卷积神经网络模型之——GoogLeNet网络结构与代码实现

转x33g5p2x  于2022-07-26 转载在 其他  
字(7.8k)|赞(0)|评价(0)|浏览(524)

GoogLeNet网络简介

GoogLeNet原文地址：Going Deeper with Convolutions：https://www.cv-foundation.org/openaccess/content_cvpr_2015/papers/Szegedy_Going_Deeper_With_2015_CVPR_paper.pdf

GoogLeNet在2014年由Christian Szegedy提出，它是一种全新的深度学习结构。

GoogLeNet网络的主要创新点在于：

  1. 提出Inception结构在多个尺寸上同时进行卷积再聚合；

  1. 使用1X1的卷积进行降维以及映射处理；
  2. 添加两个辅助分类器帮助训练；
    辅助分类器是将中间某一层的输出用作分类,并按一个较小的权重加到最终分类结果中。
  3. 使用平均池化层代替全连接层，大大减少了参数量。

GoogLeNet网络结构

GoogLeNet的完整网络结构如下所示：

下面我们将其逐层拆分讲解并结合代码分析

Inception之前的几层结构

在进入Inception结构之前，GoogLeNet网络先堆叠了两个卷积（实则3个，有一个1X1的卷积）和两个最大池化层。

  1. # input(3,224,224)
  2. self.front = nn.Sequential(
  3.     nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3),   # output(64,112,112)
  4.     nn.ReLU(inplace=True),
  6.     nn.MaxPool2d(kernel_size=3,stride=2,ceil_mode=True),    # output(64,56,56)
  8.     nn.Conv2d(64,64,kernel_size=1),
  9.     nn.Conv2d(64,192,kernel_size=3,stride=1,padding=1),     # output(192,56,56)
  10.     nn.ReLU(inplace=True),
  12.     nn.MaxPool2d(kernel_size=3,stride=2,ceil_mode=True),    # output(192,28,28)
  13. )

Inception结构

Inception模块只会改变特征图的通道数，而不会改变尺寸大小。

Inception结构相对复杂，我们重新创建一个类来构建此结构，并通过参数不同的参数来控制各层的通道数。

  1. class Inception(nn.Module):
  2.     '''
  3.      in_channels: 输入通道数
  4.      out1x1：分支1输出通道数
  5.      in3x3：分支2的3x3卷积的输入通道数
  6.      out3x3：分支2的3x3卷积的输出通道数
  7.      in5x5：分支3的5x5卷积的输入通道数
  8.      out5x5：分支3的5x5卷积的输出通道数
  9.      pool_proj：分支4的最大池化层输出通道数
  10.     '''
  11.     def __init__(self,in_channels,out1x1,in3x3,out3x3,in5x5,out5x5,pool_proj):
  12.         super(Inception, self).__init__()
  14.         self.branch1 = nn.Sequential(
  15.             nn.Conv2d(in_channels, out1x1, kernel_size=1),
  16.             nn.ReLU(inplace=True)
  17.         )
  18.         self.branch2 = nn.Sequential(
  19.             nn.Conv2d(in_channels,in3x3,kernel_size=1),
  20.             nn.ReLU(inplace=True),
  21.             nn.Conv2d(in3x3,out3x3,kernel_size=3,padding=1),
  22.             nn.ReLU(inplace=True)
  23.         )
  24.         self.branch3 = nn.Sequential(
  25.             nn.Conv2d(in_channels, in5x5, kernel_size=1),
  26.             nn.ReLU(inplace=True),
  27.             nn.Conv2d(in5x5, out5x5, kernel_size=5, padding=2),
  28.             nn.ReLU(inplace=True)
  29.         )
  31.         self.branch4 = nn.Sequential(
  32.             nn.MaxPool2d(kernel_size=3,stride=1,padding=1),
  33.             nn.Conv2d(in_channels,pool_proj,kernel_size=1),
  34.             nn.ReLU(inplace=True)
  35.         )
  37.     def forward(self,x):
  38.         branch1 = self.branch1(x)
  39.         branch2 = self.branch2(x)
  40.         branch3 = self.branch3(x)
  41.         branch4 = self.branch4(x)
  43.         outputs = [branch1,branch2,branch3,branch4]
  44.         return torch.cat(outputs,1)	# 按通道数叠加

Inception3a模块

  1. # input(192,28,28)
  2. self.inception3a = Inception(192, 64, 96, 128, 16, 32, 32)		# output(256,28,28)

Inception3b + MaxPool

  1. # input(256,28,28)
  2. self.inception3b = Inception(256, 128, 128, 192, 32, 96, 64)		# output(480,28,28)
  3. self.maxpool3 = nn.MaxPool2d(3, stride=2, ceil_mode=True)			# output(480,14,14)

Inception4a

  1. # input(480,14,14)
  2. self.inception4a = Inception(480, 192, 96, 208, 16, 48, 64)		# output(512,14,14)

Inception4b

  1. # input(512,14,14)
  2. self.inception4b = Inception(512, 160, 112, 224, 24, 64, 64)		# output(512,14,14)

Inception4c

  1. # input(512,14,14)
  2. self.inception4c = Inception(512, 160, 112, 224, 24, 64, 64)		# output(512,14,14)

Inception4d

  1. # input(512,14,14)
  2. self.inception4d = Inception(512, 112, 144, 288, 32, 64, 64)		# output(528,14,14)

Inception4e+MaxPool

  1. # input(528,14,14)
  2. self.inception4e = Inception(528, 256, 160, 320, 32, 128, 128)	# output(832,14,14)
  3. self.maxpool4 = nn.MaxPool2d(3, stride=2, ceil_mode=True)		# output(832,7,7)

Inception5a

  1. # input(832,7,7)
  2. self.inception5a = Inception(832, 256, 160, 320, 32, 128, 128)		# output(832,7,7)

Inception5b

  1. # input(832,7,7)
  2. self.inception5b = Inception(832, 384, 192, 384, 48, 128, 128)		# output(1024,7,7)

Inception之后的几层结构

辅助分类模块

除了以上主干网络结构以外，GoogLeNet还提供了两个辅助分类模块，用于将中间某一层的输出用作分类，并按一个较小的权重(0.3)加到最终分类结果。

与Inception模块一样，我们也重新创建一个类来搭建辅助分类模块结构。

  1. class AccClassify(nn.Module):
  2. 	# in_channels: 输入通道
  3. 	# num_classes： 分类数
  4.     def __init__(self,in_channels,num_classes):
  5.         self.avgpool = nn.AvgPool2d(kernel_size=5, stride=3)
  6.         self.conv = nn.MaxPool2d(in_channels, 128, kernel_size=1)  # output[batch, 128, 4, 4]
  7.         self.relu = nn.ReLU(inplace=True)
  9.         self.fc1 = nn.Linear(2048, 1024)
  10.         self.fc2 = nn.Linear(1024, num_classes)
  12.     def forward(self,x):
  13.         x = self.avgpool(x)
  14.         x = self.conv(x)
  15.         x = self.relu(x)
  16.         x = torch.flatten(x, 1)
  17.         x = F.dropout(x, 0.5, training=self.training)
  18.         x = F.relu(self.fc1(x), inplace=True)
  19.         x = F.dropout(x, 0.5, training=self.training)
  20.         x = self.fc2(x)
  22.         return x

辅助分类模块1

第一个中间层输出位于Inception4a之后，将Inception4a的输出经过平均池化，1X1卷积和全连接后等到分类结果。

  1. self.acc_classify1 = AccClassify(512,num_classes)

辅助分类模块2

  1. self.acc_classify2 = AccClassify(528,num_classes)

整体网络结构

pytorch搭建完整代码

  1. """
  2. #-*-coding:utf-8-*- 
  3. # @author: wangyu a beginner programmer, striving to be the strongest.
  4. # @date: 2022/7/5 18:37
  5. """
  6. import torch.nn as nn
  7. import torch
  8. import torch.nn.functional as F
  10. class GoogLeNet(nn.Module):
  11.     def __init__(self,num_classes=1000,aux_logits=True):
  12.         super(GoogLeNet, self).__init__()
  13.         self.aux_logits = aux_logits
  15.         # input(3,224,224)
  16.         self.front = nn.Sequential(
  17.             nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3),   # output(64,112,112)
  18.             nn.ReLU(inplace=True),
  20.             nn.MaxPool2d(kernel_size=3,stride=2,ceil_mode=True),    # output(64,56,56)
  22.             nn.Conv2d(64,64,kernel_size=1),
  23.             nn.Conv2d(64,192,kernel_size=3,stride=1,padding=1),     # output(192,56,56)
  24.             nn.ReLU(inplace=True),
  26.             nn.MaxPool2d(kernel_size=3,stride=2,ceil_mode=True),    # output(192,28,28)
  27.         )
  29.         # input(192,28,28)
  30.         self.inception3a = Inception(192, 64, 96, 128, 16, 32, 32)  # output(64+128+32+32=256,28,28)
  31.         self.inception3b = Inception(256, 128, 128, 192, 32, 96, 64)  # output(480,28,28)
  32.         self.maxpool3 = nn.MaxPool2d(3, stride=2, ceil_mode=True)  # output(480,14,14)
  34.         self.inception4a = Inception(480, 192, 96, 208, 16, 48, 64)  # output(512,14,14)
  35.         self.inception4b = Inception(512, 160, 112, 224, 24, 64, 64)  # output(512,14,14)
  36.         self.inception4c = Inception(512, 128, 128, 256, 24, 64, 64)  # output(512,14,14)
  37.         self.inception4d = Inception(512, 112, 144, 288, 32, 64, 64)  # output(528,14,14)
  38.         self.inception4e = Inception(528, 256, 160, 320, 32, 128, 128)  # output(832,14,14)
  39.         self.maxpool4 = nn.MaxPool2d(3, stride=2, ceil_mode=True)  # output(832,7,7)
  41.         self.inception5a = Inception(832, 256, 160, 320, 32, 128, 128)  # output(832,7,7)
  42.         self.inception5b = Inception(832, 384, 192, 384, 48, 128, 128)  # output(1024,7,7)
  44.         if self.training and self.aux_logits:
  45.             self.acc_classify1 = AccClassify(512,num_classes)
  46.             self.acc_classify2 = AccClassify(528,num_classes)
  48.         self.avgpool = nn.AdaptiveAvgPool2d((1,1))        # output(1024,1,1)
  49.         self.dropout = nn.Dropout(0.4)
  50.         self.fc = nn.Linear(1024,num_classes)
  52.     def forward(self,x):
  53.         # input(3,224,224)
  54.         x = self.front(x)       # output(192,28,28)
  56.         x= self.inception3a(x)  # output(256,28,28)
  57.         x = self.inception3b(x)
  58.         x = self.maxpool3(x)
  60.         x = self.inception4a(x)
  62.         if self.training and self.aux_logits:
  63.             classify1 = self.acc_classify1(x)
  65.         x = self.inception4b(x)
  66.         x = self.inception4c(x)
  67.         x = self.inception4d(x)
  69.         if self.training and self.aux_logits:
  70.             classify2 = self.acc_classify2(x)
  72.         x = self.inception4e(x)
  73.         x = self.maxpool4(x)
  75.         x = self.inception5a(x)
  76.         x = self.inception5b(x)
  78.         x = self.avgpool(x)
  79.         x = torch.flatten(x,dims=1)
  80.         x = self.dropout(x)
  81.         x= self.fc(x)
  83.         if self.training and self.aux_logits:
  84.             return x,classify1,classify2
  86.         return x
  88. class Inception(nn.Module):
  89.     '''
  90.      in_channels: 输入通道数
  91.      out1x1：分支1输出通道数
  92.      in3x3：分支2的3x3卷积的输入通道数
  93.      out3x3：分支2的3x3卷积的输出通道数
  94.      in5x5：分支3的5x5卷积的输入通道数
  95.      out5x5：分支3的5x5卷积的输出通道数
  96.      pool_proj：分支4的最大池化层输出通道数
  97.     '''
  98.     def __init__(self,in_channels,out1x1,in3x3,out3x3,in5x5,out5x5,pool_proj):
  99.         super(Inception, self).__init__()
  101.         # input(192,28,28)
  102.         self.branch1 = nn.Sequential(
  103.             nn.Conv2d(in_channels, out1x1, kernel_size=1),
  104.             nn.ReLU(inplace=True)
  105.         )
  106.         self.branch2 = nn.Sequential(
  107.             nn.Conv2d(in_channels,in3x3,kernel_size=1),
  108.             nn.ReLU(inplace=True),
  109.             nn.Conv2d(in3x3,out3x3,kernel_size=3,padding=1),
  110.             nn.ReLU(inplace=True)
  111.         )
  112.         self.branch3 = nn.Sequential(
  113.             nn.Conv2d(in_channels, in5x5, kernel_size=1),
  114.             nn.ReLU(inplace=True),
  115.             nn.Conv2d(in5x5, out5x5, kernel_size=5, padding=2),
  116.             nn.ReLU(inplace=True)
  117.         )
  119.         self.branch4 = nn.Sequential(
  120.             nn.MaxPool2d(kernel_size=3,stride=1,padding=1),
  121.             nn.Conv2d(in_channels,pool_proj,kernel_size=1),
  122.             nn.ReLU(inplace=True)
  123.         )
  125.     def forward(self,x):
  126.         branch1 = self.branch1(x)
  127.         branch2 = self.branch2(x)
  128.         branch3 = self.branch3(x)
  129.         branch4 = self.branch4(x)
  131.         outputs = [branch1,branch2,branch3,branch4]
  132.         return torch.cat(outputs,1)
  134. class AccClassify(nn.Module):
  135.     def __init__(self,in_channels,num_classes):
  136.         self.avgpool = nn.AvgPool2d(kernel_size=5, stride=3)
  137.         self.conv = nn.MaxPool2d(in_channels, 128, kernel_size=1)  # output[batch, 128, 4, 4]
  138.         self.relu = nn.ReLU(inplace=True)
  140.         self.fc1 = nn.Linear(2048, 1024)
  141.         self.fc2 = nn.Linear(1024, num_classes)
  143.     def forward(self,x):
  144.         x = self.avgpool(x)
  145.         x = self.conv(x)
  146.         x = self.relu(x)
  147.         x = torch.flatten(x, 1)
  148.         x = F.dropout(x, 0.5, training=self.training)
  149.         x = F.relu(self.fc1(x), inplace=True)
  150.         x = F.dropout(x, 0.5, training=self.training)
  151.         x = self.fc2(x)
  153.         return x
  155. # print(GoogLeNet())

结构图

版权说明 : 本文为转载文章, 版权归原作者所有 版权申明

原文链接 : https://wang11.blog.csdn.net/article/details/125626247

内容来源于网络，如有侵权，请联系作者删除！

pytorch深度学习cnn神经网络GoogLeNet

相关文章

    查看更多

    热门标签

    更多
    JavaquerypythonNode开发语言requestUtil数据库Table后端算法LoggerMessageElementParser

    最新文章

    更多
    • 蜀ICP备13028337号-1 大数据知识库 https://www.saoniuhuo.com © All rights reserved
    • 本站内容来源互联网,如果侵犯您的权益请联系我们删除, 联系方式：448109455@qq.com