本文目的：展现如何利用PyTorch进行手写数字识别。html
1 导入相关库，定义一些参数
import torch
import torch.nn as nn
import torch.nn.functional as F
from torchvision import datasets, transforms
from torch.utils.data import DataLoader
#定义一些参数
BATCH_SIZE = 64
EPOCHS = 10
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
2 准备数据
使用Pytorch自带数据集。python
#图像预处理
transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307,), (0.3081,))
])
#训练集
train_set = datasets.MNIST('data', train=True, transform=transform, download=True)
train_loader = DataLoader(train_set,
batch_size=BATCH_SIZE,
shuffle=True)
#测试集
test_set = datasets.MNIST('data', train=False, transform=transform, download=True)
test_loader = DataLoader(test_set,
batch_size=BATCH_SIZE,
shuffle=True)
3 准备模型
#搭建模型
class ConvNet(nn.Module):
#图像输入是(batch,1,28,28)
def __init__(self):
super().__init__()
self.conv1 = nn.Conv2d(1, 10, (3,3)) #输入通道数为1，输出通道数为10，卷积核(3,3)
self.conv2 = nn.Conv2d(10, 32, (3,3))
self.fc1 = nn.Linear(12*12*32, 100)
self.fc2 = nn.Linear(100, 10)
def forward(self, x):
x = self.conv1(x) #(batch,10,26,26)
x = F.relu(x)
x = self.conv2(x) #(batch,32,24,24)
x = F.relu(x)
x = F.max_pool2d(x, (2,2)) #(batch,32,12,12)
x = x.view(x.size(0), -1) #flatten (batch,12*12*32)
x = self.fc1(x) #(batch,100)
x = F.relu(x)
x = self.fc2(x) #(batch,10)
out = F.log_softmax(x, dim=1) #softmax激活并取对数，数值上更稳定
return out
4 训练
#定义模型和优化器
model = ConvNet().to(DEVICE) #模型移至GPU
optimizer = torch.optim.Adam(model.parameters())
#定义训练函数
def train(model, device, train_loader, optimizer, epoch): #跑一个epoch
model.train() #开启训练模式，即启用BatchNormalization和Dropout等
for batch_idx, (data, target) in enumerate(train_loader): #每次产生一个batch
data, target = data.to(device), target.to(device) #产生的数据移至GPU
output = model(data)
loss = F.nll_loss(output, target) #CrossEntropyLoss = log_softmax + NLLLoss
optimizer.zero_grad() #全部梯度清零
loss.backward() #反向传播求全部参数梯度
optimizer.step() #沿负梯度方向走一步
if(batch_idx+1) % 234 == 0:
print('Train Epoch: {} [{}/{} ({:.1f}%)]\tLoss: {:.6f}'.format(
epoch, (batch_idx+1) * len(data), len(train_loader.dataset),
100. * (batch_idx+1) / len(train_loader), loss.item()))
#定义测试函数
def test(model, device, test_loader):
model.eval() #测试模式，不启用BatchNormalization和Dropout
test_loss = 0
correct = 0
with torch.no_grad(): #避免梯度跟踪
for data, target in test_loader:
data, target = data.to(device), target.to(device)
output = model(data)
test_loss += F.nll_loss(output, target, reduction='sum').item() #将一批损失相加
pred = output.max(1, keepdim=True)[1] #找到几率最大的下标
#上句效果等同于 pred = torch.argmax(output, dim=1, keepdim=True)
correct += pred.eq(target.view_as(pred)).sum().item()
test_loss /= len(test_loader.dataset)
#len(train_loader)为batch数，len(train_loader.dataset)为样本总数
print('\nTest set: Average loss: {:.4f}, Accuracy: {}/{} ({:.1f}%)\n'.format(
test_loss, correct, len(test_loader.dataset),
100. * correct / len(test_loader.dataset)))
#开始训练
for epoch in range(1, EPOCHS + 1):
train(model, DEVICE, train_loader, optimizer, epoch)
test(model, DEVICE, test_loader)
注意，torch.max()有两种用法：git
- 直接传入一个tensor，则返回全局最大值；
- torch.max(a, dim, [keepdim])返回一个tuple，前者为最大值结果，后者为indices（效果同argmax）；
- 详见 https://pytorch.org/docs/stable/torch.html?highlight=max#torch.max
- 此处 output.max() 与 torch.max()相似，只不过无需传入tensor
最终结果以下：github
5 小结
- 任务流程：准备数据，准备模型，训练
- 如何使用PyTorch自带数据集进行训练
- 自定义模型须要实现forward函数
- model.train()和model.eval()做用
- 最后一层x的交叉熵两种方式等价：CrossEntropyLoss = log_softmax + nll_loss
- torch.max()有两种用法，返回值不同
Reference函数