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它是一种全新的深度学习结构。

GoogLeNet完整的网络结构如下：

下面，我们将逐层解释，并结合代码分析

在进入Inception结构之前，GoogLeNet网络首先堆叠了两个卷积(实际上3个，1个X1卷积)和两个最大池化层。

# input(3,224,224) self.front = nn.Sequential(     nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3),   # output(64,112,112)     nn.ReLU(inplace=True),      nn.MaxPool2d(kernel_size=3,stride=2,ceil_mode=True),    # output(64,56,56)      nn.Conv2d(64,64,kernel_size=1),     nn.Conv2d(64,192,kernel_size=3,stride=1,padding=1),     # output(192,56,56)     nnspan class="token punctuation">.ReLU(inplace=True),

    nn.MaxPool2d(kernel_size=3,stride=2,ceil_mode=True),    # output(192,28,28)
)

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。

class Inception(nn.Module):
    ''' in_channels: 输入通道数 out1x1：分支1输出通道数 in3x3：分支2的3x3卷积的输入通道数 out3x3：分支2的3x3卷积的输出通道数 in5x5：分支3的5x5卷积的输入通道数 out5x5：分支3的5x5卷积的输出通道数 pool_proj：分支4的最大池化层输出通道数 '''
    def __init__(self,in_channels,out1x1,in3x3,out3x3,in5x5,out5x5,pool_proj):
        super(Inception, self).__init__()

        self.branch1 = nn.Sequential(
            nn.Conv2d(in_channels, out1x1, kernel_size=1),
            nn.ReLU(inplace=True)
        )
        self.branch2 = nn.Sequential(
            nn.Conv2d(in_channels,in3x3,kernel_size=1),
            nn.ReLU(inplace=True),
            nn.Conv2d(in3x3,out3x3,kernel_size=3,padding=1),
            nn.ReLU(inplace=True)
        )
        self.branch3 = nn.Sequential(
            nn.Conv2d(in_channels, in5x5, kernel_size=1),
            nn.ReLU(inplace=True),
            nn.Conv2d(in5x5, out5x5, kernel_size=5, padding=2),
            nn.ReLU(inplace=True)
        )

        self.branch4 = nn.Sequential(
            nn.MaxPool2d(kernel_size=3,stride=1,padding=1),
            nn.Conv2d(in_channels,pool_proj,kernel_size=1),
            nn.ReLU(inplace=True)
        )

    def forward(self,x):
        branch1 = self.branch1(x)
        branch2 = self.branch2(x)
        branch3 = self.branch3(x)
        branch4 = self.branch4(x)

        outputs = [branch1,branch2,branch3,branch4]
        return torch.cat(outputs,1)	# 按通道数叠加

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

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

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

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

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

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

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

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

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

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。

class AccClassify(nn.Module):
	# in_channels: 输入通道
	# num_classes： 分类数
    def __init__(self,in_channels,num_classes):
        self.avgpool = nn.AvgPool2d(kernel_size=5, stride=3)
        self.conv = nn.MaxPool2d(in_channels, 128, kernel_size=1)  # output[batch, 128, 4, 4]
        self.relu = nn.ReLU(inplace=True)

        self.fc1 = nn.Linear(2048, 1024)
        self.fc2 = nn.Linear(1024, num_classes)

    def forward(self,x):
        x = self.avgpool(x)
        x = self.conv(x)
        x = self.relu(x)
        x = torch.flatten(x, 1)
        x = F.dropout(x, 0.5, training=self.training)
        x = F.relu(self.fc1(x), inplace=True)
        x = F.dropout(x, 0.5, training=self.training)
        x = self.fc2(x)

        return x

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。

self.acc_classify1 = AccClassify(512,num_classes)

self.acc_classify2 = AccClassify(528,num_classes)