0. 引言

CSP-Darknet53无论是其作为CV Backbone，还是说它在别的数据集上取得极好的效果。与此同时，它与别的网络的适配能力极强。这些特点都在宣告：CSP-Darknet53的重要性。
关于原理部分的内容请查看这里CV 经典主干网络 (Backbone) 系列: CSPNet

1. 网络结构图

具体网络结构可以参考YOLO V3详解(一)：网络结构介绍中使用的工具来进行操作。具体网址和对应的权重文件下载地址如下：
模型可视化工具：https://lutzroeder.github.io/netron/
cfg文件下载网址：https://github.com/WongKinYiu/CrossStagePartialNetworks

得到的部分网络结构图的如下所示。

1.1 输入部分

1.2 CSP部分结构

1.3 输出部分

2. 代码实现

2.1 代码整体实现

通过代码实现CSP-Darknet53。框架为PyTorch，代码整体框架实现如下所示：

class CsDarkNet53(nn.Module):
    def __init__(self, num_classes):
        super(CsDarkNet53, self).__init__()

        input_channels = 32

        # Network
        self.stage1 = Conv2dBatchLeaky(3, input_channels, 3, 1, activation='mish')
        self.stage2 = Stage2(input_channels)
        self.stage3 = Stage3(4*input_channels)
        self.stage4 = Stage(4*input_channels, 8)
        self.stage5 = Stage(8*input_channels, 8)
        self.stage6 = Stage(16*input_channels, 4)

        self.conv = Conv2dBatchLeaky(32*input_channels, 32*input_channels, 1, 1, activation='mish')
        self.avgpool = nn.AdaptiveAvgPool2d((1,1))
        self.fc = nn.Linear(1024, num_classes)

        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
            elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
                nn.init.constant_(m.weight, 1)
                nn.init.constant_(m.bias, 0)

    def forward(self, x):
        stage1 = self.stage1(x)
        stage2 = self.stage2(stage1)
        stage3 = self.stage3(stage2)
        stage4 = self.stage4(stage3)
        stage5 = self.stage5(stage4)
        stage6 = self.stage6(stage5)

        conv = self.conv(stage6)
        x = self.avgpool(conv)
        x = x.view(-1, 1024)
        x = self.fc(x)

        return x

2.2 代码各个阶段实现

在代码中，对各个阶段的具体实现如下所示：

class Mish(nn.Module):
    def __init__(self):
        super(Mish, self).__init__()

    def forward(self, x):
        return x * torch.tanh(F.softplus(x))

class Conv2dBatchLeaky(nn.Module):
    """
    This convenience layer groups a 2D convolution, a batchnorm and a leaky ReLU.
    """
    def __init__(self, in_channels, out_channels, kernel_size, stride, activation='leaky', leaky_slope=0.1):
        super(Conv2dBatchLeaky, self).__init__()

        # Parameters
        self.in_channels = in_channels
        self.out_channels = out_channels
        self.kernel_size = kernel_size
        self.stride = stride
        if isinstance(kernel_size, (list, tuple)):
            self.padding = [int(k/2) for k in kernel_size]
        else:
            self.padding = int(kernel_size/2)
        self.leaky_slope = leaky_slope
        # self.mish = Mish()

        # Layer
        if activation == "leaky":
            self.layers = nn.Sequential(
                nn.Conv2d(self.in_channels, self.out_channels, self.kernel_size, self.stride, self.padding, bias=False),
                nn.BatchNorm2d(self.out_channels),
                nn.LeakyReLU(self.leaky_slope, inplace=True)
            )
        elif activation == "mish":
            self.layers = nn.Sequential(
                nn.Conv2d(self.in_channels, self.out_channels, self.kernel_size, self.stride, self.padding, bias=False),
                nn.BatchNorm2d(self.out_channels),
                Mish()
            )
        elif activation == "linear":
            self.layers = nn.Sequential(
                nn.Conv2d(self.in_channels, self.out_channels, self.kernel_size, self.stride, self.padding, bias=False)
            )

    def __repr__(self):
        s = '{name} ({in_channels}, {out_channels}, kernel_size={kernel_size}, stride={stride}, padding={padding}, negative_slope={leaky_slope})'
        return s.format(name=self.__class__.__name__, **self.__dict__)

    def forward(self, x):
        x = self.layers(x)
        return x

class SmallBlock(nn.Module):

    def __init__(self, nchannels):
        super().__init__()
        self.features = nn.Sequential(
            Conv2dBatchLeaky(nchannels, nchannels, 1, 1, activation='mish'),
            Conv2dBatchLeaky(nchannels, nchannels, 3, 1, activation='mish')
        )
        # conv_shortcut
        '''
        参考 https://github.com/bubbliiiing/yolov4-pytorch
        shortcut后不接任何conv
        '''
        # self.active_linear = Conv2dBatchLeaky(nchannels, nchannels, 1, 1, activation='linear')
        # self.conv_shortcut = Conv2dBatchLeaky(nchannels, nchannels, 1, 1, activation='mish')


    def forward(self, data):
        short_cut = data + self.features(data)
        # active_linear = self.conv_shortcut(short_cut)

        return short_cut

# Stage1  conv [256,256,3]->[256,256,32]

class Stage2(nn.Module):

    def __init__(self, nchannels):
        super().__init__()
        # stage2 32
        self.conv1 = Conv2dBatchLeaky(nchannels, 2*nchannels, 3, 2, activation='mish')
        self.split0 = Conv2dBatchLeaky(2*nchannels, 2*nchannels, 1, 1, activation='mish')
        self.split1 = Conv2dBatchLeaky(2*nchannels, 2*nchannels, 1, 1, activation='mish')

        self.conv2 = Conv2dBatchLeaky(2*nchannels, nchannels, 1, 1, activation='mish')
        self.conv3 = Conv2dBatchLeaky(nchannels, 2*nchannels, 3, 1, activation='mish')

        self.conv4 = Conv2dBatchLeaky(2*nchannels, 2*nchannels, 1, 1, activation='mish')


    def forward(self, data):
        conv1 = self.conv1(data)
        split0 = self.split0(conv1)
        split1 = self.split1(conv1)
        conv2 = self.conv2(split1)
        conv3 = self.conv3(conv2)

        shortcut = split1 + conv3
        conv4 = self.conv4(shortcut)

        route = torch.cat([split0, conv4], dim=1)
        return route

class Stage3(nn.Module):
    def __init__(self, nchannels):
        super().__init__()
        # stage3 128
        self.conv1 = Conv2dBatchLeaky(nchannels, int(nchannels/2), 1, 1, activation='mish')
        self.conv2 = Conv2dBatchLeaky(int(nchannels/2), nchannels, 3, 2, activation='mish')

        self.split0 = Conv2dBatchLeaky(nchannels, int(nchannels/2), 1, 1, activation='mish')
        self.split1 = Conv2dBatchLeaky(nchannels, int(nchannels/2), 1, 1, activation='mish')

        self.block1 = SmallBlock(int(nchannels/2))
        self.block2 = SmallBlock(int(nchannels/2))

        self.conv3 = Conv2dBatchLeaky(int(nchannels/2), int(nchannels/2), 1, 1, activation='mish')

    def forward(self, data):
        conv1 = self.conv1(data)
        conv2 = self.conv2(conv1)

        split0 = self.split0(conv2)
        split1 = self.split1(conv2)

        block1 = self.block1(split1)
        block2 = self.block2(block1)

        conv3 = self.conv3(block2)

        route = torch.cat([split0, conv3], dim=1)

        return route

# Stage4 Stage5 Stage6
class Stage(nn.Module):
    def __init__(self, nchannels, nblocks):
        super().__init__()
        # stage4 : 128
        # stage5 : 256
        # stage6 : 512
        self.conv1 = Conv2dBatchLeaky(nchannels, nchannels, 1, 1, activation='mish')
        self.conv2 = Conv2dBatchLeaky(nchannels, 2*nchannels, 3, 2, activation='mish')
        self.split0 = Conv2dBatchLeaky(2*nchannels, nchannels, 1, 1, activation='mish')
        self.split1 = Conv2dBatchLeaky(2*nchannels, nchannels, 1, 1, activation='mish')
        blocks = []
        for i in range(nblocks):
            blocks.append(SmallBlock(nchannels))
        self.blocks = nn.Sequential(*blocks)
        self.conv4 = Conv2dBatchLeaky(nchannels, nchannels, 1, 1, activation='mish')

    def forward(self,data):
        conv1 = self.conv1(data)
        conv2 = self.conv2(conv1)

        split0 = self.split0(conv2)
        split1 = self.split1(conv2)
        blocks = self.blocks(split1)
        conv4 = self.conv4(blocks)
        route = torch.cat([split0, conv4], dim=1)

        return route

3. 代码测试

下面使用一个小例子来对代码进行测试。

if __name__ == "__main__":
    use_cuda = torch.cuda.is_available()
    if use_cuda:
        device = torch.device("cuda")
        cudnn.benchmark = True
    else:
        device = torch.device("cpu")

    darknet = CsDarkNet53(num_classes=10)
    darknet = darknet.cuda()
    with torch.no_grad():
        darknet.eval()
        data = torch.rand(1, 3, 256, 256)
        data = data.cuda()
        try:
            #print(darknet)
            summary(darknet,(3,256,256))
            print(darknet(data))
        except Exception as e:
            print(e)

代码的输出如下所示：

Total params: 26,627,434
Trainable params: 26,627,434
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.75
Forward/backward pass size (MB): 553.51
Params size (MB): 101.58
Estimated Total Size (MB): 655.83
----------------------------------------------------------------
tensor([[ 0.1690,  0.0798,  0.1836,  0.2414,  0.3855,  0.2437, -0.1422, -0.1855,
          0.1758, -0.2452]], device='cuda:0')

注意：输出中存在框架结构内容，这里没有将其写在博客中

4. 结论

CSP-Darknet53的代码结构结合着对应的代码实现一起看，可以有效帮助大家理解关于原理部分的内容。希望可以帮助到大家！！！
另外，关于代码中存在的一些小的部分可能会在后面进行介绍。