一、MNIST 数据集分类

1.加载数据

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torchvision import datasets, transforms
import matplotlib.pyplot as plt
import numpy

# 一个函数,用来计算模型中有多少参数
def get_n_params(model):
    np=0
    for p in list(model.parameters()):
        np += p.nelement()
    return np

# 使用GPU训练,可以在菜单 "代码执行工具" -> "更改运行时类型" 里进行设置
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
input_size  = 28*28   # MNIST上的图像尺寸是 28x28
output_size = 10      # 类别为 0 到 9 的数字,因此为十类

train_loader = torch.utils.data.DataLoader(
    datasets.MNIST('./data', train=True, download=True,
        transform=transforms.Compose(
            [transforms.ToTensor(),
             transforms.Normalize((0.1307,), (0.3081,))])),
    batch_size=64, shuffle=True)

test_loader = torch.utils.data.DataLoader(
    datasets.MNIST('./data', train=False, transform=transforms.Compose([
             transforms.ToTensor(),
             transforms.Normalize((0.1307,), (0.3081,))])),
    batch_size=1000, shuffle=True)

显示数据集中的部分图像

plt.figure(figsize=(8, 5))
for i in range(20):
    plt.subplot(4, 5, i + 1)
    image, _ = train_loader.dataset.__getitem__(i)
    plt.imshow(image.squeeze().numpy(),'gray')
    plt.axis('off');

 

 2.创建网络

class FC2Layer(nn.Module):
    def __init__(self, input_size, n_hidden, output_size):
        # nn.Module子类的函数必须在构造函数中执行父类的构造函数
        # 下式等价于nn.Module.__init__(self)        
        super(FC2Layer, self).__init__()
        self.input_size = input_size
        # 这里直接用 Sequential 就定义了网络,注意要和下面 CNN 的代码区分开
        self.network = nn.Sequential(
            nn.Linear(input_size, n_hidden), 
            nn.ReLU(), 
            nn.Linear(n_hidden, n_hidden), 
            nn.ReLU(), 
            nn.Linear(n_hidden, output_size), 
            nn.LogSoftmax(dim=1)
        )
    def forward(self, x):
        # view一般出现在model类的forward函数中,用于改变输入或输出的形状
        # x.view(-1, self.input_size) 的意思是多维的数据展成二维
        # 代码指定二维数据的列数为 input_size=784,行数 -1 表示我们不想算,电脑会自己计算对应的数字
        # 在 DataLoader 部分,我们可以看到 batch_size 是64,所以得到 x 的行数是64
        # 大家可以加一行代码:print(x.cpu().numpy().shape)
        # 训练过程中,就会看到 (64, 784) 的输出,和我们的预期是一致的

        # forward 函数的作用是,指定网络的运行过程,这个全连接网络可能看不啥意义,
        # 下面的CNN网络可以看出 forward 的作用。
        x = x.view(-1, self.input_size)
        return self.network(x)
    


class CNN(nn.Module):
    def __init__(self, input_size, n_feature, output_size):
        # 执行父类的构造函数,所有的网络都要这么写
        super(CNN, self).__init__()
        # 下面是网络里典型结构的一些定义,一般就是卷积和全连接
        # 池化、ReLU一类的不用在这里定义
        self.n_feature = n_feature
        self.conv1 = nn.Conv2d(in_channels=1, out_channels=n_feature, kernel_size=5)
        self.conv2 = nn.Conv2d(n_feature, n_feature, kernel_size=5)
        self.fc1 = nn.Linear(n_feature*4*4, 50)
        self.fc2 = nn.Linear(50, 10)    
    
    # 下面的 forward 函数,定义了网络的结构,按照一定顺序,把上面构建的一些结构组织起来
    # 意思就是,conv1, conv2 等等的,可以多次重用
    def forward(self, x, verbose=False):
        x = self.conv1(x)
        x = F.relu(x)
        x = F.max_pool2d(x, kernel_size=2)
        x = self.conv2(x)
        x = F.relu(x)
        x = F.max_pool2d(x, kernel_size=2)
        x = x.view(-1, self.n_feature*4*4)
        x = self.fc1(x)
        x = F.relu(x)
        x = self.fc2(x)
        x = F.log_softmax(x, dim=1)
        return x

定义训练和测试函数

# 训练函数
def train(model):
    model.train()
    # 主里从train_loader里,64个样本一个batch为单位提取样本进行训练
    for batch_idx, (data, target) in enumerate(train_loader):
        # 把数据送到GPU中
        data, target = data.to(device), target.to(device)

        optimizer.zero_grad()
        output = model(data)
        loss = F.nll_loss(output, target)
        loss.backward()
        optimizer.step()
        if batch_idx % 100 == 0:
            print('Train: [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
                batch_idx * len(data), len(train_loader.dataset),
                100. * batch_idx / len(train_loader), loss.item()))


def test(model):
    model.eval()
    test_loss = 0
    correct = 0
    for data, target in test_loader:
        # 把数据送到GPU中
        data, target = data.to(device), target.to(device)
        # 把数据送入模型,得到预测结果
        output = model(data)
        # 计算本次batch的损失,并加到 test_loss 中
        test_loss += F.nll_loss(output, target, reduction='sum').item()
        # get the index of the max log-probability,最后一层输出10个数,
        # 值最大的那个即对应着分类结果,然后把分类结果保存在 pred 里
        pred = output.data.max(1, keepdim=True)[1]
        # 将 pred 与 target 相比,得到正确预测结果的数量,并加到 correct 中
        # 这里需要注意一下 view_as ,意思是把 target 变成维度和 pred 一样的意思                                                
        correct += pred.eq(target.data.view_as(pred)).cpu().sum().item()

    test_loss /= len(test_loader.dataset)
    accuracy = 100. * correct / len(test_loader.dataset)
    print('\nTest set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format(
        test_loss, correct, len(test_loader.dataset),
        accuracy))

3. 在小型全连接网络上训练(Fully-connected network)

n_hidden = 8 # number of hidden units

model_fnn = FC2Layer(input_size, n_hidden, output_size)
model_fnn.to(device)
optimizer = optim.SGD(model_fnn.parameters(), lr=0.01, momentum=0.5)
print('Number of parameters: {}'.format(get_n_params(model_fnn)))

train(model_fnn)
test(model_fnn)

 

 

 3. 在卷积神经网络上训练

# Training settings 
n_features = 6 # number of feature maps

model_cnn = CNN(input_size, n_features, output_size)
model_cnn.to(device)
optimizer = optim.SGD(model_cnn.parameters(), lr=0.01, momentum=0.5)
print('Number of parameters: {}'.format(get_n_params(model_cnn)))

train(model_cnn)
test(model_cnn)

通过上面的测试结果,可以发现,
含有相同参数的 CNN 效果要明显优于 简单的全连接网络,是因为 CNN 能够更好的挖掘图像中的信息,主要通过两个手段:

卷积:Locality and stationarity in images

池化:Builds in some translation invariance

4. 打乱像素顺序再次在两个网络上训练与测试

考虑到CNN在卷积与池化上的优良特性,如果我们把图像中的像素打乱顺序,这样 卷积 和 池化 就难以发挥作用了,为了验证这个想法,我们把图像中的像素打乱顺序再试试。

#这里解释一下 torch.randperm 函数,给定参数n,返回一个从0到n-1的随机整数排列
perm = torch.randperm(784)
plt.figure(figsize=(8, 4))
for i in range(10):
    image, _ = train_loader.dataset.__getitem__(i)
    # permute pixels
    image_perm = image.view(-1, 28*28).clone()
    image_perm = image_perm[:, perm]
    image_perm = image_perm.view(-1, 1, 28, 28)
    plt.subplot(4, 5, i + 1)
    plt.imshow(image.squeeze().numpy(), 'gray')
    plt.axis('off')
    plt.subplot(4, 5, i + 11)
    plt.imshow(image_perm.squeeze().numpy(), 'gray')
    plt.axis('off')

重新定义训练与测试函数,我们写了两个函数 train_perm 和 test_perm,分别对应着加入像素打乱顺序的训练函数与测试函数。

与之前的训练与测试函数基本上完全相同,只是对 data 加入了打乱顺序操作。

# 对每个 batch 里的数据,打乱像素顺序的函数
def perm_pixel(data, perm):
    # 转化为二维矩阵
    data_new = data.view(-1, 28*28)
    # 打乱像素顺序
    data_new = data_new[:, perm]
    # 恢复为原来4维的 tensor
    data_new = data_new.view(-1, 1, 28, 28)
    return data_new

# 训练函数
def train_perm(model, perm):
    model.train()
    for batch_idx, (data, target) in enumerate(train_loader):
        data, target = data.to(device), target.to(device)
        # 像素打乱顺序
        data = perm_pixel(data, perm)

        optimizer.zero_grad()
        output = model(data)
        loss = F.nll_loss(output, target)
        loss.backward()
        optimizer.step()
        if batch_idx % 100 == 0:
            print('Train: [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
                batch_idx * len(data), len(train_loader.dataset),
                100. * batch_idx / len(train_loader), loss.item()))

# 测试函数
def test_perm(model, perm):
    model.eval()
    test_loss = 0
    correct = 0
    for data, target in test_loader:
        data, target = data.to(device), target.to(device)

        # 像素打乱顺序
        data = perm_pixel(data, perm)

        output = model(data)
        test_loss += F.nll_loss(output, target, reduction='sum').item()
        pred = output.data.max(1, keepdim=True)[1]                                            
        correct += pred.eq(target.data.view_as(pred)).cpu().sum().item()

    test_loss /= len(test_loader.dataset)
    accuracy = 100. * correct / len(test_loader.dataset)
    print('\nTest set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format(
        test_loss, correct, len(test_loader.dataset),
        accuracy))

在全连接网络上训练与测试:

perm = torch.randperm(784)
n_hidden = 8 # number of hidden units

model_fnn = FC2Layer(input_size, n_hidden, output_size)
model_fnn.to(device)
optimizer = optim.SGD(model_fnn.parameters(), lr=0.01, momentum=0.5)
print('Number of parameters: {}'.format(get_n_params(model_fnn)))

train_perm(model_fnn, perm)
test_perm(model_fnn, perm)

 

 

 在卷积神经网络上训练与测试:

perm = torch.randperm(784)
n_features = 6 # number of feature maps

model_cnn = CNN(input_size, n_features, output_size)
model_cnn.to(device)
optimizer = optim.SGD(model_cnn.parameters(), lr=0.01, momentum=0.5)
print('Number of parameters: {}'.format(get_n_params(model_cnn)))

train_perm(model_cnn, perm)
test_perm(model_cnn, perm)

从打乱像素顺序的实验结果来看,全连接网络的性能基本上没有发生变化,但是 卷积神经网络的性能明显下降。

这是因为对于卷积神经网络,会利用像素的局部关系,但是打乱顺序以后,这些像素间的关系将无法得到利用。

二、CIFAR10 数据集分类

import torch
import torchvision
import torchvision.transforms as transforms
import matplotlib.pyplot as plt
import numpy as np
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim

# 使用GPU训练,可以在菜单 "代码执行工具" -> "更改运行时类型" 里进行设置
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

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

# 注意下面代码中:训练的 shuffle 是 True,测试的 shuffle 是 false
# 训练时可以打乱顺序增加多样性,测试是没有必要
trainset = torchvision.datasets.CIFAR10(root='./data', train=True,download=True, transform=transform)

trainloader = torch.utils.data.DataLoader(trainset, batch_size=64,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=8,shuffle=False, num_workers=2)

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

定义网络,损失函数和优化器:

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)))
        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

# 网络放到GPU上
net = Net().to(device)
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(net.parameters(), lr=0.001)

训练网络:

for epoch in range(10):  # 重复多轮训练
    for i, (inputs, labels) in enumerate(trainloader):
        inputs = inputs.to(device)
        labels = labels.to(device)
        # 优化器梯度归零
        optimizer.zero_grad()
        # 正向传播 + 反向传播 + 优化 
        outputs = net(inputs)
        loss = criterion(outputs, labels)
        loss.backward()
        optimizer.step()
        # 输出统计信息
        if i % 100 == 0:   
            print('Epoch: %d Minibatch: %5d loss: %.3f' %(epoch + 1, i + 1, loss.item()))

print('Finished Training')

现在我们从测试集中取出8张图片:

# 得到一组图像
images, labels = iter(testloader).next()
# 展示图像
imshow(torchvision.utils.make_grid(images))
# 展示图像的标签
for j in range(8):
    print(classes[labels[j]])

 

 把图片输入模型,看看CNN把这些图片识别成什么:

outputs = net(images.to(device))
_, predicted = torch.max(outputs, 1)

# 展示预测的结果
for j in range(8):
    print(classes[predicted[j]])

 

 可以看到,有几个都识别错了~ 让我们看看网络在整个数据集上的表现:

correct = 0
total = 0

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

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

 

 准确率还可以,通过改进网络结构,性能还可以进一步提升。在 Kaggle 的LeaderBoard上,准确率高的达到95%以上。

三、使用 VGG16 对 CIFAR10 分类

1. 定义 dataloader

import torch
import torchvision
import torchvision.transforms as transforms
import matplotlib.pyplot as plt
import numpy as np
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim

# 使用GPU训练,可以在菜单 "代码执行工具" -> "更改运行时类型" 里进行设置
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

transform_train = transforms.Compose([
    transforms.RandomCrop(32, padding=4),
    transforms.RandomHorizontalFlip(),
    transforms.ToTensor(),
    transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010))])

transform_test = transforms.Compose([
    transforms.ToTensor(),
    transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010))])

trainset = torchvision.datasets.CIFAR10(root='./data', train=True,  download=True, transform=transform_train)
testset  = torchvision.datasets.CIFAR10(root='./data', train=False, download=True, transform=transform_test)

trainloader = torch.utils.data.DataLoader(trainset, batch_size=128, shuffle=True, num_workers=2)
testloader = torch.utils.data.DataLoader(testset, batch_size=128, shuffle=False, num_workers=2)

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

2.VGG 网络定义

class VGG(nn.Module):
    def __init__(self):
        super(VGG, self).__init__()
        self.cfg = [64, 'M', 128, 'M', 256, 256, 'M', 512, 512, 'M', 512, 512, 'M']
        self.features = self._make_layers(self.cfg) #这里如果不加self,编译器可能识别不了,后面无法初始化对象
        self.classifier = nn.Linear(512, 10)    #channel是512,512维特征(1*1)经过全连接层变为10类

    def forward(self, x):
        out = self.features(x)
        out = out.view(out.size(0), -1)
        out = self.classifier(out)
        return out

    def _make_layers(self, cfg):
        layers = []
        in_channels = 3
        for x in cfg:
            if x == 'M':
                #池化层输出变为(N/2)*(N/2),共经历5次池化变为最后变为N/32,CIFAR10数据N=32,即1*1
                layers += [nn.MaxPool2d(kernel_size=2, stride=2)]
            else:
                layers += [nn.Conv2d(in_channels, x, kernel_size=3, padding=1), #卷积层输出特征图仍为N*N (N+2*1-3)/1+1=N 
                           nn.BatchNorm2d(x),
                           nn.ReLU(inplace=True)]
                in_channels = x
        layers += [nn.AvgPool2d(kernel_size=1, stride=1)]
        return nn.Sequential(*layers)

初始化网络,根据实际需要,修改分类层。因为 tiny-imagenet 是对200类图像分类,这里把输出修改为200。

# 网络放到GPU上
net = VGG().to(device)
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(net.parameters(), lr=0.001)

3.网络训练

训练的代码和以前是完全一样的:

for epoch in range(10):  # 重复多轮训练
    for i, (inputs, labels) in enumerate(trainloader):
        inputs = inputs.to(device)
        labels = labels.to(device)
        # 优化器梯度归零
        optimizer.zero_grad()
        # 正向传播 + 反向传播 + 优化 
        outputs = net(inputs)
        loss = criterion(outputs, labels)
        loss.backward()
        optimizer.step()
        # 输出统计信息
        if i % 100 == 0:   
            print('Epoch: %d Minibratch: %5d loss: %.3f' %(epoch + 1, i + 1, loss.item()))

print('Finished Training')

4. 测试验证准确率:

correct = 0
total = 0

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

print('Accuracy of the network on the 10000 test images: %.2f %%' % (
    100 * correct / total))

可以看到,使用一个简化版的 VGG 网络,就能够显著地将准确率由 64%,提升到 84.92%

posted on 2021-10-16 16:24  东皇御前  阅读(107)  评论(0编辑  收藏  举报