上篇理解完官方的代码后,根据自己的理解,写了一份自己风格的CPU版本代码,并加上了详细的注释,方便像我一样的新手小白理解每块代码都在做些什么事情。
from __future__ import print_function
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torchvision import datasets, transforms
#parameters
batch_size = 32
lr = 0.01
momentum = 0.5
#defined a Net
class Net(nn.Module):
def __init__(self): #初始化函数
super(Net, self).__init__() #初始化
self.conv1 = nn.Sequential( #第一部分,一个卷积层,然后用relu函数激活,最后pooling
nn.Conv2d(1, 20, 5, 1), #in_channels = 1,out_channels = 20,kernel_size = 5,stride =1
nn.ReLU(),
nn.MaxPool2d(2) #kernel_size = 2
)
self.conv2 = nn.Sequential( #同上,in_channels = 20,out_channels =50
nn.Conv2d(20, 50, 5, 1),
nn.ReLU(),
nn.MaxPool2d(2)
)
self.fc1 = nn.Linear(4 * 4 * 50, 500) #fully connection,线性函数,
self.out = nn.Linear(500, 10) #ouptut层,最后输出结果有10种可能
def forward(self, x): #foward函数,往output层传播
x = self.conv1(x) #通过两层复合卷积层
x = self.conv2(x)
x = x.view(-1, 4 * 4 * 50) #用view()函数把卷积结果展平
x = self.fc1(x) #通过fc1线性函数还原
x = self.out(x) #通过一个线性函数输出
return F.log_softmax(x, dim=1) #注意这里return的值要是数值类型,这里用log_softmax优化,模仿官方的做法
def train(model, train_loader, optimizer, epoch): #训练函数
model.train() #进入训练状态
for batch_idx, (data, target) in enumerate(train_loader): #对train_loader进行批数据训练
optimizer.zero_grad() #梯度清零
output = model(data) #调用网络
loss = F.nll_loss(output, target) #计算损失函数,把output的训练结果和target进行比较
loss.backward() #计算梯度
optimizer.step() #更新参数
if batch_idx % 50 == 0: #每训练50批打印,epoch,当前进度,当前损失值
print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
epoch, batch_idx * len(data), len(train_loader.dataset),
100. * batch_idx / len(train_loader), loss.item()))
def test(model, test_loader): #测试函数
model.eval() #进入test状态
test_loss = 0 #初始化
correct = 0
with torch.no_grad(): #test不需要梯度,猜想这个函数是想节约内存
for data, target in test_loader:
output = model(data)
test_loss += F.nll_loss(output, target, reduction='sum').item() # sum up batch loss
pred = output.argmax(dim=1, keepdim=True) # get the index of the max log-probability
correct += pred.eq(target.view_as(pred)).sum().item() #比较pred和target的值,equal为1
test_loss /= len(test_loader.dataset)
print('\nTest set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format(
test_loss, correct, len(test_loader.dataset),
100. * correct / len(test_loader.dataset)))
def main():
train_loader = torch.utils.data.DataLoader(
datasets.MNIST('../mnist', train=True, download=True, #下载并创建数据集
transform=transforms.Compose([ #对数据集进行处理
transforms.ToTensor(), #转化为tensor类型
transforms.Normalize((0.1307,), (0.3081,)) #按照mean和std处理数据
])),
batch_size=batch_size, shuffle=True)
test_loader = torch.utils.data.DataLoader(
datasets.MNIST('../mnist', train=False, transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307,), (0.3081,))
])),
batch_size=batch_size, shuffle=True)
model = Net() #建模
optimizer = optim.SGD(model.parameters(), lr=lr, momentum=momentum) #用SGD优化模型
for epoch in range(1): #遍历全部数据集一次
train(model, train_loader, optimizer, epoch) #训练
test(model,test_loader) #测试
main()
下面是我的测试结果,比官方的略微好了一点点,我的batch_size定义的比较小,是32,官方64。