YOLOv5/v7 更换骨干网络之 GhostNet

2023-02-28

self conv nn

论文地址：https://arxiv.org/abs/1911.11907代码地址：https://github.com/huawei-noah/ghostnet由于内存和计算资源有限，在嵌入式设备上部署卷积神经网络（CNN）很困难。特征图中的冗余是那些成功的神经网络的重要特征，但在神经架构设计中很

论文地址：https://arxiv.org/abs/1911.11907
代码地址：https://github.com/huawei-noah/ghostnet

由于内存和计算资源有限，在嵌入式设备上部署卷积神经网络（CNN）很困难。特征图中的冗余是那些成功的神经网络的重要特征，但在神经架构设计中很少研究。本文提出了一种新的Ghost模块，以从廉价的操作中生成更多的特征图。基于一组内在特征图，我们以低成本应用一系列线性变换来生成许多重影特征图，这些重影特征可以充分揭示内在特征的信息。所提出的Ghost模块可以作为即插即用组件来升级现有的卷积神经网络。Ghost瓶颈被设计为堆叠Ghost模块，然后可以轻松地建立轻量级GhostNet。在基准上进行的实验表明，所提出的Ghost模块是基线模型中卷积层的一个令人印象深刻的替代方案，并且我们的GhostNet可以在ImageNet ILSVRC2012分类数据集上以类似的计算成本实现比MobileNetV3更高的识别性能（例如，75.7%的前1精度）。代码可从https://github.com/huawei-noah/ghostnet获取。

GhostNet网络结构

将YOLOv5主干网络替换为GhostNet：

yolov5lGhost.yaml

# YOLOv5 🚀 by Ultralytics, GPL-3.0 license

# Parameters
nc: 80  # number of classes
depth_multiple: 1.0  # model depth multiple
width_multiple: 1.0  # layer channel multiple
anchors:
  - [10,13, 16,30, 33,23]  # P3/8
  - [30,61, 62,45, 59,119]  # P4/16
  - [116,90, 156,198, 373,326]  # P5/32

# Ghostnet backbone
backbone:
  # [from, number, module, args]
  [[-1, 1, Conv, [16, 3, 2, 1]],            # 0-P1/2  ch_out, kernel, stride, padding, groups
   [-1, 1, G_bneck, [16, 16, 3, 1]],        # 1  ch_out, ch_mid, dw-kernel, stride

   [-1, 1, G_bneck, [24, 48, 3, 2]],        # 2-P2/4
   [-1, 1, G_bneck, [24, 72, 3, 1]],        # 3

   [-1, 1, G_bneck, [40, 72, 3, 2, True]],  # 4-P3/8
   [-1, 1, G_bneck, [40, 120, 3, 1, True]], # 5

   [-1, 1, G_bneck, [80, 240, 3, 2]],        # 6-P4/16
   [-1, 3, G_bneck, [80, 184, 3, 1]],        # 7
   [-1, 1, G_bneck, [112, 480, 3, 1, True]],
   [-1, 1, G_bneck, [112, 480, 3, 1, True]],

   [-1, 1, G_bneck, [160, 672, 3, 2, True]], # 10-P5/32
   [-1, 1, G_bneck, [160, 960, 3, 1]],       # 11
   [-1, 1, G_bneck, [160, 960, 3, 1, True]],
   [-1, 1, G_bneck, [160, 960, 3, 1]],
   [-1, 1, G_bneck, [160, 960, 3, 1, True]],
   [-1, 1, Conv, [960]],
  ]

# YOLOv5 v6.0 head
head:
  [[-1, 1, Conv, [512, 1, 1]],
   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
   [[-1, 9], 1, Concat, [1]],  # cat backbone P4
   [-1, 3, C3, [512, False]],  # 20

   [-1, 1, Conv, [256, 1, 1]],
   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
   [[-1, 5], 1, Concat, [1]],  # cat backbone P3
   [-1, 3, C3, [256, False]],  # 24 (P3/8-small)

   [-1, 1, Conv, [256, 3, 2]],
   [[-1, 20], 1, Concat, [1]], # cat head P4
   [-1, 3, C3, [512, False]],  # 27 (P4/16-medium)

   [-1, 1, Conv, [512, 3, 2]],
   [[-1, 16], 1, Concat, [1]],  # cat head P5
   [-1, 3, C3, [1024, False]],  # 30 (P5/32-large)

   [[23, 26, 29], 1, Detect, [nc, anchors]],  # Detect(P3, P4, P5)
  ]
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58

在YOLOv5项目中添加方式：

common.py中加入以下代码：

class SeBlock(nn.Module):
    def __init__(self, in_channel, reduction=4):
        super().__init__()
        self.Squeeze = nn.AdaptiveAvgPool2d(1)

        self.Excitation = nn.Sequential()
        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
        self.Excitation.add_module('ReLU', nn.ReLU())
        self.Excitation.add_module('FC2', nn.Conv2d(in_channel // reduction, in_channel, kernel_size=1))
        self.Excitation.add_module('Sigmoid', nn.Sigmoid())

    def forward(self, x):
        y = self.Squeeze(x)
        ouput = self.Excitation(y)
        return x * (ouput.expand_as(x))

class G_bneck(nn.Module):
    # Ghost Bottleneck https://github.com/huawei-noah/ghostnet
    def __init__(self, c1, c2, midc, k=5, s=1, use_se = False):  # ch_in, ch_mid, ch_out, kernel, stride, use_se
        super().__init__()
        assert s in [1, 2]
        c_ = midc
        self.conv = nn.Sequential(GhostConv(c1, c_, 1, 1),              # Expansion
                                  Conv(c_, c_, 3, s=2, p=1, g=c_, act=False) if s == 2 else nn.Identity(),  # dw
                                  # Squeeze-and-Excite
                                  SeBlock(c_) if use_se else nn.Sequential(),
                                  GhostConv(c_, c2, 1, 1, act=False))   # Squeeze pw-linear

        self.shortcut = nn.Identity() if (c1 == c2 and s == 1) else \
                                                nn.Sequential(Conv(c1, c1, 3, s=s, p=1, g=c1, act=False), \
                                                Conv(c1, c2, 1, 1, act=False)) # 避免stride=2时 通道数改变的情况

    def forward(self, x):
        # print(self.conv(x).shape)
        # print(self.shortcut(x).shape)
        return self.conv(x) + self.shortcut(x)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36