【人工智能概论】 构建神经网络——以用InceptionNet解决MNIST任务为例

一. 整体思路

• 两条原则，四个步骤。

1.1 两条原则

1. 从宏观到微观
2. 把握数据形状

1.2 四个步骤

1. 准备数据
2. 构建模型
3. 确定优化策略
4. 完善训练与测试代码

二. 举例——用InceptionNet解决MNIST任务

2.1 模型简介

• InceptionNet的设计思路是通过增加网络宽度来获得更好的模型性能。
• 其核心在于基本单元Inception结构块，如下图：
  
• 通过纵向堆叠Inception块构建完整网络。

2.2 MNIST任务

• MNIST是入门级的机器学习任务；
• 它是一个手写数字识别的数据集。

2.3 完整的程序

# 调包
import torch
import torch.nn as nn
import torch.nn.functional as F
from torchvision import transforms
from torchvision import datasets
from torch.utils.data import DataLoader
import torch.optim as optim"""数据准备"""
batch_size = 64transform = transforms.Compose([transforms.ToTensor(),transforms.Normalize((0.1307,),(0.3081,))
])train_dataset = datasets.MNIST(root='./mnist/',train=True,download=True,transform=transform)
train_loader = DataLoader(train_dataset,shuffle=True,batch_size=batch_size)
test_dataset = datasets.MNIST(root='./mnist/',train=False,download=True,transform=transform)
test_loader = DataLoader(test_dataset,shuffle=False,batch_size=batch_size)"""构建模型"""
# 需要指定输入的通道数class Inceptiona(nn.Module): def __init__(self,in_channels):super(Inceptiona,self).__init__()self.branch1_1 = nn.Conv2d(in_channels , 16 , kernel_size= 1)self.branch5_5_1 =nn.Conv2d(in_channels, 16, kernel_size= 1)self.branch5_5_2 =nn.Conv2d(16,24,kernel_size=5,padding=2)self.branch3_3_1 = nn.Conv2d(in_channels, 16,kernel_size=1)self.branch3_3_2 = nn.Conv2d(16,24,kernel_size=3,padding=1)self.branch3_3_3 = nn.Conv2d(24,24,kernel_size=3,padding=1)self.branch_pooling = nn.Conv2d(in_channels,24,kernel_size=1)def forward(self,x):x1 = self.branch1_1(x)x2 = self.branch5_5_1(x)x2 = self.branch5_5_2(x2)x3 = self.branch3_3_1(x)x3 = self.branch3_3_2(x3)x3 = self.branch3_3_3(x3)x4 = F.avg_pool2d(x,kernel_size=3,stride = 1, padding=1)x4 = self.branch_pooling(x4)outputs = [x1,x2,x3,x4]return torch.cat(outputs,dim=1)      # 构建完整的网络
class Net(nn.Module):def __init__(self):super(Net,self).__init__()self.conv1 = nn.Conv2d(1,10,kernel_size=5)self.conv2 = nn.Conv2d(88,20,kernel_size=5)self.incep1 = Inceptiona(in_channels=10)self.incep2 = Inceptiona(in_channels=20)self.mp = nn.MaxPool2d(2)self.fc = nn.Linear(1408,10)def forward(self,x):batch_size = x.size(0)x = F.relu(self.mp(self.conv1(x)))x = self.incep1(x)x = F.relu(self.mp(self.conv2(x)))x = self.incep2(x)x = x.view(batch_size,-1)x = self.fc(x)return x"""确定优化策略"""
model = Net()
device = torch.device('cuda:0'if torch.cuda.is_available() else 'cpu')
model.to(device) # 指定设备criterion = torch.nn.CrossEntropyLoss()
optimizer = optim.SGD(model.parameters(),lr=0.01,momentum=0.5)"""完善训练与测试代码"""
def train(epoch):running_loss = 0.0for batch_index, data in enumerate(train_loader,0):inputs, target = data# 把数据和模型送到同一个设备上inputs, target = inputs.to(device), target.to(device)optimizer.zero_grad()outputs = model(inputs)loss = criterion(outputs,target)loss.backward()optimizer.step()running_loss += loss.item() # 用loss.item不会构建计算图，得到的不是张量，而是标量if batch_index % 300 == 299:# 每三百组计算一次平均损失print('[%d,%5d] loss: %.3f' %(epoch+1,batch_index+1,running_loss/300))# 给出的是平均每一轮的损失running_loss = 0.0def test():correct = 0total = 0with torch.no_grad(): # 测试的环节不用求梯度for data in test_loader:images , labels = dataimages , labels = images.to(device),  labels.to(device)outputs = model(images)_, predicted = torch.max(outputs.data,dim=1)total += labels.size(0)correct += (predicted == labels).sum().item()print('accuracy on test set: %d,%%'%(100*correct/total))return 100*correct/total  # 将测试的准确率返回# 执行训练
if __name__=='__main__':score_best = 0for epoch in range(10):train(epoch)score = test()if score > score_best:score_best = scoretorch.save(model.state_dict(), "model.pth")