Inception Note

InceptionV3的PyTorch实现：https://github.com/pytorch/vision/blob/master/torchvision/models/inception.py

2a表示第2组的第1个Block，同一组的空间维度相同
但为何没有3a, 5a？

(299, 299, 3)

→【1a, Cout=32, f=3, s=2】→(149, 149, 32)

→【2a, Cout=32, f=3】→(147, 147, 32)→【2b, Cout=64, f=3, p=1】→(147, 147, 64)
→【max pool, f=3, s=2】→(73, 73, 64)

→【3b, Cout=80, f=1】→(73, 73, 80)

→【4a, Cout=192, f=3】→(71, 71, 192)→【max pool, f=3, s=2】→(35, 35, 192)

→【Mixed_5b, InceptionA】→(35, 35, 256)→【Mixed_5c, InceptionA】→(35, 35, 288)
→【Mixed_5d, InceptionA】→(35, 35, 288)

→【Mixed_6a, InceptionB】→(17, 17, 768)
→【Mixed_6b, InceptionC】→(17, 17, 768)→【Mixed_6c, InceptionC】→(17, 17, 768)
→【Mixed_6d, InceptionC】→(17, 17, 768)→【Mixed_6e, InceptionC】→(17, 17, 768)

→【Mixed_7a, InceptionD】→(8, 8, 1280)
→【Mixed_7b, InceptionE】→(8, 8, 2048)→【Mixed_7c, InceptionE】→(8, 8, 2048)

→【global avg pool】→(2048,)→【dropout】→(2048,)→【fc】→(1000,)

分支：→【Mixed_6e, InceptionC】→(17, 17, 768)→【InceptionAux】→(1000,)

pool层几个值得注意的地方
(147, 147, 64)→【max pool, f=3, s=2】→(73, 73, 64)
(71, 71, 192)→【max pool, f=3, s=2】→(35, 35, 192)，使用f=3（一般使用f=2）

InceptionA使用了3次，分别用在Mixed_5b, Mixed_5c, Mixed_5d中，包含参数pool_features，输入为(35, 35, in_channels)，输出固定为(35, 35, 224+pool features)

(35, 35, 192)→【Mixed_5b，InceptionA, pool_features=32】→(35, 35, 224+32=256)
(35, 35, 256)→【Mixed_5c，InceptionA, pool_features=64】→(35, 35, 224+64=288)
(35, 35, 288)→【Mixed_5d，InceptionA, pool_features=64】→(35, 35, 224+64=288)

以(35, 35, 192)→【Mixed_5b, InceptionA, pool_features=64】→(35, 35, 256)为例

输入：(35, 35, 192)

分支1：→【BasicConv2d, Cout=64, f=1】→(35, 35, 64)

分支2：→【BasicConv2d, Cout=48, f=1】→(35, 35, 48)
→【BasicConv2d, Cout=64, f=5, p=2】→(35, 35, 64)

分支3：→【BasicConv2d, Cout=64, f=1】→(35, 35, 64)
→【BasicConv2d, Cout=96, f=3, p=1】→(35, 35, 96)
→【BasicConv2d, Cout=96, f=3, p=1】→(35, 35, 96)

分支4：→【avg pool, f=3, s=1, p=1】→(35, 35, 192)
→【BasicConv2d, Cout=pool_features, f=1】→(35, 35, pool_features)

合并：(35, 35, 224+pool_features)

InceptionB只使用了1次，用在Mixed_6a中
(35, 35, 288)→【Mixed_6a, InceptionB】→(17, 17, 768)，空间维度减半，通道数增加到大约2.7倍

输入：(35, 35, 288)

分支1：→【BasicConv2d, Cout=384, f=3, s=2】→(17, 17, 384)

分支2：→【BasicConv2d, Cout=64, f=1】→(35, 35, 64)
→【BasicConv2d, Cout=96, f=3, p=1】→(35, 35, 96)
→【BasicConv2d, Cout=96, f=3, s=2】→(17, 17, 96)

分支3：→【max pool, f=3, s=2】→(17, 17, 288)

合并：(17, 17, 384+96+288=768)

InceptionC使用了4次，分别用在Mixed_6b, Mixed_6c, Mixed_6d, Mixed_6e中，输入和输出均为(17, 17, 768)，只是参数channels_7x7不同，参数channels_7x7简记为c7

Mixed_6b，c7=128
Mixed_6c，c7=160
Mixed_6d，c7=160
Mixed_6e，c7=192

输入：(17, 17, 768)

分支1：→【BasicConv2d, Cout=192, f=1】→(17, 17, 192)

分支2：→【BasicConv2d, Cout=c7, f=1】→(17, 17, c7)
→【BasicConv2d, Cout=c7, f=(1, 7), p=(0, 3)】→(17, 17, c7)
→【BasicConv2d, Cout=192, f=(7, 1), p=(3, 0)】→(17, 17, 192)

分支3：→【BasicConv2d, Cout=c7, f=1】→(17, 17, c7)
→【BasicConv2d, Cout=c7, f=(7, 1), p=(3, 0)】→(17, 17, c7)
→【BasicConv2d, Cout=c7, f=(1, 7), p=(0, 3)】→(17, 17, c7)
→【BasicConv2d, Cout=c7, f=(7, 1), p=(3, 0)】→(17, 17, c7)
→【BasicConv2d, Cout=192, f=(1, 7), p=(0, 3)】→(17, 17, 192)

分支4：→【avg pool, f=3, s=1, p=1】→(17, 17, 768)
→【BasicConv2d, Cout=192, f=1】→(17, 17, 192)

合并：(17, 17, 192×4=768)

InceptionD只使用了1次，用在Mixed_7a中
(17, 17, 768)→【Mixed_7a, InceptionD】→(8, 8, 1280)，空间维度减半，通道数增加到大约1.7倍

输入：(17, 17, 768)

分支1：→【BasicConv2d, Cout=192, f=1】→(17, 17, 192)
→【BasicConv2d, Cout=320, f=3, s=2】→(8, 8, 320)

分支2：→【BasicConv2d, Cout=192, f=1】→(17, 17, 192)
→【BasicConv2d, Cout=192, f=(1, 7), p=(0, 3)】→(17, 17, 192)
→【BasicConv2d, Cout=192, f=(7, 1), p=(3, 0)】→(17, 17, 192)
→【BasicConv2d, Cout=192, f=3, s=2】→(8, 8, 192)

分支3：→【max pool, f=3, s=2】→(8, 8, 768)

合并：(8, 8, 320+192+768=1280)

InceptionE使用了2次， 分别用在Mixed_7b, Mixed_7c中，输入为(8, 8, in_channels)，输出固定为(8, 8, 2048)
以(8, 8, 1280)→【Mixed_7b, InceptionE】→(8, 8, 2048)为例

输入：(8, 8, 1280)

分支1：→【BasicConv2d, Cout=320, f=1】→(8, 8, 320)

分支2：→【BasicConv2d, Cout=384, f=1】→(8, 8, 384)
  分支2-1：→【BasicConv2d, Cout=384, f=(1, 3), p=(0, 1)】→(8, 8, 384)
  分支2-2：→【BasicConv2d, Cout=384, f=(3, 1), p=(1, 0)】→(8, 8, 384)
  合并：(8, 8, 384×2=768)

分支3：→【BasicConv2d, Cout=448, f=1】→(8, 8, 448)
      →【BasicConv2d, Cout=384, f=3, p=1】→(8, 8, 384)
  分支3-1：→【BasicConv2d, Cout=384, f=(1, 3), p=(0, 1)】→(8, 8, 384)
  分支3-2：→【BasicConv2d, Cout=384, f=(3, 1), p=(1, 0)】→(8, 8, 384)
  合并：(8, 8, 384×2=768)

分支4：→【avg pool, f=3, s=1, p=1】→(8, 8, 1280)
→【BasicConv2d, Cout=192, f=1】→(8, 8, 192)

合并：(8, 8, 320+768+768+192=2048)

InceptionAux只用了1次，连接在【Mixed_6e, InceptionC】的输出(17, 17, 768)上

(17, 17, 768)
→【avg pool, f=5, s=3】→(5, 5, 768)
→【BasicConv2d, Cout=128, f=1】→(5, 5, 128)
→【BasicConv2d, Cout=768, f=5】→(1, 1, 768)
→【reshape】→(768,)→【fc】→(1000,)