使用ResNet18预训练模型

由于笔记本性能太差，所以在服务器上运行的，显卡配置为4090。经大量实验判断，初始学习率为0.01最后效果较差，所以初始学习率应设为0.001。全部代码代码已上传到：https://github.com/wp-a/-CIFAR10-.git

库函数导入

1
2
3
4
5
6
7
8
9
import matplotlib.pyplot as plt
import torch
import torch.nn as nn
import torchvision
import torchvision.transforms as transforms
from sklearn.metrics import confusion_matrix, classification_report
from itertools import chain
import multiprocessing
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

数据集加载及增强操作

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
transform_train = transforms.Compose([
    transforms.Resize(224),	#图像从32x32放大到224x224是通过在原始像素之间插入新的像素点来实现的。插入的像素值是根据周围原始像素的值，通过不同的插值算法计算出来的，默认使用双线性插值。
    transforms.RandomHorizontalFlip(), #以 0.5 的概率水平随机翻转图像
    transforms.RandomRotation(10),  #将图像随机旋转 -10 到 10 度之间的角度
    transforms.ToTensor(),  #将图像从 PIL 图像格式转换为 PyTorch 张量
    transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2470, 0.2435, 0.2616)) 
    #使用为三个颜色通道中的每一个提供的平均和标准差来标准化图像的像素值(归一化)
])

transform_test = transforms.Compose([
    transforms.Resize(224),
    transforms.ToTensor(),
    transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2470, 0.2435, 0.2616))
])

# 加载CIFAR-10数据集
train_set = torchvision.datasets.CIFAR10("./data", download=True, transform=transform_train)
test_set = torchvision.datasets.CIFAR10("./data", download=True, train=False, transform=transform_test)

batch_size = 64  
train_loader = torch.utils.data.DataLoader(train_set, batch_size=batch_size, shuffle=True, num_workers=0)
test_loader = torch.utils.data.DataLoader(test_set, batch_size=batch_size, shuffle=False, num_workers=0)

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

#返回数字对应的类名
def output_label(label):
    output_mapping = {
        0: "plane", 1: "car", 2: "bird", 3: "cat", 4: "deer",
        5: "dog", 6: "frog", 7: "horse", 8: "ship", 9: "truck"
    }
    input = (label.item() if type(label) == torch.Tensor else label)
    return output_mapping[input]

加载预训练模型

1
2
3
4
5
6
7
def get_resnet18(pretrained=True):
    model = torchvision.models.resnet18(pretrained=pretrained)
    model.fc = nn.Linear(512, 10)  # 修改最后一层以适应CIFAR10的10个类别
    return model

model = get_resnet18()
model = model.to(device)

模型训练超参数设置

1
2
3
4
5
6
7
8
9
10
11
12
13
#定义损失函数，初始学习率，优化器
criterion = nn.CrossEntropyLoss()
learning_rate = 0.001
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)

#学习率调节器，前期需要较大步长来快速收敛，后期如还用较大步长可能会在极小值周围震荡，所以后期学习率逐步衰减。可以提高训练效率和性能。Adam负责在参数层面进行自适应的学习率调整，以加快收敛。学习率调度器负责在训练过程的宏观层面调整全局学习率，以提高模型的最终性能和泛化能力。二者不会产生冲突。
def exp_lr_scheduler(optimizer, epoch, init_lr=0.001, lr_decay_epoch=2):
    lr = init_lr * (0.1 ** (epoch // lr_decay_epoch))
    if epoch % lr_decay_epoch == 0:
        print(f'LR is set to {lr}')
    for param_group in optimizer.param_groups:
        param_group['lr'] = lr
    return optimizer

模型训练

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
num_epochs = 6
count = 0
loss_list = []
iteration_list = []
accuracy_list = []
predictions_list = []
labels_list = []

for epoch in range(num_epochs):
    model.train() 
    running_loss = 0.0

    optimizer = exp_lr_scheduler(optimizer, epoch, init_lr=learning_rate, lr_decay_epoch=2)

    for i, (images, labels) in enumerate(train_loader):
        images, labels = images.to(device), labels.to(device)

        outputs = model(images)
        loss = criterion(outputs, labels)

        optimizer.zero_grad()
        loss.backward()
        optimizer.step()

        running_loss += loss.item()
        count += 1

        if (i + 1) % 50 == 0:
            print(f'Epoch [{epoch + 1}/{num_epochs}], Step [{i + 1}/{len(train_loader)}], Loss: {loss.item():.4f}')

            model.eval()  
            total = 0
            correct = 0
            test_predictions = []
            test_labels = []

            with torch.no_grad():
                for images, labels in test_loader:
                    images, labels = images.to(device), labels.to(device)
                    test_labels.extend(labels.cpu().numpy())

                    outputs = model(images)
                    _, predicted = torch.max(outputs.data, 1)
                    test_predictions.extend(predicted.cpu().numpy())

                    total += labels.size(0)
                    correct += (predicted == labels).sum().item()

            accuracy = 100 * correct / total
            print(f'Accuracy on test set: {accuracy:.2f}%')

            loss_list.append(loss.item())
            iteration_list.append(count)
            accuracy_list.append(accuracy)

            predictions_list.append(test_predictions)
            labels_list.append(test_labels)

            model.train()

print('Finished Training')

实验结果保存

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
torch.save(model.state_dict(), 'resnet18_cifar10.pth')
print('Model saved to resnet18_cifar10.pth')

plt.figure(figsize=(10, 5))
plt.plot(iteration_list, loss_list)
plt.xlabel("No. of Iteration")
plt.ylabel("Loss")
plt.title("Iterations vs Loss")
plt.savefig('loss_curve.png')
plt.show()

plt.figure(figsize=(10, 5))
plt.plot(iteration_list, accuracy_list)
plt.xlabel("No. of Iteration")
plt.ylabel("Accuracy")
plt.title("Iterations vs Accuracy")
plt.savefig('accuracy_curve.png')
plt.show()

class_correct = [0. for _ in range(10)]
total_correct = [0. for _ in range(10)]

model.eval()
with torch.no_grad():
    for images, labels in test_loader:
        images, labels = images.to(device), labels.to(device)
        outputs = model(images)
        _, predicted = torch.max(outputs, 1)
        c = (predicted == labels).squeeze()

        for i in range(labels.size(0)):
            label = labels[i]
            class_correct[label] += c[i].item()
            total_correct[label] += 1

for i in range(10):
    print(f"Accuracy of {classes[i]}: {100 * class_correct[i] / total_correct[i]:.2f}%")

flat_predictions = list(chain.from_iterable(predictions_list))
flat_labels = list(chain.from_iterable(labels_list))

cm = confusion_matrix(flat_labels, flat_predictions)
print("Confusion Matrix:")
print(cm)

print("Classification report for ResNet on CIFAR-10:")
print(classification_report(flat_labels, flat_predictions, target_names=classes))

实验结果分析

lr分析

由实验结果可得lr=0.01时，效果较差，所以直接设置初始学习率为0.001可以更节省时间，提高效率。当学习率为1e-6时，发现准确率更新较小，所以准确率最小设置为1e-6。即学习率梯度为1e-3,1e-4,1e-5,1e-6，可以有较好的效果。即令epoch / lr_decay_epoch = 3或4 都可以，具体以epoch大小为准。

lr=0.01
lr=0.001
lr=1e6

batchsize分析

Batchsize：64 epoch：6 lr_decay_epoch：2 初始学习率为0.001 94.46%

Batchsize：128 epoch：12 lr_decay_epoch：3 初始学习率为0.001 95.23%

Batchsize：128 epoch：20 lr_decay_epoch：5 初始学习率为0.01 85.87%

和分析的一致，对比下面实验，控制变量，只有初始学习率改变，精度提升10%左右。

Batchsize：128 epoch：20 lr_decay_epoch：5 初始学习率为0.001 95.44%

Batchsize：64 epoch：20 lr_decay_epoch：5 初始学习率为0.001 95.66%

使用LeNet网络

与Resnet相比换了一个网络和改了数据加载模块，其他没啥变化。

库函数导入

1
2
3
4
5
6
7
8
9
import matplotlib.pyplot as plt
import torch
import torch.nn as nn
import torchvision
import torchvision.transforms as transforms
from sklearn.metrics import confusion_matrix, classification_report
from itertools import chain
import multiprocessing
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

数据集加载及增强操作

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
transform = transforms.Compose(
        [transforms.RandomHorizontalFlip(),
         transforms.RandomCrop(32, padding=4),
         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)
train_loader = torch.utils.data.DataLoader(trainset, batch_size=32,
                                              shuffle=True, num_workers=2)

testset = torchvision.datasets.CIFAR10(root='./data', train=False,
                                           download=True, transform=transform)
test_loader = torch.utils.data.DataLoader(testset, batch_size=32,
                                             shuffle=False, num_workers=2)

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

def output_label(label):
    output_mapping = {
        0: "plane", 1: "car", 2: "bird", 3: "cat", 4: "deer",
        5: "dog", 6: "frog", 7: "horse", 8: "ship", 9: "truck"
    }
    input = (label.item() if type(label) == torch.Tensor else label)
    return output_mapping[input]

LeNet模型

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
class C1(nn.Module):
    def __init__(self):
        super(C1, self).__init__()
        self.c1 = nn.Sequential(OrderedDict([
            ('conv1', nn.Conv2d(3, 6, kernel_size=(5, 5))),
            ('relu1', nn.ReLU()),
            ('pool1', nn.MaxPool2d(kernel_size=(2, 2), stride=2))
        ]))

    def forward(self, img):
        output = self.c1(img)
        return output

class C3(nn.Module):
    def __init__(self):
        super(C3, self).__init__()
        self.c3 = nn.Sequential(OrderedDict([
            ('conv3', nn.Conv2d(6, 16, kernel_size=(5, 5))),
            ('relu3', nn.ReLU()),
            ('pool3', nn.MaxPool2d(kernel_size=(2, 2), stride=2))
        ]))

    def forward(self, img):
        output = self.c3(img)
        return output

class F4(nn.Module):
    def __init__(self):
        super(F4, self).__init__()
        self.f4 = nn.Sequential(OrderedDict([
            ('fc4', nn.Linear(16 * 5 * 5, 120)),
            ('relu4', nn.ReLU())
        ]))

    def forward(self, img):
        output = self.f4(img)
        return output

class F5(nn.Module):
    def __init__(self):
        super(F5, self).__init__()
        self.f5 = nn.Sequential(OrderedDict([
            ('fc5', nn.Linear(120, 84)),
            ('relu5', nn.ReLU())
        ]))

    def forward(self, img):
        output = self.f5(img)
        return output

class F6(nn.Module):
    def __init__(self):
        super(F6, self).__init__()
        self.f6 = nn.Sequential(OrderedDict([
            ('fc6', nn.Linear(84, 10))
        ]))

    def forward(self, img):
        output = self.f6(img)
        return output

class LeNet5(nn.Module):
    def __init__(self):
        super(LeNet5, self).__init__()
        self.c1 = C1()
        self.c3 = C3()
        self.f4 = F4()
        self.f5 = F5()
        self.f6 = F6()

    def forward(self, img):
        output = self.c1(img)
        output = self.c3(output)
        output = output.view(-1, 16 * 5 * 5)
        output = self.f4(output)
        output = self.f5(output)
        output = self.f6(output)
        return output

模型训练超参数设置

1
2
3
4
5
6
7
8
9
10
11
12
13
#定义损失函数，初始学习率，优化器
criterion = nn.CrossEntropyLoss()
learning_rate = 0.001
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)

#学习率调节器，前期需要较大步长来快速收敛，后期如还用较大步长可能会在极小值周围震荡，所以后期学习率逐步衰减。可以提高训练效率和性能。
def exp_lr_scheduler(optimizer, epoch, init_lr=0.001, lr_decay_epoch=2):
    lr = init_lr * (0.1 ** (epoch // lr_decay_epoch))
    if epoch % lr_decay_epoch == 0:
        print(f'LR is set to {lr}')
    for param_group in optimizer.param_groups:
        param_group['lr'] = lr
    return optimizer

模型训练

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
num_epochs = 20
count = 0
loss_list = []
iteration_list = []
accuracy_list = []
predictions_list = []
labels_list = []

for epoch in range(num_epochs):
    model.train() 
    running_loss = 0.0

    optimizer = exp_lr_scheduler(optimizer, epoch, init_lr=learning_rate, lr_decay_epoch=2)

    for i, (images, labels) in enumerate(train_loader):
        images, labels = images.to(device), labels.to(device)

        outputs = model(images)
        loss = criterion(outputs, labels)

        optimizer.zero_grad()
        loss.backward()
        optimizer.step()

        running_loss += loss.item()
        count += 1

        if (i + 1) % 50 == 0:
            print(f'Epoch [{epoch + 1}/{num_epochs}], Step [{i + 1}/{len(train_loader)}], Loss: {loss.item():.4f}')

            model.eval() 
            total = 0
            correct = 0
            test_predictions = []
            test_labels = []

            with torch.no_grad():
                for images, labels in test_loader:
                    images, labels = images.to(device), labels.to(device)
                    test_labels.extend(labels.cpu().numpy())

                    outputs = model(images)
                    _, predicted = torch.max(outputs.data, 1)
                    test_predictions.extend(predicted.cpu().numpy())

                    total += labels.size(0)
                    correct += (predicted == labels).sum().item()

            accuracy = 100 * correct / total
            print(f'Accuracy on test set: {accuracy:.2f}%')

            loss_list.append(loss.item())
            iteration_list.append(count)
            accuracy_list.append(accuracy)

            predictions_list.append(test_predictions)
            labels_list.append(test_labels)

            model.train()

print('Finished Training')

实验结果保存

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
torch.save(model.state_dict(), 'lenet5_cifar10.pth') 
print('Model saved to lenet5_cifar10.pth')

plt.figure(figsize=(10, 5))
plt.plot(iteration_list, loss_list)
plt.xlabel("No. of Iteration")
plt.ylabel("Loss")
plt.title("Iterations vs Loss")
plt.savefig('loss_curve.png')
plt.show()

plt.figure(figsize=(10, 5))
plt.plot(iteration_list, accuracy_list)
plt.xlabel("No. of Iteration")
plt.ylabel("Accuracy")
plt.title("Iterations vs Accuracy")
plt.savefig('accuracy_curve.png')
plt.show()

class_correct = [0. for _ in range(10)]
total_correct = [0. for _ in range(10)]

model.eval()
with torch.no_grad():
    for images, labels in test_loader:
        images, labels = images.to(device), labels.to(device)
        outputs = model(images)
        _, predicted = torch.max(outputs, 1)
        c = (predicted == labels).squeeze()

        for i in range(labels.size(0)):
            label = labels[i]
            class_correct[label] += c[i].item()
            total_correct[label] += 1

for i in range(10):
    print(f"Accuracy of {classes[i]}: {100 * class_correct[i] / total_correct[i]:.2f}%")

flat_predictions = list(chain.from_iterable(predictions_list))
flat_labels = list(chain.from_iterable(labels_list))

cm = confusion_matrix(flat_labels, flat_predictions)
print("Confusion Matrix:")
print(cm)

print("Classification report for LeNet on CIFAR-10:") 
print(classification_report(flat_labels, flat_predictions, target_names=classes))

实验结果分析

使用图像增强操作

Batchsize：128 epoch：20 lr_decay_epoch：5 初始学习率为 0.001 60.56%

未使用图像增强操作

Batchsize：128 epoch：20 lr_decay_epoch：5 初始学习率为 0.001 62.66%

使用FashionNet

代码详见github

图像增强操作训练结果分析

使用图像增强操作

Batchsize：128 epoch：20 lr_decay_epoch：5 初始学习率为 0.001 71.51%

未使用图像增强操作

Batchsize：128 epoch：30 lr_decay_epoch：6 初始学习率为 0.01 71.69%

Batchsize：128 epoch：20 lr_decay_epoch：5 初始学习率为 0.001 74.25%

Gradcam实现图像特征可视化

Grad-CAM 的目标层均为最后一个卷积层

Resnet效果图

参数配置 Batch_size：64 epoch：20 lr_decay_epoch：5 lr：0.001 95.66%

图片来自谷歌。

FashionNet效果图

参数配置 Bachsize：128 epoch：20 lr_decay_epoch：5 初始学习率为 0.001 74.25%

LeNet效果图

参数配置 Bachsize：128 epoch：20 lr_decay_epoch：5 初始学习率为 0.001 60.56%

总结

数据集在不同模型上的准确率

模型 准确率
ResNet 95.66%
LeNet 62.66%
FashionNet 74.25%

LeNet对数据集进行图像增强操作的影响

分析：由于模型过于简单，本来就会出现欠拟合现象，继续进行图像增强操作后会导致模型更加欠拟合。所以出现三个模型精度差距较大，主要是因为模型复杂度的差距。虽然是欠拟合，但是继续增加epoch也并不会提高精度，因为模型对数据已经学到了尽可能多的知识。

未归一化 归一化 图像增强操作 大量图像增强操作
58.75% 62.66% 60.56% 56.78%

神经网络参数对模型准确率的影响

batch_size：64或128并无太大影响。

初始学习率：由实验结果可得lr=0.01时，效果较差，三个网络都使用了0.01进行测试得出的结果，有的实验结果没贴图。所以直接设置初始学习率为0.001可以更节省时间，提高效率。当学习率为1e-6时，发现准确率更新较小，所以准确率最小设置为1e-6。即学习率梯度为1e-3,1e-4,1e-5,1e-6，可以有较好的效果。

lr_decay_epoch和epoch：即学习率梯度为1e-3,1e-4,1e-5,1e-6，可以有较好的效果。即令epoch / lr_decay_epoch = 3或4 都可以，具体以epoch大小为准。