CNN-GoogLeNet

2014年，ImageNet图像识别挑战赛，GoogleNet脱颖而出。

GoogleNet

名字上致敬了LeNet网络，但内在吸收了NiN中网络串联的思想。

Inception块

GoogleNet中的基础卷积块叫做Inception块，（Inception：盗梦空间，贼好看，可以了解下！~哈哈）

如果说NiN是串联，那么汲取了NiN网络的GoogleNet网络则是融入了并联的思想。

Inception块里有4条并行的线路。

从输入开始：

1. 1x1卷积层
2. 1x1卷积层 + 3x3卷积层
3. 1x1卷积层 + 5x5卷积层
4. 3x3最大池化层 + 1x1卷积层

1x1、3x3、5x5主要目的是为了抽取不同空间尺寸下的信息，而3x3、5x5卷积前的1x1卷积，目的是为了改变通道数的同时，降低模型复杂度；最后一个则是使用了3x3的最大池化层；最终将四条线路进行汇总，合并作为输出。

GoogleNet模型

与VGG一样，在主体卷积上使用了5个模块（block），每个模块之间都使用了步长为2的3x3最大池化层来减小输出高宽。

详细见如下：

Sequential(
  (0): Sequential(
    (0): Conv2d(1, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3))
    (1): ReLU()
    (2): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)
  )
  (1): Sequential(
    (0): Conv2d(64, 64, kernel_size=(1, 1), stride=(1, 1))
    (1): Conv2d(64, 192, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (2): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)
  )
  (2): Sequential(
    (0): Inception(
      (p1_1): Conv2d(192, 64, kernel_size=(1, 1), stride=(1, 1))
      (p2_1): Conv2d(192, 96, kernel_size=(1, 1), stride=(1, 1))
      (p2_2): Conv2d(96, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (p3_1): Conv2d(192, 16, kernel_size=(1, 1), stride=(1, 1))
      (p3_2): Conv2d(16, 32, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2))
      (p4_1): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)
      (p4_2): Conv2d(192, 32, kernel_size=(1, 1), stride=(1, 1))
    )
    (1): Inception(
      (p1_1): Conv2d(192, 64, kernel_size=(1, 1), stride=(1, 1))
      (p2_1): Conv2d(192, 96, kernel_size=(1, 1), stride=(1, 1))
      (p2_2): Conv2d(96, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (p3_1): Conv2d(192, 16, kernel_size=(1, 1), stride=(1, 1))
      (p3_2): Conv2d(16, 32, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2))
      (p4_1): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)
      (p4_2): Conv2d(192, 32, kernel_size=(1, 1), stride=(1, 1))
    )
    (2): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)
  )
  (3): Sequential(
    (0): Inception(
      (p1_1): Conv2d(480, 192, kernel_size=(1, 1), stride=(1, 1))
      (p2_1): Conv2d(480, 96, kernel_size=(1, 1), stride=(1, 1))
      (p2_2): Conv2d(96, 208, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (p3_1): Conv2d(480, 16, kernel_size=(1, 1), stride=(1, 1))
      (p3_2): Conv2d(16, 48, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2))
      (p4_1): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)
      (p4_2): Conv2d(480, 64, kernel_size=(1, 1), stride=(1, 1))
    )
    (1): Inception(
      (p1_1): Conv2d(512, 160, kernel_size=(1, 1), stride=(1, 1))
      (p2_1): Conv2d(512, 112, kernel_size=(1, 1), stride=(1, 1))
      (p2_2): Conv2d(112, 224, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (p3_1): Conv2d(512, 24, kernel_size=(1, 1), stride=(1, 1))
      (p3_2): Conv2d(24, 64, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2))
      (p4_1): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)
      (p4_2): Conv2d(512, 64, kernel_size=(1, 1), stride=(1, 1))
    )
    (2): Inception(
      (p1_1): Conv2d(512, 128, kernel_size=(1, 1), stride=(1, 1))
      (p2_1): Conv2d(512, 128, kernel_size=(1, 1), stride=(1, 1))
      (p2_2): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (p3_1): Conv2d(512, 24, kernel_size=(1, 1), stride=(1, 1))
      (p3_2): Conv2d(24, 64, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2))
      (p4_1): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)
      (p4_2): Conv2d(512, 64, kernel_size=(1, 1), stride=(1, 1))
    )
    (3): Inception(
      (p1_1): Conv2d(512, 112, kernel_size=(1, 1), stride=(1, 1))
      (p2_1): Conv2d(512, 144, kernel_size=(1, 1), stride=(1, 1))
      (p2_2): Conv2d(144, 288, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (p3_1): Conv2d(512, 32, kernel_size=(1, 1), stride=(1, 1))
      (p3_2): Conv2d(32, 64, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2))
      (p4_1): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)
      (p4_2): Conv2d(512, 64, kernel_size=(1, 1), stride=(1, 1))
    )
    (4): Inception(
      (p1_1): Conv2d(528, 256, kernel_size=(1, 1), stride=(1, 1))
      (p2_1): Conv2d(528, 160, kernel_size=(1, 1), stride=(1, 1))
      (p2_2): Conv2d(160, 320, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (p3_1): Conv2d(528, 32, kernel_size=(1, 1), stride=(1, 1))
      (p3_2): Conv2d(32, 128, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2))
      (p4_1): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)
      (p4_2): Conv2d(528, 128, kernel_size=(1, 1), stride=(1, 1))
    )
    (5): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)
  )
  (4): Sequential(
    (0): Inception(
      (p1_1): Conv2d(832, 256, kernel_size=(1, 1), stride=(1, 1))
      (p2_1): Conv2d(832, 160, kernel_size=(1, 1), stride=(1, 1))
      (p2_2): Conv2d(160, 320, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (p3_1): Conv2d(832, 32, kernel_size=(1, 1), stride=(1, 1))
      (p3_2): Conv2d(32, 128, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2))
      (p4_1): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)
      (p4_2): Conv2d(832, 128, kernel_size=(1, 1), stride=(1, 1))
    )
    (1): Inception(
      (p1_1): Conv2d(832, 384, kernel_size=(1, 1), stride=(1, 1))
      (p2_1): Conv2d(832, 192, kernel_size=(1, 1), stride=(1, 1))
      (p2_2): Conv2d(192, 384, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (p3_1): Conv2d(832, 48, kernel_size=(1, 1), stride=(1, 1))
      (p3_2): Conv2d(48, 128, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2))
      (p4_1): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)
      (p4_2): Conv2d(832, 128, kernel_size=(1, 1), stride=(1, 1))
    )
    (2): GlobalAvgPool2d()
  )
  # 和NiN一样使用了全局平均池化层，置高和宽为1
  (5): FlattenLayer()
  (6): Linear(in_features=1024, out_features=10, bias=True)
)