Points for attention in adding spp module to the network

spp The module draws on the idea of spatial pyramid , adopt SPP The module realizes the fusion of local features and global features , Enriching the expression ability of feature map , It is beneficial to the case that the target size difference in the image to be detected is large , Therefore, the accuracy of detection has been greatly improved . There's a lot of yolov3 The version of has been added spp modular , So I'm also in my yolov3（bubbliiiing edition ） Try to add spp modular . The specific code is as follows ：

class SPPBottleneck(nn.Module):
    """Spatial pyramid pooling layer used in YOLOv3-SPP"""

    def __init__(
        self, in_channels, out_channels, kernel_sizes=(5, 9, 13)
     ):
        super().__init__()
        hidden_channels = in_channels // 2
        self.conv1 = nn.Sequential(nn.Conv2d(in_channels, hidden_channels, 1, stride=1),
                                   nn.ReLU(),
                                   nn.BatchNorm2d(hidden_channels)
                                   )
        self.m = nn.ModuleList(
            [
                nn.MaxPool2d(kernel_size=ks, stride=1, padding=ks // 2)
                for ks in kernel_sizes
            ]
        )
        conv2_channels = hidden_channels * (len(kernel_sizes) + 1)
        self.conv2 = nn.Sequential(nn.Conv2d(conv2_channels, out_channels, 1, stride=1),
                                   nn.ReLU(),
                                   nn.BatchNorm2d(out_channels)
                                   )

    def forward(self, x):
       x = self.conv1(x)
       x = torch.cat([x] + [m(x) for m in self.m], dim=1)
       x = self.conv2(x)
       return x

Where the input and output are darknet Output of the last convolution layer in channel, these two items. channel It's the same size .

On the Internet body Add self.spp

self.spp = SPPBottleneck(1024,1024)

And then darknet First output load spp Module

x0 = self.spp(x0)

Besides , There is a small problem in adding the network , Forget to join bn The layer , Lead to loss stay in a high position without going down , Another small detail , Xi xi xi .