神经网络模型训练过程

通过学习前面几个例子，我们可以总结一下神经网络的典型训练过程如下：

第一步：定义具有一些可学习参数（或权重）的神经网络
第二步：迭代输入数据集
第三步：通过网络处理输入
第四步：计算损失(loss)
第五步：反向传播网络的参数
第六步：更新网络的参数，通常使用一个简单的更新规则：weight = weight - learning_rate * gradient

使用PyTorch训练分类器

本文通过官网上一个更为具体的例子来更好的学习使用PyTorch训练分类模型的具体过程。

数据预处理

一般来说，当我们在处理图像，文本，音频或视频数据时，可以使用标准的Python包将数据加载到一个numpy数组中。然后把这个数组转换成 torch.* Tensor。

1.在处理图片数据方面，可以选择使用Pillow，OpenCV等。
2.在处理音频数据方面，可以选择使用scipy，librosa等。
3.在处理文本数据方面，可以选择使用NLTK，SpaCy等。

在PyTorch这个框架中，对于计算机视觉，已经创建了一个名为torchvision的包，包含了常见数据集的数据加载器，如Imagenet，CIFAR10，MNIST等，以及用于图像的数据转换器，即torchvision.datasets和torch.utils.data.DataLoader。

在本文中，我们将使用CIFAR10数据集。 它有 ‘airplane’, ‘automobile’, ‘bird’, ‘cat’, ‘deer’, ‘dog’, ‘frog’, ‘horse’, ‘ship’, ‘truck’类。 CIFAR-10中的图像大小为3x32x32，即32x32像素的3通道彩色图像。

训练图像分类器

根据前面提到的，我们将按顺序执行以下步骤：

第一步：使用torchvision加载和规范CIFAR10训练和测试数据集
第二步：定义模型：选用卷积神经网络
第三步：定义损失函数
第四步：在训练数据上训练网络
第五步：在测试数据上测试网络

加载CIFAR10数据集

使用torchvision可以很方面的加载CIFAR10数据集。torchvision数据集的输出是范围[0，1]的PILImage图像。 我们将它们转换为归一化范围的张量[-1,1]。下面代码可视化一个batch的图片。

import torch
import torchvision
import torchvision.transforms as transforms

transform = transforms.Compose(
    [transforms.ToTensor(),
     transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])

trainset = torchvision.datasets.CIFAR10(root='./data', train=True,
                                        download=True, transform=transform)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=4,
                                          shuffle=True, num_workers=2)

testset = torchvision.datasets.CIFAR10(root='./data', train=False,
                                       download=True, transform=transform)
testloader = torch.utils.data.DataLoader(testset, batch_size=4,
                                         shuffle=False, num_workers=2)

classes = ('plane', 'car', 'bird', 'cat',
           'deer', 'dog', 'frog', 'horse', 'ship', 'truck')

import matplotlib.pyplot as plt
import numpy as np

# functions to show an image
def imgshow(img):
    img = img / 2 + 0.5     # unnormalize
    npimg = img.numpy()
    plt.imshow(np.transpose(npimg, (1, 2, 0)))
    plt.show()

if __name__ ==  '__main__':
    # get some random training images
    dataiter = iter(trainloader)
    images, labels = dataiter.next()
    # show images
  imgshow(torchvision.utils.make_grid(images))
    # print labels
    print(' '.join('%5s' % classes[labels[j]] for j in range(4)))

运行结果：

定义卷积神经网络

此时为3通道图像（而不是单通道图像）:

from torch.autograd import  Variable
import torch.nn as nn
import torch.nn.functional as F

# 3*32*32
class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = nn.Conv2d(3, 6, 5)
        self.pool = nn.MaxPool2d(2, 2)
        self.conv2 = nn.Conv2d(6, 16, 5)
        self.fc1 = nn.Linear(16 * 5 * 5, 120)
        self.fc2 = nn.Linear(120, 84)
        self.fc3 = nn.Linear(84, 10)

    def forward(self, x):
        x = self.pool(F.relu(self.conv1(x)))
        x = self.pool(F.relu(self.conv2(x)))
        # fully connect
        x = x.view(-1, 16 * 5 * 5)
        x = F.relu(self.fc1(x))
        x = F.relu(self.fc2(x))
        x = self.fc3(x)
        return x

定义损失函数和优化器

# Define loss (Cross-Entropy)
import torch.optim as optim

criterion = nn.CrossEntropyLoss()
# SGD with momentum
optimizer = optim.SGD(net.parameters(), lr=0.001, momentum=0.9)

训练网络

循环遍历数据迭代器，并将数据输入到网络并进行优化。

# Train the network
    for epoch in range(5):
        running_loss = 0.0
        for i, data in enumerate(trainloader, 0):
            # get the inputs
            inputs, labels = data
            # warp them in Variable
            inputs, labels = Variable(inputs), Variable(labels)
            # zero the parameter gradients
            optimizer.zero_grad()
            # forward
            outputs = net(inputs)
            # loss
            loss = criterion(outputs, labels)
            # backward
            loss.backward()
            # update weights
            optimizer.step()
            # print statistics
            running_loss += loss.data[0]
            if i % 2000 == 1999:  # print every 2000 mini-batches
                print('[%d, %5d] loss: %.3f' %
                      (epoch + 1, i + 1, running_loss / 2000))
                running_loss = 0.0

    print("Finished Training")

在测试数据上测试训练好的模型

correct = 0
    total = 0
    for data in testloader:
        images, labels = data
        outputs = net(Variable(images))
        _, predicted = torch.max(outputs.data, 1)
        total += labels.size(0)
        correct += (predicted == labels).sum()

完整代码

#author: yuquanle
#date: 2018.2.5
#Classifier use PyTorch (CIFAR10 dataset)

import torch
import torchvision
import torchvision.transforms as transforms

transform = transforms.Compose(
    [transforms.ToTensor(),
     transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])

trainset = torchvision.datasets.CIFAR10(root='./data', train=True,
                                        download=True, transform=transform)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=4,
                                          shuffle=True, num_workers=1)

testset = torchvision.datasets.CIFAR10(root='./data', train=False,
                                       download=True, transform=transform)
testloader = torch.utils.data.DataLoader(testset, batch_size=4,
                                         shuffle=False, num_workers=1)

classes = ('plane', 'car', 'bird', 'cat',
           'deer', 'dog', 'frog', 'horse', 'ship', 'truck')

# 3*32*32
from torch.autograd import  Variable
import torch.nn as nn
import torch.nn.functional as F

class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = nn.Conv2d(3, 6, 5)
        self.pool = nn.MaxPool2d(2, 2)
        self.conv2 = nn.Conv2d(6, 16, 5)
        self.fc1 = nn.Linear(16 * 5 * 5, 120)
        self.fc2 = nn.Linear(120, 84)
        self.fc3 = nn.Linear(84, 10)

    def forward(self, x):
        x = self.pool(F.relu(self.conv1(x)))
        x = self.pool(F.relu(self.conv2(x)))
        # fully connect
        x = x.view(-1, 16 * 5 * 5)
        x = F.relu(self.fc1(x))
        x = F.relu(self.fc2(x))
        x = self.fc3(x)
        return x

# our model
net = Net()

# Define loss (Cross-Entropy)
import torch.optim as optim

criterion = nn.CrossEntropyLoss()
# SGD with momentum
optimizer = optim.SGD(net.parameters(), lr=0.001, momentum=0.9)

if __name__ ==  '__main__':
    # Train the network
    for epoch in range(5):
        running_loss = 0.0
        for i, data in enumerate(trainloader, 0):
            # get the inputs
            inputs, labels = data

            # warp them in Variable
            inputs, labels = Variable(inputs), Variable(labels)

            # zero the parameter gradients
            optimizer.zero_grad()

            # forward
            outputs = net(inputs)
            # loss
            loss = criterion(outputs, labels)
            # backward
            loss.backward()
            # update weights
            optimizer.step()

            # print statistics
            running_loss += loss.data[0]
            if i % 2000 == 1999:  # print every 2000 mini-batches
                print('[%d, %5d] loss: %.3f' %
                      (epoch + 1, i + 1, running_loss / 2000))
                running_loss = 0.0
    print("Finished Training")

    print("Beginning Testing")
    correct = 0
    total = 0
    for data in testloader:
        images, labels = data
        outputs = net(Variable(images))
        _, predicted = torch.max(outputs.data, 1)
        total += labels.size(0)
        correct += (predicted == labels).sum()

    print('Accuracy of the network on the 10000 test images: %d %%' % (
        100 * correct / total))

运行结果：

Files already downloaded and verified
Files already downloaded and verified
Files already downloaded and verified
Files already downloaded and verified
[1, 2000] loss: 2.190
[1, 4000] loss: 1.860
[1, 6000] loss: 1.687
[1, 8000] loss: 1.591
[1, 10000] loss: 1.524
[1, 12000] loss: 1.471
Files already downloaded and verified
Files already downloaded and verified
[2, 2000] loss: 1.385
[2, 4000] loss: 1.367
[2, 6000] loss: 1.343
[2, 8000] loss: 1.311
[2, 10000] loss: 1.285
[2, 12000] loss: 1.281
Files already downloaded and verified
Files already downloaded and verified
[3, 2000] loss: 1.198
[3, 4000] loss: 1.194
[3, 6000] loss: 1.193
[3, 8000] loss: 1.172
[3, 10000] loss: 1.168
[3, 12000] loss: 1.142
Files already downloaded and verified
Files already downloaded and verified
[4, 2000] loss: 1.063
[4, 4000] loss: 1.096
[4, 6000] loss: 1.070
[4, 8000] loss: 1.074
[4, 10000] loss: 1.086
[4, 12000] loss: 1.066
Files already downloaded and verified
Files already downloaded and verified
[5, 2000] loss: 0.988
[5, 4000] loss: 0.996
[5, 6000] loss: 1.010
[5, 8000] loss: 0.999
[5, 10000] loss: 1.012
[5, 12000] loss: 1.016
Finished Training
Beginning Testing
Files already downloaded and verified
Files already downloaded and verified
Accuracy of the network on the 10000 test images: 62 %

Process finished with exit code 0

参考：http://pytorch.org/tutorials/beginner/blitz/cifar10_tutorial.html