文章目录
- 一、CSPDarknet整体结构
- 二、CSPNet结构
- 2.1 里面小残差结构Bottleneck
- 2.2 外层大残差结构CSP
- 三、SPP结构
- 四、CSPDarknet结构
- 4.1 stem结构
- 4.2 dark2,3,4结构，dark3为例：
- 4.3 dark5结构
- 4.4 CSPDarkNet整体结构
- 四、完整代码
一、CSPDarknet整体结构
CSPDarknet主要借鉴Darknet53（yolov3中使用）的网络框架，并使用了SPP和CSPNet结构做了改进。
二、CSPNet结构
论文链接:https://arxiv.org/pdf/1911.11929
CSPNet由里层小残差Bottleneck结构和外层大残差结构CSP组成。
2.1 里面小残差结构Bottleneck
Bottleneck小残差结构图如下，其中Conv2D可以根据需要设置为普通卷积BaseConv或者是深度可分离卷积DWConv，激活函数可以选择relu或者leaky_relu。
代码实现
#--------------------------------------------------#
# activation func.
#--------------------------------------------------#
def get_activation(name="lrelu", inplace=True):if name == "relu":module = nn.ReLU(inplace=inplace)elif name == "lrelu":module = nn.LeakyReLU(0.1, inplace=inplace)else:raise AttributeError("Unsupported act type: {}".format(name))return module#--------------------------------------------------#
# BaseConv
#--------------------------------------------------#
class BaseConv(nn.Module):def __init__(self, in_channels, out_channels, ksize, stride, groups=1, bias=False, act="lrelu"):super().__init__()pad = (ksize - 1) // 2self.conv = nn.Conv2d(in_channels, out_channels, kernel_size=ksize, stride=stride, padding=pad, groups=groups, bias=bias)self.bn = nn.BatchNorm2d(out_channels, eps=0.001, momentum=0.03)self.act = get_activation(act, inplace=True)def forward(self, x):return self.act(self.bn(self.conv(x)))def fuseforward(self, x):return self.act(self.conv(x))#--------------------------------------------------#
# DWConv
#--------------------------------------------------#
class DWConv(nn.Module):def __init__(self, in_channels, out_channels, ksize, stride=1, act="lrelu"):super().__init__()self.dconv = BaseConv(in_channels, in_channels, ksize=ksize, stride=stride, groups=in_channels, act=act,)self.pconv = BaseConv(in_channels, out_channels, ksize=1, stride=1, groups=1, act=act)def forward(self, x):x = self.dconv(x)return self.pconv(x)#--------------------------------------------------#
# 残差结构的构建，小的残差结构
#--------------------------------------------------#
class Bottleneck(nn.Module):# Standard bottleneckdef __init__(self, in_channels, out_channels, shortcut=True, expansion=0.5, depthwise=False, act="lrelu"):super().__init__()hidden_channels = int(out_channels * expansion)Conv = DWConv if depthwise else BaseConv#--------------------------------------------------## 利用1x1卷积进行通道数的缩减。缩减率一般是50%#--------------------------------------------------#self.conv1 = BaseConv(in_channels, hidden_channels, 1, stride=1, act=act)#--------------------------------------------------## 利用3x3卷积进行通道数的拓张。并且完成特征提取#--------------------------------------------------#self.conv2 = Conv(hidden_channels, out_channels, 3, stride=1, act=act)self.use_add = shortcut and in_channels == out_channelsdef forward(self, x):y = self.conv2(self.conv1(x))if self.use_add:y = y + xreturn y
2.2 外层大残差结构CSP
CSPLayer外层大残差结构图如下
代码实现
#--------------------------------------------------#
# CSP
#--------------------------------------------------#
class CSPLayer(nn.Module):def __init__(self, in_channels, out_channels, n=1, shortcut=True, expansion=0.5, depthwise=False, act="lrelu"):# ch_in, ch_out, number, shortcut, groups, expansionsuper().__init__()hidden_channels = int(out_channels * expansion) #--------------------------------------------------## 主干部分的初次卷积#--------------------------------------------------#self.conv1 = BaseConv(in_channels, hidden_channels, 1, stride=1, act=act)#--------------------------------------------------## 大的残差边部分的初次卷积#--------------------------------------------------#self.conv2 = BaseConv(in_channels, hidden_channels, 1, stride=1, act=act)#-----------------------------------------------## 对堆叠的结果进行卷积的处理#-----------------------------------------------#self.conv3 = BaseConv(2 * hidden_channels, out_channels, 1, stride=1, act=act)#--------------------------------------------------## 根据循环的次数构建上述Bottleneck残差结构#--------------------------------------------------#module_list = [Bottleneck(hidden_channels, hidden_channels, shortcut, 1.0, depthwise, act=act) for _ in range(n)]self.m = nn.Sequential(*module_list)def forward(self, x):#-------------------------------## x_1是主干部分#-------------------------------#x_1 = self.conv1(x)#-------------------------------## x_2是大的残差边部分#-------------------------------#x_2 = self.conv2(x)#-----------------------------------------------## 主干部分利用残差结构堆叠继续进行特征提取#-----------------------------------------------#x_1 = self.m(x_1)#-----------------------------------------------## 主干部分和大的残差边部分进行堆叠#-----------------------------------------------#x = torch.cat((x_1, x_2), 1)#-----------------------------------------------## 对堆叠的结果进行卷积的处理#-----------------------------------------------#return self.conv3(x)
三、SPP结构
在一般的CNN结构中，在卷积层后面通常连接着全连接。而全连接层的特征数是固定的，所以在网络输入的时候，会固定输入的大小(fixed-size)。但在现实中，我们的输入的图像尺寸总是不能满足输入时要求的大小。然而通常的手法就是裁剪(crop)和拉伸(warp)。
这样做总是不好的：图像的纵横比(ratio aspect) 和 输入图像的尺寸是被改变的。这样就会扭曲原始的图像。而Kaiming He在这里提出了一个SPP(Spatial Pyramid Pooling)层能很好的解决这样的问题， 但SPP通常连接在最后一层卷基层。
论文链接：https://arxiv.org/pdf/1406.4729
代码实现
#--------------------------------------------------#
# SPP
#--------------------------------------------------#
class SPPBottleneck(nn.Module):def __init__(self, in_channels, out_channels, kernel_sizes=(5, 9, 13), activation="lrelu"):super().__init__()hidden_channels = in_channels // 2self.conv1 = BaseConv(in_channels, hidden_channels, 1, stride=1, act=activation)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 = BaseConv(conv2_channels, out_channels, 1, stride=1, act=activation)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
四、CSPDarknet结构
参考yolov3中的Darknet53结构，结合二，三节的CSP和SPP结构，就可以进行CSPDarknet完整网络结构的搭建。CSPDarkenet主体由input，stem，dark2，dark3，dark4，dark5组成。其中dark2,3,4结构类似，stem，dark5稍微有点区别。
4.1 stem结构
结构
代码实现
#-----------------------------------------------#
# 输入图片是640, 640, 3
# 初始的基本通道是64
#-----------------------------------------------#
base_channels = int(wid_mul * 64) # 64
base_depth = max(round(dep_mul * 3), 1) # 3#-----------------------------------------------#
# 利用卷积结构进行特征提取
# 640, 640, 3 -> 320, 320, 64
#-----------------------------------------------#
self.stem = Conv(3, base_channels, 6, 2)
4.2 dark2,3,4结构，dark3为例：
结构
代码实现
#-----------------------------------------------#
# 完成卷积之后，160, 160, 128 -> 80, 80, 256
# 完成CSPlayer之后，80, 80, 256 -> 80, 80, 256
#-----------------------------------------------#
self.dark3 = nn.Sequential(Conv(base_channels * 2, base_channels * 4, 3, 2, act=act),CSPLayer(base_channels * 4, base_channels * 4, n=base_depth * 3, depthwise=depthwise, act=act),
)
4.3 dark5结构
结构
代码实现
#-----------------------------------------------#
# 完成卷积之后，40, 40, 512 -> 20, 20, 1024
# 完成SPP之后，20, 20, 1024 -> 20, 20, 1024
# 完成CSPlayer之后，20, 20, 1024 -> 20, 20, 1024
#-----------------------------------------------#
self.dark5 = nn.Sequential(Conv(base_channels * 8, base_channels * 16, 3, 2, act=act),SPPBottleneck(base_channels * 16, base_channels * 16, activation=act),# SPPF(base_channels * 16, base_channels * 16,activation=act),CSPLayer(base_channels * 16, base_channels * 16, n=base_depth, shortcut=False, depthwise=depthwise, act=act),
)
4.4 CSPDarkNet整体结构
结构
代码实现
#--------------------------------------------------#
# CSPDarknet
#--------------------------------------------------#
class CSPDarknet(nn.Module):def __init__(self, dep_mul, wid_mul, out_features=("dark3", "dark4", "dark5"), depthwise=False, act="lrelu"):super().__init__()assert out_features, "please provide output features of Darknet"self.out_features = out_featuresConv = DWConv if depthwise else BaseConv#-----------------------------------------------## 输入图片是640, 640, 3# 初始的基本通道是64#-----------------------------------------------#base_channels = int(wid_mul * 64) # 64base_depth = max(round(dep_mul * 3), 1) # 3#-----------------------------------------------## 利用卷积结构进行特征提取# 640, 640, 3 -> 320, 320, 64#-----------------------------------------------#self.stem = Conv(3, base_channels, 6, 2)#-----------------------------------------------## 完成卷积之后，320, 320, 64 -> 160, 160, 128# 完成CSPlayer之后，160, 160, 128 -> 160, 160, 128#-----------------------------------------------#self.dark2 = nn.Sequential(Conv(base_channels, base_channels * 2, 3, 2, act=act),CSPLayer(base_channels * 2, base_channels * 2, n=base_depth, depthwise=depthwise, act=act),)#-----------------------------------------------## 完成卷积之后，160, 160, 128 -> 80, 80, 256# 完成CSPlayer之后，80, 80, 256 -> 80, 80, 256#-----------------------------------------------#self.dark3 = nn.Sequential(Conv(base_channels * 2, base_channels * 4, 3, 2, act=act),CSPLayer(base_channels * 4, base_channels * 4, n=base_depth * 3, depthwise=depthwise, act=act),)#-----------------------------------------------## 完成卷积之后，80, 80, 256 -> 40, 40, 512# 完成CSPlayer之后，40, 40, 512 -> 40, 40, 512#-----------------------------------------------#self.dark4 = nn.Sequential(Conv(base_channels * 4, base_channels * 8, 3, 2, act=act),CSPLayer(base_channels * 8, base_channels * 8, n=base_depth * 3, depthwise=depthwise, act=act),)#-----------------------------------------------## 完成卷积之后，40, 40, 512 -> 20, 20, 1024# 完成SPP之后，20, 20, 1024 -> 20, 20, 1024# 完成CSPlayer之后，20, 20, 1024 -> 20, 20, 1024#-----------------------------------------------#self.dark5 = nn.Sequential(Conv(base_channels * 8, base_channels * 16, 3, 2, act=act),SPPBottleneck(base_channels * 16, base_channels * 16, activation=act),# SPPF(base_channels * 16, base_channels * 16,activation=act),CSPLayer(base_channels * 16, base_channels * 16, n=base_depth, shortcut=False, depthwise=depthwise, act=act),)def forward(self, x):outputs = {}x = self.stem(x)outputs["stem"] = xx = self.dark2(x)outputs["dark2"] = x#-----------------------------------------------## dark3的输出为80, 80, 256，是一个有效特征层#-----------------------------------------------#x = self.dark3(x)outputs["dark3"] = x#-----------------------------------------------## dark4的输出为40, 40, 512，是一个有效特征层#-----------------------------------------------#x = self.dark4(x)outputs["dark4"] = x#-----------------------------------------------## dark5的输出为20, 20, 1024，是一个有效特征层#-----------------------------------------------#x = self.dark5(x)outputs["dark5"] = xreturn [v for k, v in outputs.items() if k in self.out_features]
四、完整代码
#!/usr/bin/env python3
# -*- coding:utf-8 -*-
# Copyright (c) Megvii, Inc. and its affiliates.import torch
from torch import nn
from torchsummary import summary#--------------------------------------------------#
# activation func.
#--------------------------------------------------#
def get_activation(name="lrelu", inplace=True):if name == "relu":module = nn.ReLU(inplace=inplace)elif name == "lrelu":module = nn.LeakyReLU(0.1, inplace=inplace)else:raise AttributeError("Unsupported act type: {}".format(name))return module#--------------------------------------------------#
# BaseConv (CBL)
#--------------------------------------------------#
class BaseConv(nn.Module):def __init__(self, in_channels, out_channels, ksize, stride, groups=1, bias=False, act="lrelu"):super().__init__()pad = (ksize - 1) // 2self.conv = nn.Conv2d(in_channels, out_channels, kernel_size=ksize, stride=stride, padding=pad, groups=groups, bias=bias)self.bn = nn.BatchNorm2d(out_channels, eps=0.001, momentum=0.03)self.act = get_activation(act, inplace=True)def forward(self, x):return self.act(self.bn(self.conv(x)))def fuseforward(self, x):return self.act(self.conv(x))#--------------------------------------------------#
# DWConv
#--------------------------------------------------#
class DWConv(nn.Module):def __init__(self, in_channels, out_channels, ksize, stride=1, act="lrelu"):super().__init__()self.dconv = BaseConv(in_channels, in_channels, ksize=ksize, stride=stride, groups=in_channels, act=act,)self.pconv = BaseConv(in_channels, out_channels, ksize=1, stride=1, groups=1, act=act)def forward(self, x):x = self.dconv(x)return self.pconv(x)#--------------------------------------------------#
# SPP
#--------------------------------------------------#
class SPPBottleneck(nn.Module):def __init__(self, in_channels, out_channels, kernel_sizes=(5, 9, 13), activation="lrelu"):super().__init__()hidden_channels = in_channels // 2self.conv1 = BaseConv(in_channels, hidden_channels, 1, stride=1, act=activation)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 = BaseConv(conv2_channels, out_channels, 1, stride=1, act=activation)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#--------------------------------------------------#
# 残差结构的构建，小的残差结构
#--------------------------------------------------#
class Bottleneck(nn.Module):# Standard bottleneckdef __init__(self, in_channels, out_channels, shortcut=True, expansion=0.5, depthwise=False, act="lrelu"):super().__init__()hidden_channels = int(out_channels * expansion)Conv = DWConv if depthwise else BaseConv#--------------------------------------------------## 利用1x1卷积进行通道数的缩减。缩减率一般是50%#--------------------------------------------------#self.conv1 = BaseConv(in_channels, hidden_channels, 1, stride=1, act=act)#--------------------------------------------------## 利用3x3卷积进行通道数的拓张。并且完成特征提取#--------------------------------------------------#self.conv2 = Conv(hidden_channels, out_channels, 3, stride=1, act=act)self.use_add = shortcut and in_channels == out_channelsdef forward(self, x):y = self.conv2(self.conv1(x))if self.use_add:y = y + xreturn y#--------------------------------------------------#
# CSP
#--------------------------------------------------#
class CSPLayer(nn.Module):def __init__(self, in_channels, out_channels, n=1, shortcut=True, expansion=0.5, depthwise=False, act="lrelu"):# ch_in, ch_out, number, shortcut, groups, expansionsuper().__init__()hidden_channels = int(out_channels * expansion) #--------------------------------------------------## 主干部分的初次卷积#--------------------------------------------------#self.conv1 = BaseConv(in_channels, hidden_channels, 1, stride=1, act=act)#--------------------------------------------------## 大的残差边部分的初次卷积#--------------------------------------------------#self.conv2 = BaseConv(in_channels, hidden_channels, 1, stride=1, act=act)#-----------------------------------------------## 对堆叠的结果进行卷积的处理#-----------------------------------------------#self.conv3 = BaseConv(2 * hidden_channels, out_channels, 1, stride=1, act=act)#--------------------------------------------------## 根据循环的次数构建上述Bottleneck残差结构#--------------------------------------------------#module_list = [Bottleneck(hidden_channels, hidden_channels, shortcut, 1.0, depthwise, act=act) for _ in range(n)]self.m = nn.Sequential(*module_list)def forward(self, x):#-------------------------------## x_1是主干部分#-------------------------------#x_1 = self.conv1(x)#-------------------------------## x_2是大的残差边部分#-------------------------------#x_2 = self.conv2(x)#-----------------------------------------------## 主干部分利用残差结构堆叠继续进行特征提取#-----------------------------------------------#x_1 = self.m(x_1)#-----------------------------------------------## 主干部分和大的残差边部分进行堆叠#-----------------------------------------------#x = torch.cat((x_1, x_2), 1)#-----------------------------------------------## 对堆叠的结果进行卷积的处理#-----------------------------------------------#return self.conv3(x)#--------------------------------------------------#
# CSPDarknet
#--------------------------------------------------#
class CSPDarknet(nn.Module):def __init__(self, dep_mul, wid_mul, out_features=("dark3", "dark4", "dark5"), depthwise=False, act="lrelu"):super().__init__()assert out_features, "please provide output features of Darknet"self.out_features = out_featuresConv = DWConv if depthwise else BaseConv#-----------------------------------------------## 输入图片是640, 640, 3# 初始的基本通道是64#-----------------------------------------------#base_channels = int(wid_mul * 64) # 64base_depth = max(round(dep_mul * 3), 1) # 3#-----------------------------------------------## 利用卷积结构进行特征提取# 640, 640, 3 -> 320, 320, 64#-----------------------------------------------#self.stem = Conv(3, base_channels, 6, 2)#-----------------------------------------------## 完成卷积之后，320, 320, 64 -> 160, 160, 128# 完成CSPlayer之后，160, 160, 128 -> 160, 160, 128#-----------------------------------------------#self.dark2 = nn.Sequential(Conv(base_channels, base_channels * 2, 3, 2, act=act),CSPLayer(base_channels * 2, base_channels * 2, n=base_depth, depthwise=depthwise, act=act),)#-----------------------------------------------## 完成卷积之后，160, 160, 128 -> 80, 80, 256# 完成CSPlayer之后，80, 80, 256 -> 80, 80, 256#-----------------------------------------------#self.dark3 = nn.Sequential(Conv(base_channels * 2, base_channels * 4, 3, 2, act=act),CSPLayer(base_channels * 4, base_channels * 4, n=base_depth * 3, depthwise=depthwise, act=act),)#-----------------------------------------------## 完成卷积之后，80, 80, 256 -> 40, 40, 512# 完成CSPlayer之后，40, 40, 512 -> 40, 40, 512#-----------------------------------------------#self.dark4 = nn.Sequential(Conv(base_channels * 4, base_channels * 8, 3, 2, act=act),CSPLayer(base_channels * 8, base_channels * 8, n=base_depth * 3, depthwise=depthwise, act=act),)#-----------------------------------------------## 完成卷积之后，40, 40, 512 -> 20, 20, 1024# 完成SPP之后，20, 20, 1024 -> 20, 20, 1024# 完成CSPlayer之后，20, 20, 1024 -> 20, 20, 1024#-----------------------------------------------#self.dark5 = nn.Sequential(Conv(base_channels * 8, base_channels * 16, 3, 2, act=act),SPPBottleneck(base_channels * 16, base_channels * 16, activation=act),# SPPF(base_channels * 16, base_channels * 16,activation=act),CSPLayer(base_channels * 16, base_channels * 16, n=base_depth, shortcut=False, depthwise=depthwise, act=act),)def forward(self, x):outputs = {}x = self.stem(x)outputs["stem"] = xx = self.dark2(x)outputs["dark2"] = x#-----------------------------------------------## dark3的输出为80, 80, 256，是一个有效特征层#-----------------------------------------------#x = self.dark3(x)outputs["dark3"] = x#-----------------------------------------------## dark4的输出为40, 40, 512，是一个有效特征层#-----------------------------------------------#x = self.dark4(x)outputs["dark4"] = x#-----------------------------------------------## dark5的输出为20, 20, 1024，是一个有效特征层#-----------------------------------------------#x = self.dark5(x)outputs["dark5"] = xreturn [v for k, v in outputs.items() if k in self.out_features]if __name__ == '__main__':dep_mul =1wid_mul = 1net = CSPDarknet(dep_mul,wid_mul,out_features=("dark3", "dark4", "dark5"), depthwise=False, act="lrelu")summary(net, input_size=(3, 320, 320), batch_size=2, device="cpu")
运行查看CSPDarknet完整结构信息
----------------------------------------------------------------Layer (type) Output Shape Param #
================================================================Conv2d-1 [2, 32, 160, 160] 3,456BatchNorm2d-2 [2, 32, 160, 160] 64LeakyReLU-3 [2, 32, 160, 160] 0BaseConv-4 [2, 32, 160, 160] 0Conv2d-5 [2, 64, 80, 80] 18,432BatchNorm2d-6 [2, 64, 80, 80] 128LeakyReLU-7 [2, 64, 80, 80] 0BaseConv-8 [2, 64, 80, 80] 0Conv2d-9 [2, 32, 80, 80] 2,048BatchNorm2d-10 [2, 32, 80, 80] 64LeakyReLU-11 [2, 32, 80, 80] 0BaseConv-12 [2, 32, 80, 80] 0Conv2d-13 [2, 32, 80, 80] 2,048BatchNorm2d-14 [2, 32, 80, 80] 64LeakyReLU-15 [2, 32, 80, 80] 0BaseConv-16 [2, 32, 80, 80] 0Conv2d-17 [2, 32, 80, 80] 1,024BatchNorm2d-18 [2, 32, 80, 80] 64LeakyReLU-19 [2, 32, 80, 80] 0BaseConv-20 [2, 32, 80, 80] 0Conv2d-21 [2, 32, 80, 80] 9,216BatchNorm2d-22 [2, 32, 80, 80] 64LeakyReLU-23 [2, 32, 80, 80] 0BaseConv-24 [2, 32, 80, 80] 0Bottleneck-25 [2, 32, 80, 80] 0Conv2d-26 [2, 64, 80, 80] 4,096BatchNorm2d-27 [2, 64, 80, 80] 128LeakyReLU-28 [2, 64, 80, 80] 0BaseConv-29 [2, 64, 80, 80] 0CSPLayer-30 [2, 64, 80, 80] 0Conv2d-31 [2, 128, 40, 40] 73,728BatchNorm2d-32 [2, 128, 40, 40] 256LeakyReLU-33 [2, 128, 40, 40] 0BaseConv-34 [2, 128, 40, 40] 0Conv2d-35 [2, 64, 40, 40] 8,192BatchNorm2d-36 [2, 64, 40, 40] 128LeakyReLU-37 [2, 64, 40, 40] 0BaseConv-38 [2, 64, 40, 40] 0Conv2d-39 [2, 64, 40, 40] 8,192BatchNorm2d-40 [2, 64, 40, 40] 128LeakyReLU-41 [2, 64, 40, 40] 0BaseConv-42 [2, 64, 40, 40] 0Conv2d-43 [2, 64, 40, 40] 4,096BatchNorm2d-44 [2, 64, 40, 40] 128LeakyReLU-45 [2, 64, 40, 40] 0BaseConv-46 [2, 64, 40, 40] 0Conv2d-47 [2, 64, 40, 40] 36,864BatchNorm2d-48 [2, 64, 40, 40] 128LeakyReLU-49 [2, 64, 40, 40] 0BaseConv-50 [2, 64, 40, 40] 0Bottleneck-51 [2, 64, 40, 40] 0Conv2d-52 [2, 64, 40, 40] 4,096BatchNorm2d-53 [2, 64, 40, 40] 128LeakyReLU-54 [2, 64, 40, 40] 0BaseConv-55 [2, 64, 40, 40] 0Conv2d-56 [2, 64, 40, 40] 36,864BatchNorm2d-57 [2, 64, 40, 40] 128LeakyReLU-58 [2, 64, 40, 40] 0BaseConv-59 [2, 64, 40, 40] 0Bottleneck-60 [2, 64, 40, 40] 0Conv2d-61 [2, 64, 40, 40] 4,096BatchNorm2d-62 [2, 64, 40, 40] 128LeakyReLU-63 [2, 64, 40, 40] 0BaseConv-64 [2, 64, 40, 40] 0Conv2d-65 [2, 64, 40, 40] 36,864BatchNorm2d-66 [2, 64, 40, 40] 128LeakyReLU-67 [2, 64, 40, 40] 0BaseConv-68 [2, 64, 40, 40] 0Bottleneck-69 [2, 64, 40, 40] 0Conv2d-70 [2, 128, 40, 40] 16,384BatchNorm2d-71 [2, 128, 40, 40] 256LeakyReLU-72 [2, 128, 40, 40] 0BaseConv-73 [2, 128, 40, 40] 0CSPLayer-74 [2, 128, 40, 40] 0Conv2d-75 [2, 256, 20, 20] 294,912BatchNorm2d-76 [2, 256, 20, 20] 512LeakyReLU-77 [2, 256, 20, 20] 0BaseConv-78 [2, 256, 20, 20] 0Conv2d-79 [2, 128, 20, 20] 32,768BatchNorm2d-80 [2, 128, 20, 20] 256LeakyReLU-81 [2, 128, 20, 20] 0BaseConv-82 [2, 128, 20, 20] 0Conv2d-83 [2, 128, 20, 20] 32,768BatchNorm2d-84 [2, 128, 20, 20] 256LeakyReLU-85 [2, 128, 20, 20] 0BaseConv-86 [2, 128, 20, 20] 0Conv2d-87 [2, 128, 20, 20] 16,384BatchNorm2d-88 [2, 128, 20, 20] 256LeakyReLU-89 [2, 128, 20, 20] 0BaseConv-90 [2, 128, 20, 20] 0Conv2d-91 [2, 128, 20, 20] 147,456BatchNorm2d-92 [2, 128, 20, 20] 256LeakyReLU-93 [2, 128, 20, 20] 0BaseConv-94 [2, 128, 20, 20] 0Bottleneck-95 [2, 128, 20, 20] 0Conv2d-96 [2, 128, 20, 20] 16,384BatchNorm2d-97 [2, 128, 20, 20] 256LeakyReLU-98 [2, 128, 20, 20] 0BaseConv-99 [2, 128, 20, 20] 0Conv2d-100 [2, 128, 20, 20] 147,456BatchNorm2d-101 [2, 128, 20, 20] 256LeakyReLU-102 [2, 128, 20, 20] 0BaseConv-103 [2, 128, 20, 20] 0Bottleneck-104 [2, 128, 20, 20] 0Conv2d-105 [2, 128, 20, 20] 16,384BatchNorm2d-106 [2, 128, 20, 20] 256LeakyReLU-107 [2, 128, 20, 20] 0BaseConv-108 [2, 128, 20, 20] 0Conv2d-109 [2, 128, 20, 20] 147,456BatchNorm2d-110 [2, 128, 20, 20] 256LeakyReLU-111 [2, 128, 20, 20] 0BaseConv-112 [2, 128, 20, 20] 0Bottleneck-113 [2, 128, 20, 20] 0Conv2d-114 [2, 256, 20, 20] 65,536BatchNorm2d-115 [2, 256, 20, 20] 512LeakyReLU-116 [2, 256, 20, 20] 0BaseConv-117 [2, 256, 20, 20] 0CSPLayer-118 [2, 256, 20, 20] 0Conv2d-119 [2, 512, 10, 10] 1,179,648BatchNorm2d-120 [2, 512, 10, 10] 1,024LeakyReLU-121 [2, 512, 10, 10] 0BaseConv-122 [2, 512, 10, 10] 0Conv2d-123 [2, 256, 10, 10] 131,072BatchNorm2d-124 [2, 256, 10, 10] 512LeakyReLU-125 [2, 256, 10, 10] 0BaseConv-126 [2, 256, 10, 10] 0MaxPool2d-127 [2, 256, 10, 10] 0MaxPool2d-128 [2, 256, 10, 10] 0MaxPool2d-129 [2, 256, 10, 10] 0Conv2d-130 [2, 512, 10, 10] 524,288BatchNorm2d-131 [2, 512, 10, 10] 1,024LeakyReLU-132 [2, 512, 10, 10] 0BaseConv-133 [2, 512, 10, 10] 0SPPBottleneck-134 [2, 512, 10, 10] 0Conv2d-135 [2, 256, 10, 10] 131,072BatchNorm2d-136 [2, 256, 10, 10] 512LeakyReLU-137 [2, 256, 10, 10] 0BaseConv-138 [2, 256, 10, 10] 0Conv2d-139 [2, 256, 10, 10] 131,072BatchNorm2d-140 [2, 256, 10, 10] 512LeakyReLU-141 [2, 256, 10, 10] 0BaseConv-142 [2, 256, 10, 10] 0Conv2d-143 [2, 256, 10, 10] 65,536BatchNorm2d-144 [2, 256, 10, 10] 512LeakyReLU-145 [2, 256, 10, 10] 0BaseConv-146 [2, 256, 10, 10] 0Conv2d-147 [2, 256, 10, 10] 589,824BatchNorm2d-148 [2, 256, 10, 10] 512LeakyReLU-149 [2, 256, 10, 10] 0BaseConv-150 [2, 256, 10, 10] 0Bottleneck-151 [2, 256, 10, 10] 0Conv2d-152 [2, 512, 10, 10] 262,144BatchNorm2d-153 [2, 512, 10, 10] 1,024LeakyReLU-154 [2, 512, 10, 10] 0BaseConv-155 [2, 512, 10, 10] 0CSPLayer-156 [2, 512, 10, 10] 0
================================================================
Total params: 4,212,672
Trainable params: 4,212,672
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 2.34
Forward/backward pass size (MB): 303.91
Params size (MB): 16.07
Estimated Total Size (MB): 322.32
----------------------------------------------------------------
)
![[安洵杯 2019]JustBase（替换表）](http://pic.xiahunao.cn/[安洵杯 2019]JustBase（替换表）)
分层解耦(三层架构/IOC-DI引入))
)
