目录
一、GoogLeNet 简介
GoogLeNet 吸收了 NiN 中串联网络的思想，并在此基础上做了改进。GoogLeNet 的一个观点是，有时使用不同大小的卷积核的组合是有利的。在本节中，我们将介绍一个稍微简化的 GoogLeNet 版本。
二、Inception 块
在 GoogLeNet 中，基本的卷积块被称为 Inception 块，其结构如下：
Inception 块由四条并行路径组成。 前三条路径使用窗口大小为
1
×
1
1\times1
1×1、
3
×
3
3\times3
3×3 和
5
×
5
5\times5
5×5 的卷积层，从不同空间大小中提取信息。中间的两条路径在输入上执行
1
×
1
1\times1
1×1 卷积，以减少通道数，从而降低模型的复杂性。第四条路径使用
3
×
3
3\times3
3×3 最大汇聚层，然后使用
1
×
1
1\times1
1×1 卷积层来改变通道数。这四条路径都使用合适的填充来**使得输入与输出的高和宽一致**，最后我们将每条线路的输出**在通道维度上进行连接**，并构成 Inception 块的输出。
可以看出，Inception 块只改变输入的通道数而不改变其高和宽。
在 Inception 块中，通常调整的超参数是每条路径的输出通道数。
设四条路径从左往右的输出通道参数分别为
c
1
,
c
2
,
c
3
,
c
4
c_1,c_2,c_3,c_4
c1,c2,c3,c4。因第一条路径只有一个卷积层，所以
c
1
c_1
c1 是一个标量。而第二、三条路径各有两个卷积层，因此
c
2
,
c
3
c_2,c_3
c2,c3 各是元组（大小为 2）。由于最大池化不影响通道数，因此
c
4
c_4
c4 也是标量。
Inception 块的代码实现如下：
import torch
from torch import nn
from torch.nn import functional as F
classInception(nn.Module):def__init__(self, in_channels, c1, c2, c3, c4):super().__init__()
self.path_1 = nn.Conv2d(in_channels, c1, kernel_size=1)
self.path_2 = nn.Sequential(
nn.Conv2d(in_channels, c2[0], kernel_size=1),
nn.ReLU(),
nn.Conv2d(c2[0], c2[1], kernel_size=3, padding=1),)
self.path_3 = nn.Sequential(
nn.Conv2d(in_channels, c3[0], kernel_size=1),
nn.ReLU(),
nn.Conv2d(c3[0], c3[1], kernel_size=5, padding=2),)
self.path_4 = nn.Sequential(
nn.MaxPool2d(kernel_size=3, stride=1, padding=1),
nn.Conv2d(in_channels, c4, kernel_size=1),)defforward(self, x):
output_1 = F.relu(self.path_1(x))
output_2 = F.relu(self.path_2(x))
output_3 = F.relu(self.path_3(x))
output_4 = F.relu(self.path_4(x))return torch.cat((output_1, output_2, output_3, output_4), dim=1)
三、GoogLeNet 架构
GoogLeNet 架构如下：
其实现如下：
classGoogLeNet(nn.Module):def__init__(self):super().__init__()
self.block_1 = nn.Sequential(
nn.Conv2d(1,64, kernel_size=7, stride=2, padding=3),
nn.ReLU(),
nn.MaxPool2d(kernel_size=3, stride=2, padding=1),)
self.block_2 = nn.Sequential(
nn.Conv2d(64,64, kernel_size=1),
nn.ReLU(),
nn.Conv2d(64,192, kernel_size=3, padding=1),
nn.ReLU(),
nn.MaxPool2d(kernel_size=3, stride=2, padding=1),)
self.block_3 = nn.Sequential(
Inception(192,64,(96,128),(16,32),32),
Inception(256,128,(128,192),(32,96),64),
nn.MaxPool2d(kernel_size=3, stride=2, padding=1),)
self.block_4 = nn.Sequential(
Inception(480,192,(96,208),(16,48),64),
Inception(512,160,(112,224),(24,64),64),
Inception(512,128,(128,256),(24,64),64),
Inception(512,112,(144,288),(32,64),64),
Inception(528,256,(160,320),(32,128),128),
nn.MaxPool2d(kernel_size=3, stride=2, padding=1),)
self.block_5 = nn.Sequential(
Inception(832,256,(160,320),(32,128),128),
Inception(832,384,(192,384),(48,128),128),
nn.AdaptiveAvgPool2d((1,1)),
nn.Flatten(),
nn.Linear(1024,10),)defforward(self, x):
x = self.block_1(x)
x = self.block_2(x)
x = self.block_3(x)
x = self.block_4(x)
x = self.block_5(x)return x
四、训练/测试 GoogLeNet
GoogLeNet 模型的计算复杂，而且不如 VGG 那样便于修改通道数。 为此，我们将输入的高和宽从 224 降到 96，这简化了计算。
使用 FashionMNIST 数据集，设置 batch size 为 128，学习率为 0.07，在 NVIDIA GeForce RTX 3090 上训练 20 个 Epoch 的结果如下：
Epoch 20--------------------------------------------------
Train Avg Loss:0.168895, Train Accuracy:0.935033
Test Avg Loss:0.269051, Test Accuracy:0.907700--------------------------------------------------3514.5 samples/sec
--------------------------------------------------
Done!
附录：完整代码
import torch
import torchvision
from torch import nn
from torch.utils.data import DataLoader
from torchvision.transforms import ToTensor, Resize
from Experiment import Experiment as E
from torch.nn import functional as F
classInception(nn.Module):def__init__(self, in_channels, c1, c2, c3, c4):super().__init__()
self.path_1 = nn.Conv2d(in_channels, c1, kernel_size=1)
self.path_2 = nn.Sequential(
nn.Conv2d(in_channels, c2[0], kernel_size=1),
nn.ReLU(),
nn.Conv2d(c2[0], c2[1], kernel_size=3, padding=1),)
self.path_3 = nn.Sequential(
nn.Conv2d(in_channels, c3[0], kernel_size=1),
nn.ReLU(),
nn.Conv2d(c3[0], c3[1], kernel_size=5, padding=2),)
self.path_4 = nn.Sequential(
nn.MaxPool2d(kernel_size=3, stride=1, padding=1),
nn.Conv2d(in_channels, c4, kernel_size=1),)defforward(self, x):
output_1 = F.relu(self.path_1(x))
output_2 = F.relu(self.path_2(x))
output_3 = F.relu(self.path_3(x))
output_4 = F.relu(self.path_4(x))return torch.cat((output_1, output_2, output_3, output_4), dim=1)classGoogLeNet(nn.Module):def__init__(self):super().__init__()
self.block_1 = nn.Sequential(
nn.Conv2d(1,64, kernel_size=7, stride=2, padding=3),
nn.ReLU(),
nn.MaxPool2d(kernel_size=3, stride=2, padding=1),)
self.block_2 = nn.Sequential(
nn.Conv2d(64,64, kernel_size=1),
nn.ReLU(),
nn.Conv2d(64,192, kernel_size=3, padding=1),
nn.ReLU(),
nn.MaxPool2d(kernel_size=3, stride=2, padding=1),)
self.block_3 = nn.Sequential(
Inception(192,64,(96,128),(16,32),32),
Inception(256,128,(128,192),(32,96),64),
nn.MaxPool2d(kernel_size=3, stride=2, padding=1),)
self.block_4 = nn.Sequential(
Inception(480,192,(96,208),(16,48),64),
Inception(512,160,(112,224),(24,64),64),
Inception(512,128,(128,256),(24,64),64),
Inception(512,112,(144,288),(32,64),64),
Inception(528,256,(160,320),(32,128),128),
nn.MaxPool2d(kernel_size=3, stride=2, padding=1),)
self.block_5 = nn.Sequential(
Inception(832,256,(160,320),(32,128),128),
Inception(832,384,(192,384),(48,128),128),
nn.AdaptiveAvgPool2d((1,1)),
nn.Flatten(),
nn.Linear(1024,10),)defforward(self, x):
x = self.block_1(x)
x = self.block_2(x)
x = self.block_3(x)
x = self.block_4(x)
x = self.block_5(x)return x
definit_net(m):iftype(m)== nn.Linear ortype(m)== nn.Conv2d:
nn.init.xavier_uniform_(m.weight)
transformer = torchvision.transforms.Compose([Resize(96), ToTensor()])
train_data = torchvision.datasets.FashionMNIST('/mnt/mydataset', train=True, transform=transformer, download=True)
test_data = torchvision.datasets.FashionMNIST('/mnt/mydataset', train=False, transform=transformer, download=True)
train_loader = DataLoader(train_data, batch_size=128, shuffle=True, num_workers=4)
test_loader = DataLoader(test_data, batch_size=128, num_workers=4)
googlenet = GoogLeNet()
googlenet.apply(init_net)
e = E(train_loader, test_loader, googlenet,30,0.07)
e.main()
e.show()
本文转载自: https://blog.csdn.net/raelum/article/details/124766391
版权归原作者 raelum 所有， 如有侵权，请联系我们删除。
版权归原作者 raelum 所有， 如有侵权，请联系我们删除。