7.4.googlenet - how to

https://www.cv-foundation.org/openaccess/content_cvpr_2015/html/Szegedy_Going_Deeper_With_2015_CVPR_paper.html

不错的论文解读: https://www.cnblogs.com/harrymore/p/17263443.html

相比于 nin，google net 的 inception 块有 4 条路径。注意在合并时是直接在 channels 维度拼接而不是相加。但 googlenet 的最后重新引入了 Linear 层。

1
import torch
2
from torch import nn
3
from torch.nn import functional as F
4
from d2l import torch as d2l
5

6

7
class Inception(nn.Module):
8
    # c1--c4是每条路径的输出通道数
9
    def __init__(self, in_channels, c1, c2, c3, c4, **kwargs):
10
        super(Inception, self).__init__(**kwargs)
11
        # 线路1，单1x1卷积层
12
        self.p1_1 = nn.Conv2d(in_channels, c1, kernel_size=1)
13
        # 线路2，1x1卷积层后接3x3卷积层
14
        self.p2_1 = nn.Conv2d(in_channels, c2[0], kernel_size=1)
15
        self.p2_2 = nn.Conv2d(c2[0], c2[1], kernel_size=3, padding=1)
40 collapsed lines
16
        # 线路3，1x1卷积层后接5x5卷积层
17
        self.p3_1 = nn.Conv2d(in_channels, c3[0], kernel_size=1)
18
        self.p3_2 = nn.Conv2d(c3[0], c3[1], kernel_size=5, padding=2)
19
        # 线路4，3x3最大汇聚层后接1x1卷积层
20
        self.p4_1 = nn.MaxPool2d(kernel_size=3, stride=1, padding=1)
21
        self.p4_2 = nn.Conv2d(in_channels, c4, kernel_size=1)
22

23
    def forward(self, x):
24
        p1 = F.relu(self.p1_1(x))
25
        p2 = F.relu(self.p2_2(F.relu(self.p2_1(x))))
26
        p3 = F.relu(self.p3_2(F.relu(self.p3_1(x))))
27
        p4 = F.relu(self.p4_2(self.p4_1(x)))
28
        # 在通道维度上连结输出
29
        return torch.cat((p1, p2, p3, p4), dim=1)
30

31
b1 = nn.Sequential(nn.Conv2d(1, 64, kernel_size=7, stride=2, padding=3),
32
            nn.ReLU(),
33
            nn.MaxPool2d(kernel_size=3, stride=2, padding=1))
34

35
b2 = nn.Sequential(nn.Conv2d(64, 64, kernel_size=1),
36
                   nn.ReLU(),
37
                   nn.Conv2d(64, 192, kernel_size=3, padding=1),
38
                   nn.ReLU(),
39
                   nn.MaxPool2d(kernel_size=3, stride=2, padding=1))
40

41
b3 = nn.Sequential(Inception(192, 64, (96, 128), (16, 32), 32),
42
                   Inception(256, 128, (128, 192), (32, 96), 64),
43
                   nn.MaxPool2d(kernel_size=3, stride=2, padding=1))
44
b4 = nn.Sequential(Inception(480, 192, (96, 208), (16, 48), 64),
45
                   Inception(512, 160, (112, 224), (24, 64), 64),
46
                   Inception(512, 128, (128, 256), (24, 64), 64),
47
                   Inception(512, 112, (144, 288), (32, 64), 64),
48
                   Inception(528, 256, (160, 320), (32, 128), 128),
49
                   nn.MaxPool2d(kernel_size=3, stride=2, padding=1))
50
b5 = nn.Sequential(Inception(832, 256, (160, 320), (32, 128), 128),
51
                   Inception(832, 384, (192, 384), (48, 128), 128),
52
                   nn.AdaptiveAvgPool2d((1,1)),
53
                   nn.Flatten())
54

55
net = nn.Sequential(b1, b2, b3, b4, b5, nn.Linear(1024, 10))