这是pytorch官方的一个例子
官方教程地址：http://pytorch.org/tutorials/beginner/blitz/cifar10_tutorial.html#sphx-glr-beginner-blitz-cifar10-tutorial-py
代码如下
1 # coding=utf-8 2 import torch.nn as nn 3 import torch.nn.functional as F 4 from torch.autograd import Variable 5 import torch 6 import torchvision 7 import torchvision.transforms as transforms 8 import torch.optim as optim 9 10 # The output of torchvision datasets are PILImage images of range [0, 1]. 11 # We transform them to Tensors of normalized range [-1, 1] 12 transform = transforms.Compose([transforms.ToTensor(), 13 transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)), 14 ]) 15 16 # 训练集，将相对目录./data下的cifar-10-batches-py文件夹中的全部数据（50000张图片作为训练数据）加载到内存中，若download为True时，会自动从网上下载数据并解压 17 trainset = torchvision.datasets.CIFAR10(root=r'E:\Face Recognition\cifar-10-python', train=True, download=False, transform=transform) 18 19 # 将训练集的50000张图片划分成12500份，每份4张图，用于mini-batch输入。shffule=True在表示不同批次的数据遍历时，打乱顺序。num_workers=2表示使用两个子进程来加载数据 20 trainloader = torch.utils.data.DataLoader(trainset, batch_size=4, 21 shuffle=True) 22 23 # 测试集，将相对目录./data下的cifar-10-batches-py文件夹中的全部数据（10000张图片作为测试数据）加载到内存中，若download为True时，会自动从网上下载数据并解压 24 testset = torchvision.datasets.CIFAR10(root=r'E:\Face Recognition\cifar-10-python', train=False, download=False, transform=transform) 25 26 # 将测试集的10000张图片划分成2500份，每份4张图，用于mini-batch输入。 27 testloader = torch.utils.data.DataLoader(testset, batch_size=4, 28 shuffle=False) 29 classes = ('plane', 'car', 'bird', 'cat', 30 'deer', 'dog', 'frog', 'horse', 'ship', 'truck') 31 32 33 class Net(nn.Module): 34 def __init__(self): 35 super(Net, self).__init__() 36 self.conv1 = nn.Conv2d(3, 6, 5) # 定义conv1函数的是图像卷积函数：输入为图像（3个频道，即彩色图）,输出为6张特征图, 卷积核为5x5正方形 37 self.pool = nn.MaxPool2d(2, 2) 38 self.conv2 = nn.Conv2d(6, 16, 5) 39 self.fc1 = nn.Linear(16 * 5 * 5, 120) 40 self.fc2 = nn.Linear(120, 84) 41 self.fc3 = nn.Linear(84, 10) 42 43 def forward(self, x): 44 x = self.pool(F.relu(self.conv1(x))) 45 x = self.pool(F.relu(self.conv2(x))) 46 x = x.view(-1, 16 * 5 * 5) 47 x = F.relu(self.fc1(x)) 48 x = F.relu(self.fc2(x)) 49 x = self.fc3(x) 50 return x 51 52 53 net = Net() 54 55 criterion = nn.CrossEntropyLoss() # 叉熵损失函数 56 optimizer = optim.SGD(net.parameters(), lr=0.001, momentum=0.9) # 使用SGD（随机梯度下降）优化，学习率为0.001，动量为0.9 57 58 for epoch in range(10): # 遍历数据集两次 59 60 running_loss = 0.0 61 # enumerate(sequence, [start=0])，i序号，data是数据 62 for i, data in enumerate(trainloader, 0): 63 # get the inputs 64 inputs, labels = data # data的结构是：[4x3x32x32的张量,长度4的张量] 65 66 # wrap them in Variable 67 inputs, labels = Variable(inputs), Variable(labels) # 把input数据从tensor转为variable 68 69 # zero the parameter gradients 70 optimizer.zero_grad() # 将参数的grad值初始化为0 71 72 # forward + backward + optimize 73 outputs = net(inputs) 74 loss = criterion(outputs, labels) # 将output和labels使用叉熵计算损失 75 loss.backward() # 反向传播 76 optimizer.step() # 用SGD更新参数 77 78 # 每2000批数据打印一次平均loss值 79 running_loss += loss.data[0] # loss本身为Variable类型，所以要使用data获取其Tensor，因为其为标量，所以取0 80 if i % 2000 == 1999: # 每2000批打印一次 81 print('[%d, %5d] loss: %.3f' % (epoch + 1, i + 1, running_loss / 2000)) 82 running_loss = 0.0 83 84 print('Finished Training') 85 86 correct = 0 87 total = 0 88 for data in testloader: 89 images, labels = data 90 outputs = net(Variable(images)) 91 # print outputs.data 92 _, predicted = torch.max(outputs.data, 1) # outputs.data是一个4x10张量，将每一行的最大的那一列的值和序号各自组成一个一维张量返回，第一个是值的张量，第二个是序号的张量。 93 total += labels.size(0) 94 correct += (predicted == labels).sum() # 两个一维张量逐行对比，相同的行记为1，不同的行记为0，再利用sum(),求总和，得到相同的个数。 95 96 print('Accuracy of the network on the 10000 test images: %d %%' % (100 * correct / total))
1.由于windows平台的pytorch存在很多问题，例如多线程无法正常工作，所以DataLoader中的num_worker得去掉
2.代码以cifar-10数据测试集为例，但是训练的效果并不是很理想，loss函数数据如下，两次重复训练后的准确率为56%，10次重复训练后的准确率为61%,(个人表示原图片像素太差，至少一半，我都分不清是啥，真是为难了神经网络了)
[1, 2000] loss: 2.219 [1, 4000] loss: 1.869 [1, 6000] loss: 1.669 [1, 8000] loss: 1.581 [1, 10000] loss: 1.537 [1, 12000] loss: 1.488 [2, 2000] loss: 1.406 [2, 4000] loss: 1.385 [2, 6000] loss: 1.343 [2, 8000] loss: 1.318 [2, 10000] loss: 1.348 [2, 12000] loss: 1.305 [3, 2000] loss: 1.234 [3, 4000] loss: 1.206 [3, 6000] loss: 1.219 [3, 8000] loss: 1.213 [3, 10000] loss: 1.205 [3, 12000] loss: 1.199 [4, 2000] loss: 1.115 [4, 4000] loss: 1.127 [4, 6000] loss: 1.123 [4, 8000] loss: 1.118 [4, 10000] loss: 1.143 [4, 12000] loss: 1.106 [5, 2000] loss: 1.023 [5, 4000] loss: 1.022 [5, 6000] loss: 1.073 [5, 8000] loss: 1.076 [5, 10000] loss: 1.060 [5, 12000] loss: 1.048 [6, 2000] loss: 0.965 [6, 4000] loss: 0.985 [6, 6000] loss: 0.988 [6, 8000] loss: 1.008 [6, 10000] loss: 1.017 [6, 12000] loss: 0.999 [7, 2000] loss: 0.902 [7, 4000] loss: 0.925 [7, 6000] loss: 0.974 [7, 8000] loss: 0.955 [7, 10000] loss: 0.968 [7, 12000] loss: 0.979 [8, 2000] loss: 0.866 [8, 4000] loss: 0.893 [8, 6000] loss: 0.909 [8, 8000] loss: 0.932 [8, 10000] loss: 0.934 [8, 12000] loss: 0.937 [9, 2000] loss: 0.837 [9, 4000] loss: 0.858 [9, 6000] loss: 0.865 [9, 8000] loss: 0.873 [9, 10000] loss: 0.906 [9, 12000] loss: 0.907 [10, 2000] loss: 0.809 [10, 4000] loss: 0.810 [10, 6000] loss: 0.832 [10, 8000] loss: 0.865 [10, 10000] loss: 0.878 [10, 12000] loss: 0.877
