1、PyTorch实例入门(1):图像分类,代码例子。
参考文章:PyTorch实例入门(1):图像分类 - 知乎,我稍微整理了一下,方便初学者理解。代码可以直接跑。
import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvision
import torchvision.transforms as transforms
# optim中定义了各种各样的优化方法,包括SGD
import torch.optim as optim
import datetime
class LeNet(nn.Module):
# 一般在__init__中定义网络需要的操作算子,比如卷积、全连接算子等等
def __init__(self):
super(LeNet, self).__init__()
# Conv2d的第一个参数是输入的channel数量,第二个是输出的channel数量,第三个是kernel size
self.conv1 = nn.Conv2d(3, 6, 5)
self.conv2 = nn.Conv2d(6, 16, 5)
# 由于上一层有16个channel输出,每个feature map大小为5*5,所以全连接层的输入是16*5*5
self.fc1 = nn.Linear(16*5*5, 120)
self.fc2 = nn.Linear(120, 84)
# 最终有10类,所以最后一个全连接层输出数量是10
self.fc3 = nn.Linear(84, 10)
self.pool = nn.MaxPool2d(2, 2)
# forward这个函数定义了前向传播的运算,只需要像写普通的python算数运算那样就可以了
def forward(self, x):
# 输入x:torch.Size([32, 3, 32, 32])
# self.conv1(x), Conv2d(3, 6, 5): torch.Size([32, 6, 28, 28])
# F.relu输出x: torch.Size([32, 6, 28, 28])
x = F.relu(self.conv1(x))
# self.pool(x):torch.Size([32, 6, 14, 14])
x = self.pool(x)
# self.conv2(x):torch.Size([32, 16, 10, 10])
# F.relu输出x:torch.Size([32, 16, 10, 10])
x = F.relu(self.conv2(x))
# self.pool(x):torch.Size([32, 16, 5, 5])
x = self.pool(x)
# 下面这步把二维特征图变为一维,这样全连接层才能处理
# x.view(-1, 16*5*5):torch.Size([32, 400])
x = x.view(-1, 16*5*5)
# self.fc1(x):torch.Size([32, 120])
# F.relu:torch.Size([32, 120])
x = F.relu(self.fc1(x))
# self.fc1(x):torch.Size([32, 84])
# F.relu:torch.Size([32, 84])
x = F.relu(self.fc2(x))
# self.fc3:torch.Size([32, 10])
x = self.fc3(x)
return x
# cifar-10官方提供的数据集是用numpy array存储的
# 下面这个transform会把numpy array变成torch tensor,然后把rgb值归一到[0, 1]这个区间
transform = transforms.Compose(
[transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])
# 在构建数据集的时候指定transform,就会应用我们定义好的transform
# root是存储数据的文件夹,download=True指定如果数据不存在先下载数据
cifar_train = torchvision.datasets.CIFAR10(root='./data', train=True,
download=True, transform=transform)
cifar_test = torchvision.datasets.CIFAR10(root='./data', train=False,
transform=transform)
print(cifar_train)
print(cifar_test)
trainloader = torch.utils.data.DataLoader(cifar_train, batch_size=32, shuffle=True)
testloader = torch.utils.data.DataLoader(cifar_test, batch_size=32, shuffle=True)
# 由于需要用到GPU,所以先获取device,然后再把网络的参数复制到GPU上
# 如果你没有GPU,那么可以忽略device相关的代码
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
net = LeNet().to(device)
# CrossEntropyLoss就是我们需要的损失函数
criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(net.parameters(), lr=0.001, momentum=0.9)
if __name__ == '__main__':
print("Start Training...")
start_time = datetime.datetime.now()
print("start_time:", start_time)
for epoch in range(30):
# 我们用一个变量来记录每100个batch的平均loss
loss100 = 0.0
# 我们的dataloader派上了用场
for i, data in enumerate(trainloader):
inputs, labels = data
inputs, labels = inputs.to(device), labels.to(device) # 注意需要复制到GPU
optimizer.zero_grad()
outputs = net(inputs)
loss = criterion(outputs, labels)
loss.backward()
optimizer.step()
loss100 += loss.item()
if i % 100 == 99:
print('[Epoch %d, Batch %5d] loss: %.3f' %
(epoch + 1, i + 1, loss100 / 100))
loss100 = 0.0
end_time = datetime.datetime.now()
print("end_time:", end_time)
print("Done Training!")
delta = end_time - start_time
# 获取timedelta对象包含的总秒数
print("total_seconds: ", delta.total_seconds())