7):
  30.         super().__init__()
  31.         assert kernel_size in (3, 7), 'kernel size must be 3 or 7'
  32.         padding = 3 if kernel_size == 7 else 1
  33.         self.cv1 = nn.Conv2d(2, 1, kernel_size, padding=padding, bias=False)
  34.         self.act = nn.Sigmoid()
  35.     def forward(self, x):
  36.         return x * self.act(self.cv1(torch.cat([torch.mean(x, 1, keepdim=True), torch.max(x, 1, keepdim=True)[0]], 1)))
  37.  
  38. class CBAM(nn.Module):
  39.     def __init__(self, c1,c2,k=7):
  40.         super().__init__()
  41.         self.channel_attention = ChannelAttentionModule(c1)
  42.         self.spatial_attention = SpatialAttentionModule(k)
  43.  
  44.     def forward(self, x):
  45.         return self.spatial_attention(self.channel_attention(x))

2.2  GAM加入common.py中

  1. def channel_shuffle(x, groups=2):   ##shuffle channel 
  2.         #RESHAPE----->transpose------->Flatten 
  3.         B, C, H, W = x.size()
  4.         out = x.view(B, groups, C // groups, H, W).permute(0, 2, 1, 3, 4).contiguous()
  5.         out=out.view(B, C, H, W) 
  6.         return out
  7.  
  8. class GAM_Attention(nn.Module):
  9.    #https://paperswithcode.com/paper/global-attention-mechanism-retain-information
  10.     def __init__(self, c1, c2, group=True,rate=4):
  11.         super(GAM_Attention, self).__init__()
  12.         
  13.         self.channel_attention = nn.Sequential(
  14.             nn.Linear(c1, int(c1 / rate)),
  15.             nn.ReLU(inplace=True),
  16.             nn.Linear(int(c1 / rate), c1)
  17.         )
  18.         
  19.         
  20.         self.spatial_attention = nn.Sequential(
  21.             
  22.             nn.Conv2d(c1, c1//rate, kernel_size=7, padding=3,groups=rate)if group else nn.Conv2d(c1, int(c1 / rate), kernel_size=7, padding=3), 
  23.             nn.BatchNorm2d(int(c1 /rate)),
  24.             nn.ReLU(inplace=True),
  25.             nn.Conv2d(c1//rate, c2, kernel_size=7, padding=3,groups=rate) if group else nn.Conv2d(int(c1 / rate), c2, kernel_size=7, padding=3), 
  26.             nn.BatchNorm2d(c2)
  27.         )
  28.  
  29.     def forward(self, x):
  30.         
  31.         b, c, h, w = x.shape
  32.         x_permute = x.permute(0, 2, 3, 1).view(b, -1, c)
  33.         x_att_permute = self.channel_attention(x_permute).view(b, h, w, c)
  34.         x_channel_att = x_att_permute.permute(0, 3, 1, 2)
  35.        # x_channel_att=channel_shuffle(x_channel_att,4) #last shuffle 
  36.         x = x * x_channel_att
  37.  
  38.         x_spatial_att = self.spatial_attention(x).sigmoid()
  39.         x_spatial_att=channel_shuffle(x_spatial_att,4) #last shuffle 
  40.         out = x * x_spatial_att
  41.         #out=channel_shuffle(out,4) #last shuffle 
  42.         return out

2.4 GAM加入common.py中加入common.py中

  1. class ResBlock_CBAM(nn.Module):
  2.     def __init__(self, in_places, places, stride=1, downsampling=False, expansion=4):
  3.         super(ResBlock_CBAM, self).__init__()
  4.         self.expansion = expansion
  5.         self.downsampling = downsampling
  6.  
  7.         self.bottleneck = nn.Sequential(
  8.             nn.Conv2d(in_channels=in_places, out_channels=places, kernel_size=1, stride=1, bias=False),
  9.             nn.BatchNorm2d(places),
  10.             nn.LeakyReLU(0.1, inplace=True),
  11.             nn.Conv2d(in_channels=places, out_channels=places, kernel_size=3, stride=stride, padding=1, bias=False),
  12.             nn.BatchNorm2d(places),
  13.             nn.LeakyReLU(0.1, inplace=True),
  14.             nn.Conv2d(in_channels=places, out_channels=places * self.expansion, kernel_size=1, stride=1,
  15.                         bias=False),
  16.             nn.BatchNorm2d(places * self.expansion),
  17.         )
  18.         self.cbam = CBAM(c1=places * self.expansion, c2=places * self.expansion, )
  19.  
  20.         if self.downsampling:
  21.             self.downsample = nn.Sequential(
  22.                 nn.Conv2d(in_channels=in_places, out_channels=places * self.expansion, kernel_size=1, stride=stride,
  23.                             bias=False),
  24.                 nn.BatchNorm2d(places * self.expansion)
  25.             )
  26.         self.relu = nn.ReLU(inplace=True)
  27.  
  28.     def forward(self, x):
  29.         residual = x
  30.         out = self.bottleneck(x)
  31.         out = self.cbam(out)
  32.         if self.downsampling:
  33.             residual = self.downsample(x)
  34.  
  35.         out += residual
  36.         out = self.relu(out)
  37.         return out

2.3 CBAM、GAM加入yolo.py中

  1. if m in {
  2.                 Conv, GhostConv, Bottleneck, GhostBottleneck, SPP, SPPF, DWConv, MixConv2d, Focus, CrossConv,
  3.                 BottleneckCSP, C3, C3TR, C3SPP, C3Ghost, nn.ConvTranspose2d, DWConvTranspose2d, C3x, C2f,CBAM,ResBlock_CBAM,GAM_Attention}:

2.4 CBAM、GAM修改对应yaml

2.4.1 修改 yolov5s_cbam.yaml

  1. # parameters
  2. nc: 10  # number of classes
  3. depth_multiple: 0.33  # model depth multiple
  4. width_multiple: 0.50  # layer channel multiple
  5.  
  6. # anchors
  7. anchors:
  8.   #- [5,6, 7,9, 12,10]      # P2/4
  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. # YOLOv5 backbone
  14. backbone:
  15.   # [from, number, module, args]               # [c=channels,module,kernlsize,strides]
  16.   [[-1, 1, Conv, [64, 6, 2, 2]],  # 0-P1/2           [c=3,64*0.5=32,3]
  17.    [-1, 1, Conv, [128, 3, 2]],  # 1-P2/4    
  18.    [-1, 3, C3, [128]],                                
  19.    [-1, 1, Conv, [256, 3, 2]],  # 3-P3/8            
  20.    [-1, 6, C3, [256]],                         
  21.    [-1, 1, Conv, [512, 3, 2]],  # 5-P4/16       
  22.    [-1, 9, C3, [512]],                     
  23.    [-1, 1, Conv, [1024, 3, 2]],  # 7-P5/32
  24.    [-1, 3, C3, [1024]], 
  25.    [-1, 1, CBAM, [1024,7]], #9
  26.    [-1, 1, SPPF, [1024,5]], #10
  27.   ]
  28.  
  29. # YOLOv5 head
  30. head:
  31.   [[-1, 1, Conv, [512, 1, 1]], 
  32.    [-1, 1, nn.Upsample, [None, 2, 'nearest']],
  33.    [[-1, 6], 1, Concat, [1]],  # cat backbone P4
  34.    [-1, 3, C3, [512, False]],  # 14
  35.  
  36.    [-1, 1, Conv, [256, 1, 1]],
  37.    [-1, 1, nn.Upsample, [None, 2, 'nearest']],
  38.    [[-1, 4], 1, Concat, [1]],  # cat backbone P3
  39.    [-1, 3, C3, [256, False]],  # 18 (P3/8-small)
  40.    [-1, 1, CBAM, [256,7]],   #19
  41.    
  42.  
  43.    [-1, 1, Conv, [256, 3, 2]],
  44.    [[-1, 15], 1, Concat, [1]],  # cat head P4
  45.    [-1, 3, C3, [512, False]],  # 22 (P4/16-medium)       [256, 256, 1, False]  
  46.    [-1, 1, CBAM, [512,7]],
  47.  
  48.  
  49.    [-1, 1, Conv, [512, 3, 2]],                           #[256, 256, 3, 2] 
  50.    [[-1, 11], 1, Concat, [1]],  # cat head P5
  51.    [-1, 3, C3, [1024, False]],  # 25 (P5/32-large)       [512, 512, 1, False]
  52.    [-1, 1, CBAM, [1024,7]],     
  53.  
  54.  
  55.    [[19, 23, 27], 1, Detect, [nc, anchors]],  # Detect(P3, P4, P5)
  56.   ]

2.4.2 修改 yolov5s_gam.yaml

  1. # parameters
  2. nc: 1  # number of classes
  3. depth_multiple: 0.33  # model depth multiple
  4. width_multiple: 0.50  # layer channel multiple
  5.  
  6. # anchors
  7. anchors:
  8.   #- [5,6, 7,9, 12,10]      # P2/4
  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. # YOLOv5 backbone
  14. backbone:
  15.   # [from, number, module, args]               # [c=channels,module,kernlsize,strides]
  16.   [[-1, 1, Conv, [64, 6, 2, 2]],  # 0-P1/2           [c=3,64*0.5=32,3]
  17.    [-1, 1, Conv, [128, 3, 2]],  # 1-P2/4    
  18.    [-1, 3, C3, [128]],                                
  19.    [-1, 1, Conv, [256, 3, 2]],  # 3-P3/8            
  20.    [-1, 6, C3, [256]],                         
  21.    [-1, 1, Conv, [512, 3, 2]],  # 5-P4/16       
  22.    [-1, 9, C3, [512]],                     
  23.    [-1, 1, Conv, [1024, 3, 2]],  # 7-P5/32
  24.    [-1, 3, C3, [1024]], 
  25.    [-1, 1, GAM_Attention, [1024,1024]], #9
  26.    [-1, 1, SPPF, [1024,5]], #10
  27.   ]
  28.  
  29. # YOLOv5 head
  30. head:
  31.   [[-1, 1, Conv, [512, 1, 1]], 
  32.    [-1, 1, nn.Upsample, [None, 2, 'nearest']],
  33.    [[-1, 6], 1, Concat, [1]],  # cat backbone P4
  34.    [-1, 3, C3, [512, False]],  # 14
  35.   
  36.  
  37.    [-1, 1, Conv, [256, 1, 1]],
  38.    [-1, 1, nn.Upsample, [None, 2, 'nearest']],
  39.    [[-1, 4], 1, Concat, [1]],  # cat backbone P3
  40.    [-1, 3, C3, [256, False]],  # 18 (P3/8-small)
  41.    [-1, 1, GAM_Attention, [256,256]],   #19
  42.    
  43.  
  44.    [-1, 1, Conv, [256, 3, 2]],
  45.    [[-1, 15], 1, Concat, [1]],  # cat head P4
  46.    [-1, 3, C3, [512, False]],  # 22 (P4/16-medium)       [256, 256, 1, False]  
  47.    [-1, 1, GAM_Attention, [512,512]],
  48.  
  49.  
  50.    [-1, 1, Conv, [512, 3, 2]],                           #[256, 256, 3, 2] 
  51.    [[-1, 11], 1, Concat, [1]],  # cat head P5
  52.    [-1, 3, C3, [1024, False]],  # 25 (P5/32-large)       [512, 512, 1, False]
  53.    [-1, 1, GAM_Attention, [1024,1024]], #
  54.  
  55.  
  56.    [[19, 23, 27], 1, Detect, [nc, anchors]],  # Detect(P3, P4, P5)
  57.   ]

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/IT小白/article/detail/873800
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。

  

闽ICP备14008679号