一个调试通过的用pytorch实现mnist数据集数字识别

手写数字识别是深度学习界的“HELLO WORLD”。网上代码很多&＃xff0c;找一份自己读懂&＃xff0c;对整个学习网络理解会有帮助。
这是自动下载mnist数据集的&＃xff0c;如果下载过慢&＃xff0c;可以自己把mnist数据集下载好放到相应目录下面
MNIST数据集 提取码&＃xff1a;1234

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torchvision import datasets, transforms
from torch.autograd import Variable# Training settings
batch_size &＃61; 64# MNIST Dataset# MNIST数据集已经集成在pytorch datasets中&＃xff0c;可以直接调用train_dataset &＃61; datasets.MNIST(root&＃61;&＃39;./data/&＃39;,train&＃61;True,transform&＃61;transforms.ToTensor(),download&＃61;True)test_dataset &＃61; datasets.MNIST(root&＃61;&＃39;./data/&＃39;,train&＃61;False,transform&＃61;transforms.ToTensor())# Data Loader (Input Pipeline)
train_loader &＃61; torch.utils.data.DataLoader(dataset&＃61;train_dataset,batch_size&＃61;batch_size,shuffle&＃61;True)test_loader &＃61; torch.utils.data.DataLoader(dataset&＃61;test_dataset,batch_size&＃61;batch_size,shuffle&＃61;False)class Net(nn.Module):def __init__(self):super(Net, self).__init__()# 输入1通道&＃xff0c;输出10通道&＃xff0c;kernel 5*5self.conv1 &＃61; nn.Conv2d(in_channels&＃61;1, out_channels&＃61;10, kernel_size&＃61;5)self.conv2 &＃61; nn.Conv2d(10, 20, 5)self.conv3 &＃61; nn.Conv2d(20, 40, 3)self.mp &＃61; nn.MaxPool2d(2)# fully connectself.fc &＃61; nn.Linear(40, 10)#&＃xff08;in_features, out_features&＃xff09;def forward(self, x):# in_size &＃61; 64in_size &＃61; x.size(0) # one batch 此时的x是包含batchsize维度为4的tensor&＃xff0c;即(batchsize&＃xff0c;channels&＃xff0c;x&＃xff0c;y)&＃xff0c;x.size(0)指batchsize的值 把batchsize的值作为网络的in_size# x: 64*1*28*28x &＃61; F.relu(self.mp(self.conv1(x)))# x: 64*10*12*12 feature map &＃61;[(28-4)/2]^2&＃61;12*12x &＃61; F.relu(self.mp(self.conv2(x)))# x: 64*20*4*4x &＃61; F.relu(self.mp(self.conv3(x)))x &＃61; x.view(in_size, -1) # flatten the tensor 相当于resharp# print(x.size())# x: 64*320x &＃61; self.fc(x)# x:64*10# print(x.size())return F.log_softmax(x) #64*10
model &＃61; Net()
optimizer &＃61; optim.SGD(model.parameters(), lr&＃61;0.01, momentum&＃61;0.5)def train(epoch):for batch_idx, (data, target) in enumerate(train_loader):#batch_idx是enumerate&＃xff08;&＃xff09;函数自带的索引&＃xff0c;从0开始# data.size():[64, 1, 28, 28]# target.size():[64]output &＃61; model(data)#output:64*10loss &＃61; F.nll_loss(output, target)if batch_idx % 200 &＃61;&＃61; 0:print(&＃39;Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}&＃39;.format(epoch, batch_idx * len(data), len(train_loader.dataset),# 100. * batch_idx / len(train_loader), loss.data[0]))100. * batch_idx / len(train_loader), loss.data.item()))optimizer.zero_grad() # 所有参数的梯度清零loss.backward() #即反向传播求梯度optimizer.step() #调用optimizer进行梯度下降更新参数def test():test_loss &＃61; 0correct &＃61; 0for data, target in test_loader:data, target &＃61; Variable(data, volatile&＃61;True), Variable(target)output &＃61; model(data)# sum up batch loss# test_loss &＃43;&＃61; F.nll_loss(output, target, size_average&＃61;False).data[0]test_loss &＃43;&＃61; F.nll_loss(output, target, size_average&＃61;False).data.item()# get the index of the max log-probabilitypred &＃61; output.data.max(1, keepdim&＃61;True)[1]print(pred)correct &＃43;&＃61; pred.eq(target.data.view_as(pred)).cpu().sum()test_loss /&＃61; len(test_loader.dataset)print(&＃39;\nTest set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n&＃39;.format(test_loss, correct, len(test_loader.dataset),100. * correct / len(test_loader.dataset)))for epoch in range(1, 10):train(epoch)test()