devbox
从前慢现在也慢
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

yolov5-7.0修改骨干网络Ghostnet\efficientnet_b0\MobileNetv3_small\ShuffleNetV2_yolov7 更换骨干网

作者：从前慢现在也慢 | 2024-04-11 02:54:19

赞

踩

yolov7 更换骨干网

修改骨干网络主要需要添加模块到models/common.py文件中，在models/yolo.py文件夹中进行注册，修改models下面的yaml文件中的网络结构。

一、添加Ghostnet网络，替换Backbone

（yolov5-7.0的models/hu/yolov5s-ghost已经包含整体替换）:

  1. class C3Ghost(C3):
  2.  
  3.     # C3 module with GhostBottleneck()
  4.  
  5.     def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):
  6.  
  7.         super().__init__(c1, c2, n, shortcut, g, e)
  8.  
  9.         c_ = int(c2 * e)  # hidden channels
  10.  
  11.         self.m = nn.Sequential(*(GhostBottleneck(c_, c_) for _ in range(n)))
  12.  
  13.  
  14.  

  1. class GhostConv(nn.Module):
  2.  
  3.     # Ghost Convolution https://github.com/huawei-noah/ghostnet
  4.  
  5.     def __init__(self, c1, c2, k=1, s=1, g=1, act=True):  # ch_in, ch_out, kernel, stride, groups
  6.  
  7.         super().__init__()
  8.  
  9.         c_ = c2 // 2  # hidden channels
  10.  
  11.         self.cv1 = Conv(c1, c_, k, s, None, g, act=act)
  12.  
  13.         self.cv2 = Conv(c_, c_, 5, 1, None, c_, act=act)
  14.  
  15.  
  16.  
  17.     def forward(self, x):
  18.  
  19.         y = self.cv1(x)
  20.  
  21.         return torch.cat((y, self.cv2(y)), 1)

  1. class GhostBottleneck(nn.Module):
  2.  
  3.     # Ghost Bottleneck https://github.com/huawei-noah/ghostnet
  4.  
  5.     def __init__(self, c1, c2, k=3, s=1):  # ch_in, ch_out, kernel, stride
  6.  
  7.         super().__init__()
  8.  
  9.         c_ = c2 // 2
  10.  
  11.         self.conv = nn.Sequential(
  12.  
  13.             GhostConv(c1, c_, 1, 1),  # pw
  14.  
  15.             DWConv(c_, c_, k, s, act=False) if s == 2 else nn.Identity(),  # dw
  16.  
  17.             GhostConv(c_, c2, 1, 1, act=False))  # pw-linear
  18.  
  19.         self.shortcut = nn.Sequential(DWConv(c1, c1, k, s, act=False), Conv(c1, c2, 1, 1,
  20.  
  21.                                                                             act=False)) if s == 2 else nn.Identity()
  22.  
  23.  
  24.  
  25.     def forward(self, x):
  26.  
  27.         return self.conv(x) + self.shortcut(x)

修改步骤：

1. common.py与yolo.py文件已经包含ghost相关模块的配置和引用；

2.yolov5s_Ghost_backnone.yaml文件配置

  1. # YOLOv5 ��� by Ultralytics, AGPL-3.0 license
  2.  
  3.  
  4.  
  5. # Parameters
  6.  
  7. nc: 80  # number of classes
  8.  
  9. depth_multiple: 0.33  # model depth multiple
  10.  
  11. width_multiple: 0.50  # layer channel multiple
  12.  
  13. anchors:
  14.  
  15.   - [10,13, 16,30, 33,23]  # P3/8
  16.  
  17.   - [30,61, 62,45, 59,119]  # P4/16
  18.  
  19.   - [116,90, 156,198, 373,326]  # P5/32
  20.  
  21.  
  22.  
  23. # YOLOv5 v6.0 backbone
  24.  
  25. backbone:
  26.  
  27.   # [from, number, module, args]
  28.  
  29.   [[-1, 1, Conv, [64, 6, 2, 2]],  # 0-P1/2
  30.  
  31.    [-1, 1, GhostConv, [128, 3, 2]],  # 1-P2/4
  32.  
  33.    [-1, 3, C3Ghost, [128]],
  34.  
  35.    [-1, 1, GhostConv, [256, 3, 2]],  # 3-P3/8
  36.  
  37.    [-1, 6, C3Ghost, [256]],
  38.  
  39.    [-1, 1, GhostConv, [512, 3, 2]],  # 5-P4/16
  40.  
  41.    [-1, 9, C3Ghost, [512]],
  42.  
  43.    [-1, 1, GhostConv, [1024, 3, 2]],  # 7-P5/32
  44.  
  45.    [-1, 3, C3Ghost, [1024]],
  46.  
  47.    [-1, 1, SPPF, [1024, 5]],  # 9
  48.  
  49.   ]
  50.  
  51.  
  52.  
  53. # YOLOv5 v6.0 head
  54.  
  55. head:
  56.  
  57.   [[-1, 1, Conv, [512, 1, 1]],
  58.  
  59.    [-1, 1, nn.Upsample, [None, 2, 'nearest']],
  60.  
  61.    [[-1, 6], 1, Concat, [1]],  # cat backbone P4
  62.  
  63.    [-1, 3, C3, [512, False]],  # 13
  64.  
  65.  
  66.  
  67.    [-1, 1, Conv, [256, 1, 1]],
  68.  
  69.    [-1, 1, nn.Upsample, [None, 2, 'nearest']],
  70.  
  71.    [[-1, 4], 1, Concat, [1]],  # cat backbone P3
  72.  
  73.    [-1, 3, C3, [256, False]],  # 17 (P3/8-small)
  74.  
  75.  
  76.  
  77.    [-1, 1, Conv, [256, 3, 2]],
  78.  
  79.    [[-1, 14], 1, Concat, [1]],  # cat head P4
  80.  
  81.    [-1, 3, C3, [512, False]],  # 20 (P4/16-medium)
  82.  
  83.  
  84.  
  85.    [-1, 1, Conv, [512, 3, 2]],
  86.  
  87.    [[-1, 10], 1, Concat, [1]],  # cat head P5
  88.  
  89.    [-1, 3, C3, [1024, False]],  # 23 (P5/32-large)
  90.  
  91.  
  92.  
  93.    [[17, 20, 23], 1, Detect, [nc, anchors]],  # Detect(P3, P4, P5)
  94.  
  95.   ]

二、添加efficientnet_b0网络

修改步骤：

1. common.py

  1. # ---------------------------- efficientnet_b0 start -------------------------------
  2.  
  3.  
  4.  
  5. class effb0(nn.Module):
  6.  
  7.     def __init__(self, *args) -> None:
  8.  
  9.         super().__init__()
  10.  
  11.         model = models.efficientnet_b0(pretrained=True)
  12.  
  13.         modules = list(model.children())
  14.  
  15.         modules = modules[0][args[1]:args[2]]
  16.  
  17.         self.model = nn.Sequential(*modules)
  18.  
  19.     def forward(self, x):
  20.  
  21.         return self.model(x)
  22.  
  23.  
  24.  
  25. # ---------------------------- efficientnet_b0 end ---------------------------------

2. yolo.py文件

parse_model函数中添加

  1.  elif m is effb0:    
  2.               
  3.             c2 = args[0]

3. yolov5s_efficientnet_b0.yaml文件配置

  1. # YOLOv5 ��� by Ultralytics, GPL-3.0 license
  2.  
  3.  
  4.  
  5. # Parameters
  6.  
  7. nc: 80  # number of classes
  8.  
  9. depth_multiple: 0.33  # model depth multiple
  10.  
  11. width_multiple: 0.25  # layer channel multiple
  12.  
  13. anchors:
  14.  
  15.   - [10,13, 16,30, 33,23]  # P3/8
  16.  
  17.   - [30,61, 62,45, 59,119]  # P4/16
  18.  
  19.   - [116,90, 156,198, 373,326]  # P5/32
  20.  
  21.  
  22.  
  23. # YOLOv5 v6.0 backbone
  24.  
  25. backbone:
  26.  
  27.   # [from, number, module, args]
  28.  
  29.   [[-1, 1, effb0, [40,0,4]],  # 0
  30.  
  31.    [-1, 1, effb0, [112,4,6]],  # 1
  32.  
  33.    [-1, 1, effb0, [1280,6,9]],  # 2
  34.  
  35.    [-1, 1, SPPF, [1024, 5]],  # 3
  36.  
  37.   ]
  38.  
  39.  
  40.  
  41. # YOLOv5 v6.0 head
  42.  
  43. head:
  44.  
  45.   [[-1, 1, Conv, [512, 1, 1]],
  46.  
  47.    [-1, 1, nn.Upsample, [None, 2, 'nearest']],
  48.  
  49.    [[-1, 1], 1, Concat, [1]],  # cat backbone P4
  50.  
  51.    [-1, 3, C3, [512, False]],  # 7
  52.  
  53.  
  54.  
  55.    [-1, 1, Conv, [256, 1, 1]],
  56.  
  57.    [-1, 1, nn.Upsample, [None, 2, 'nearest']],
  58.  
  59.    [[-1, 0], 1, Concat, [1]],  # cat backbone P3
  60.  
  61.    [-1, 3, C3, [256, False]],  # 11 (P3/8-small)
  62.  
  63.  
  64.  
  65.    [-1, 1, Conv, [256, 3, 2]],
  66.  
  67.    [[-1, 7], 1, Concat, [1]],  # cat head P4
  68.  
  69.    [-1, 3, C3, [512, False]],  # 14 (P4/16-medium)
  70.  
  71.  
  72.  
  73.    [-1, 1, Conv, [512, 3, 2]],
  74.  
  75.    [[-1, 3], 1, Concat, [1]],  # cat head P5
  76.  
  77.    [-1, 3, C3, [1024, False]],  # 17 (P5/32-large)
  78.  
  79.  
  80.  
  81.    [[11, 14, 17], 1, Detect, [nc, anchors]],  # Detect(P3, P4, P5)
  82.  
  83.   ]

三、添加MobileNetv3_small网络

修改步骤：

1. common.py

  1. # ---------------------------- MobileBlock small start -------------------------------
  2.  
  3. # Mobilenetv3Small
  4.  
  5. class SeBlock(nn.Module):
  6.  
  7.     def __init__(self, in_channel, reduction=4):
  8.  
  9.         super().__init__()
  10.  
  11.         self.Squeeze = nn.AdaptiveAvgPool2d(1)
  12.  
  13.  
  14.  
  15.         self.Excitation = nn.Sequential()
  16.  
  17.         self.Excitation.add_module('FC1', nn.Conv2d(in_channel, in_channel // reduction, kernel_size=1))  # 1*1卷积与此效果相同
  18.  
  19.         self.Excitation.add_module('ReLU', nn.ReLU())
  20.  
  21.         self.Excitation.add_module('FC2', nn.Conv2d(in_channel // reduction, in_channel, kernel_size=1))
  22.  
  23.         self.Excitation.add_module('Sigmoid', nn.Sigmoid())
  24.  
  25.  
  26.  
  27.     def forward(self, x):
  28.  
  29.         y = self.Squeeze(x)
  30.  
  31.         ouput = self.Excitation(y)
  32.  
  33.         return x * (ouput.expand_as(x))
  34.  
  35.  
  36.  
  37.  
  38.  
  39. class Conv_BN_HSwish(nn.Module):
  40.  
  41.     """
  43.     This equals to
  45.     def conv_3x3_bn(inp, oup, stride):
  47.         return nn.Sequential(
  49.             nn.Conv2d(inp, oup, 3, stride, 1, bias=False),
  51.             nn.BatchNorm2d(oup),
  53.             h_swish()
  55.         )
  57.     """
  58.  
  59.  
  60.  
  61.     def __init__(self, c1, c2, stride):
  62.  
  63.         super(Conv_BN_HSwish, self).__init__()
  64.  
  65.         self.conv = nn.Conv2d(c1, c2, 3, stride, 1, bias=False)
  66.  
  67.         self.bn = nn.BatchNorm2d(c2)
  68.  
  69.         self.act = nn.Hardswish()
  70.  
  71.  
  72.  
  73.     def forward(self, x):
  74.  
  75.         return self.act(self.bn(self.conv(x)))
  76.  
  77.  
  78.  
  79.  
  80.  
  81. class MobileNetV3_InvertedResidual(nn.Module):
  82.  
  83.     def __init__(self, inp, oup, hidden_dim, kernel_size, stride, use_se, use_hs):
  84.  
  85.         super(MobileNetV3_InvertedResidual, self).__init__()
  86.  
  87.         assert stride in [1, 2]
  88.  
  89.  
  90.  
  91.         self.identity = stride == 1 and inp == oup
  92.  
  93.  
  94.  
  95.         if inp == hidden_dim:
  96.  
  97.             self.conv = nn.Sequential(
  98.  
  99.                 # dw
  100.  
  101.                 nn.Conv2d(hidden_dim, hidden_dim, kernel_size, stride, (kernel_size - 1) // 2, groups=hidden_dim,
  102.  
  103.                           bias=False),
  104.  
  105.                 nn.BatchNorm2d(hidden_dim),
  106.  
  107.                 nn.Hardswish() if use_hs else nn.ReLU(),
  108.  
  109.                 # Squeeze-and-Excite
  110.  
  111.                 SeBlock(hidden_dim) if use_se else nn.Sequential(),
  112.  
  113.                 # pw-linear
  114.  
  115.                 nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False),
  116.  
  117.                 nn.BatchNorm2d(oup),
  118.  
  119.             )
  120.  
  121.         else:
  122.  
  123.             self.conv = nn.Sequential(
  124.  
  125.                 # pw
  126.  
  127.                 nn.Conv2d(inp, hidden_dim, 1, 1, 0, bias=False),
  128.  
  129.                 nn.BatchNorm2d(hidden_dim),
  130.  
  131.                 nn.Hardswish() if use_hs else nn.ReLU(),
  132.  
  133.                 # dw
  134.  
  135.                 nn.Conv2d(hidden_dim, hidden_dim, kernel_size, stride, (kernel_size - 1) // 2, groups=hidden_dim,
  136.  
  137.                           bias=False),
  138.  
  139.                 nn.BatchNorm2d(hidden_dim),
  140.  
  141.                 # Squeeze-and-Excite
  142.  
  143.                 SeBlock(hidden_dim) if use_se else nn.Sequential(),
  144.  
  145.                 nn.Hardswish() if use_hs else nn.ReLU(),
  146.  
  147.                 # pw-linear
  148.  
  149.                 nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False),
  150.  
  151.                 nn.BatchNorm2d(oup),
  152.  
  153.             )
  154.  
  155.  
  156.  
  157.     def forward(self, x):
  158.  
  159.         y = self.conv(x)
  160.  
  161.         if self.identity:
  162.  
  163.             return x + y
  164.  
  165.         else:
  166.  
  167.             return y
  168.  
  169.  
  170.  
  171. # ---------------------------- MobileBlock end ---------------------------------

2. yolo.py文件

在parse_model函数中引入 Conv_BN_HSwish, MobileNetV3_InvertedResidual模块   

  1. if m in {
  2.  
  3.                 Conv, GhostConv, Bottleneck, GhostBottleneck, SPP, SPPF, DWConv, MixConv2d, Focus, CrossConv,     #cyz
  4.  
  5.                 BottleneckCSP, C3, C3TR, C3SPP, C3Ghost, nn.ConvTranspose2d, DWConvTranspose2d, C3x,
  6.  
  7.                 Conv_BN_HSwish, MobileNetV3_InvertedResidual,conv_bn_relu_maxpool, Shuffle_Block}

 3.  yolov5s_MobileNetv3_small.yaml文件配置

  1. # YOLOv5 ��� by Ultralytics, GPL-3.0 license
  2.  
  3.  
  4.  
  5. # Parameters
  6.  
  7. nc: 1  # number of classes
  8.  
  9. depth_multiple: 1.0  # model depth multiple
  10.  
  11. width_multiple: 1.0  # layer channel multiple
  12.  
  13. anchors:
  14.  
  15.   - [10,13, 16,30, 33,23]  # P3/8
  16.  
  17.   - [30,61, 62,45, 59,119]  # P4/16
  18.  
  19.   - [116,90, 156,198, 373,326]  # P5/32
  20.  
  21.  
  22.  
  23.    # Mobilenetv3-small backbone
  24.  
  25.    # MobileNetV3_InvertedResidual [out_ch, hid_ch, k_s, stride, SE, HardSwish]
  26.  
  27. backbone:
  28.  
  29.   # [from, number, module, args]
  30.  
  31.   [[-1, 1, Conv_BN_HSwish, [16, 2]],                              # 0-p1/2
  32.  
  33.    [-1, 1, MobileNetV3_InvertedResidual, [16,  16, 3, 2, 1, 0]],  # 1-p2/4
  34.  
  35.    [-1, 1, MobileNetV3_InvertedResidual, [24,  72, 3, 2, 0, 0]],  # 2-p3/8
  36.  
  37.    [-1, 1, MobileNetV3_InvertedResidual, [24,  88, 3, 1, 0, 0]],  # 3
  38.  
  39.    [-1, 1, MobileNetV3_InvertedResidual, [40,  96, 5, 2, 1, 1]],  # 4-p4/16
  40.  
  41.    [-1, 1, MobileNetV3_InvertedResidual, [40, 240, 5, 1, 1, 1]],  # 5
  42.  
  43.    [-1, 1, MobileNetV3_InvertedResidual, [40, 240, 5, 1, 1, 1]],  # 6
  44.  
  45.    [-1, 1, MobileNetV3_InvertedResidual, [48, 120, 5, 1, 1, 1]],  # 7
  46.  
  47.    [-1, 1, MobileNetV3_InvertedResidual, [48, 144, 5, 1, 1, 1]],  # 8
  48.  
  49.    [-1, 1, MobileNetV3_InvertedResidual, [96, 288, 5, 2, 1, 1]],  # 9-p5/32
  50.  
  51.    [-1, 1, MobileNetV3_InvertedResidual, [96, 576, 5, 1, 1, 1]],  # 10
  52.  
  53.    [-1, 1, MobileNetV3_InvertedResidual, [96, 576, 5, 1, 1, 1]],  # 11
  54.  
  55.   ]
  56.  
  57.  
  58.  
  59. # YOLOv5 v6.0 head
  60.  
  61. head:
  62.  
  63.   [[-1, 1, Conv, [96, 1, 1]],  # 12
  64.  
  65.    [-1, 1, nn.Upsample, [None, 2, 'nearest']],
  66.  
  67.    [[-1, 8], 1, Concat, [1]],  # cat backbone P4
  68.  
  69.    [-1, 3, C3, [144, False]],  # 15
  70.  
  71.  
  72.  
  73.    [-1, 1, Conv, [144, 1, 1]], # 16
  74.  
  75.    [-1, 1, nn.Upsample, [None, 2, 'nearest']],
  76.  
  77.    [[-1, 3], 1, Concat, [1]],  # cat backbone P3
  78.  
  79.    [-1, 3, C3, [168, False]],  # 19 (P3/8-small)
  80.  
  81.  
  82.  
  83.    [-1, 1, Conv, [168, 3, 2]],
  84.  
  85.    [[-1, 16], 1, Concat, [1]], # cat head P4
  86.  
  87.    [-1, 3, C3, [312, False]],  # 22 (P4/16-medium)
  88.  
  89.  
  90.  
  91.    [-1, 1, Conv, [312, 3, 2]],
  92.  
  93.    [[-1, 12], 1, Concat, [1]], # cat head P5
  94.  
  95.    [-1, 3, C3, [408, False]],  # 25 (P5/32-large)
  96.  
  97.  
  98.  
  99.    [[19, 22, 25], 1, Detect, [nc, anchors]],  # Detect(P3, P4, P5)
  100.  
  101.   ]

四、添加ShuffleNetV2网络

修改步骤：

1.common.py

  1. # ---------------------------- ShuffleBlock start -------------------------------
  2.  
  3.  
  4.  
  5. # 通道重排，跨group信息交流
  6.  
  7. def channel_shuffle(x, groups):
  8.  
  9.     batchsize, num_channels, height, width = x.data.size()
  10.  
  11.     channels_per_group = num_channels // groups
  12.  
  13.  
  14.  
  15.     # reshape
  16.  
  17.     x = x.view(batchsize, groups,
  18.  
  19.                channels_per_group, height, width)
  20.  
  21.  
  22.  
  23.     x = torch.transpose(x, 1, 2).contiguous()
  24.  
  25.  
  26.  
  27.     # flatten
  28.  
  29.     x = x.view(batchsize, -1, height, width)
  30.  
  31.  
  32.  
  33.     return x
  34.  
  35.  
  36.  
  37.  
  38.  
  39. class conv_bn_relu_maxpool(nn.Module):
  40.  
  41.     def __init__(self, c1, c2):  # ch_in, ch_out
  42.  
  43.         super(conv_bn_relu_maxpool, self).__init__()
  44.  
  45.         self.conv = nn.Sequential(
  46.  
  47.             nn.Conv2d(c1, c2, kernel_size=3, stride=2, padding=1, bias=False),
  48.  
  49.             nn.BatchNorm2d(c2),
  50.  
  51.             nn.ReLU(inplace=True),
  52.  
  53.         )
  54.  
  55.         self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)
  56.  
  57.  
  58.  
  59.     def forward(self, x):
  60.  
  61.         return self.maxpool(self.conv(x))
  62.  
  63.  
  64.  
  65.  
  66.  
  67. class Shuffle_Block(nn.Module):
  68.  
  69.     def __init__(self, inp, oup, stride):
  70.  
  71.         super(Shuffle_Block, self).__init__()
  72.  
  73.  
  74.  
  75.         if not (1 <= stride <= 3):
  76.  
  77.             raise ValueError('illegal stride value')
  78.  
  79.         self.stride = stride
  80.  
  81.  
  82.  
  83.         branch_features = oup // 2
  84.  
  85.         assert (self.stride != 1) or (inp == branch_features << 1)
  86.  
  87.  
  88.  
  89.         if self.stride > 1:
  90.  
  91.             self.branch1 = nn.Sequential(
  92.  
  93.                 self.depthwise_conv(inp, inp, kernel_size=3, stride=self.stride, padding=1),
  94.  
  95.                 nn.BatchNorm2d(inp),
  96.  
  97.                 nn.Conv2d(inp, branch_features, kernel_size=1, stride=1, padding=0, bias=False),
  98.  
  99.                 nn.BatchNorm2d(branch_features),
  100.  
  101.                 nn.ReLU(inplace=True),
  102.  
  103.             )
  104.  
  105.  
  106.  
  107.         self.branch2 = nn.Sequential(
  108.  
  109.             nn.Conv2d(inp if (self.stride > 1) else branch_features,
  110.  
  111.                       branch_features, kernel_size=1, stride=1, padding=0, bias=False),
  112.  
  113.             nn.BatchNorm2d(branch_features),
  114.  
  115.             nn.ReLU(inplace=True),
  116.  
  117.             self.depthwise_conv(branch_features, branch_features, kernel_size=3, stride=self.stride, padding=1),
  118.  
  119.             nn.BatchNorm2d(branch_features),
  120.  
  121.             nn.Conv2d(branch_features, branch_features, kernel_size=1, stride=1, padding=0, bias=False),
  122.  
  123.             nn.BatchNorm2d(branch_features),
  124.  
  125.             nn.ReLU(inplace=True),
  126.  
  127.         )
  128.  
  129.  
  130.  
  131.     @staticmethod
  132.  
  133.     def depthwise_conv(i, o, kernel_size, stride=1, padding=0, bias=False):
  134.  
  135.         return nn.Conv2d(i, o, kernel_size, stride, padding, bias=bias, groups=i)
  136.  
  137.  
  138.  
  139.     def forward(self, x):
  140.  
  141.         if self.stride == 1:
  142.  
  143.             x1, x2 = x.chunk(2, dim=1)  # 按照维度1进行split
  144.  
  145.             out = torch.cat((x1, self.branch2(x2)), dim=1)
  146.  
  147.         else:
  148.  
  149.             out = torch.cat((self.branch1(x), self.branch2(x)), dim=1)
  150.  
  151.  
  152.  
  153.         out = channel_shuffle(out, 2)
  154.  
  155.  
  156.  
  157.         return out
  158.  
  159.  
  160.  
  161.  
  162.  
  163. # ---------------------------- ShuffleBlock end --------------------------------

2. yolo.py文件

在parse_model函数中引入 conv_bn_relu_maxpool, Shuffle_Block模块：

  1.  if m in {
  2.  
  3.                 Conv, GhostConv, Bottleneck, GhostBottleneck, SPP, SPPF, DWConv, MixConv2d, Focus, CrossConv,     #cyz
  4.  
  5.                 BottleneckCSP, C3, C3TR, C3SPP, C3Ghost, nn.ConvTranspose2d, DWConvTranspose2d, C3x,
  6.  
  7.                 Conv_BN_HSwish, MobileNetV3_InvertedResidual,conv_bn_relu_maxpool, Shuffle_Block}:
  8.  
  9.             c1, c2 = ch[f], args[0]
  10.  
  11.  
  12.

3. yolov5s_ShuffleNetV2.yaml文件配置

  1. # YOLOv5 ��� by Ultralytics, GPL-3.0 license
  2.  
  3.  
  4.  
  5. # Parameters
  6.  
  7. nc: 20  # number of classes
  8.  
  9. depth_multiple: 1.0  # model depth multiple
  10.  
  11. width_multiple: 1.0  # layer channel multiple
  12.  
  13. anchors:
  14.  
  15.   - [10,13, 16,30, 33,23]  # P3/8
  16.  
  17.   - [30,61, 62,45, 59,119]  # P4/16
  18.  
  19.   - [116,90, 156,198, 373,326]  # P5/32
  20.  
  21.  
  22.  
  23. # YOLOv5 v6.0 backbone
  24.  
  25. backbone:
  26.  
  27.   # [from, number, module, args]
  28.  
  29.   # Shuffle_Block: [out, stride]
  30.  
  31.   [[ -1, 1, conv_bn_relu_maxpool, [ 32 ] ], # 0-P2/4
  32.  
  33.    [ -1, 1, Shuffle_Block, [ 128, 2 ] ],  # 1-P3/8
  34.  
  35.    [ -1, 3, Shuffle_Block, [ 128, 1 ] ],  # 2
  36.  
  37.    [ -1, 1, Shuffle_Block, [ 256, 2 ] ],  # 3-P4/16
  38.  
  39.    [ -1, 7, Shuffle_Block, [ 256, 1 ] ],  # 4
  40.  
  41.    [ -1, 1, Shuffle_Block, [ 512, 2 ] ],  # 5-P5/32
  42.  
  43.    [ -1, 3, Shuffle_Block, [ 512, 1 ] ],  # 6
  44.  
  45.   ]
  46.  
  47.  
  48.  
  49. # YOLOv5 v6.0 head
  50.  
  51. head:
  52.  
  53.   [[-1, 1, Conv, [256, 1, 1]],
  54.  
  55.    [-1, 1, nn.Upsample, [None, 2, 'nearest']],
  56.  
  57.    [[-1, 4], 1, Concat, [1]],  # cat backbone P4
  58.  
  59.    [-1, 1, C3, [256, False]],  # 10
  60.  
  61.  
  62.  
  63.    [-1, 1, Conv, [128, 1, 1]],
  64.  
  65.    [-1, 1, nn.Upsample, [None, 2, 'nearest']],
  66.  
  67.    [[-1, 2], 1, Concat, [1]],  # cat backbone P3
  68.  
  69.    [-1, 1, C3, [128, False]],  # 14 (P3/8-small)
  70.  
  71.  
  72.  
  73.    [-1, 1, Conv, [128, 3, 2]],
  74.  
  75.    [[-1, 11], 1, Concat, [1]],  # cat head P4
  76.  
  77.    [-1, 1, C3, [256, False]],  # 17 (P4/16-medium)
  78.  
  79.  
  80.  
  81.    [-1, 1, Conv, [256, 3, 2]],
  82.  
  83.    [[-1, 7], 1, Concat, [1]],  # cat head P5
  84.  
  85.    [-1, 1, C3, [512, False]],  # 20 (P5/32-large)
  86.  
  87.  
  88.  
  89.    [[14, 17, 20], 1, Detect, [nc, anchors]],  # Detect(P3, P4, P5)
  90.  
  91.   ]

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/从前慢现在也慢/article/detail/402506
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