（动手学习深度学习）第13章 计算机视觉---图像增广与微调

13.1 图像增广

总结

• 数据增广通过变形数据来获取多样性从而使得模型泛化性能更好
• 常见图片增广包裹翻转、切割、变色。

图像增广代码实现

导入相关库

%matplotlib inline
import torch
import torchvision
from torch import nn
from d2l import torch as d2l
1
2
3
4
5

查看图像

# 查看图像
d2l.set_figsize()
img = d2l.Image.open('../img/cat1.jpg')
d2l.plt.imshow(img)
1
2
3
4

定义函数apply: 在输入图像img上多次运行图像增广方法aug并显示所有结果。

def apply(img, aug, num_rows=2, num_cols=4, scale=1.52):
    Y = [aug(img) for _ in range(num_rows * num_cols)]
    d2l.show_images(Y, num_rows, num_cols, scale=scale)
1
2
3

• 水平翻转图像（常用）

# RandomHorizontalFlip():水平翻转,默认50%
apply(img, torchvision.transforms.RandomHorizontalFlip())
1
2

• 上下翻转（不常用）

# RandomVerticalFlip():上下翻转，默认50%（针对不同的数据集，选择）
apply(img, torchvision.transforms.RandomVerticalFlip())
1
2

• 缩放

apply(img, torchvision.transforms.Resize(256))  # 只有一个值时，按长宽比缩放
1

test_augs = torchvision.transforms.Compose(
    [torchvision.transforms.Resize([256, 256]),
     torchvision.transforms.CenterCrop(224)]
)
apply(img, test_augs)
1
2
3
4
5

apply(img, torchvision.transforms.Resize([256, 256]))
1

test_augs2 = torchvision.transforms.Compose(
    [torchvision.transforms.Resize(256),  # 先按比例放大
     torchvision.transforms.CenterCrop(224)]  # 然后中心裁剪
)
apply(img, test_augs2)
1
2
3
4
5

• 随机剪裁

# 随机裁剪：scale(裁剪面积与原始面积的百分比)、ratio(裁剪区域的宽高比)
shape_aug = torchvision.transforms.RandomResizedCrop(
    (200, 200), scale=(0.1, 1), ratio=(0.5, 2)
)
apply(img, shape_aug)
1
2
3
4
5

• 随机更改图像的亮度

# 更改图像的：brightness(亮度)、contrast（对比度）、saturation(色调)
apply(img, torchvision.transforms.ColorJitter(
    brightness=0.5, contrast=0, saturation=0, hue=0
))
1
2
3
4

• 随机更改图像的色调

apply(img, torchvision.transforms.ColorJitter(
    brightness=0, contrast=0, saturation=0, hue=0.5
))
1
2
3

• 随机更改图像的亮度、色调

apply(img, torchvision.transforms.ColorJitter(
    brightness=0.5, contrast=0.5, saturation=0.5, hue=0.5
))
1
2
3

• 结合多种图像增广的方法

augs = torchvision.transforms.Compose(
    [torchvision.transforms.RandomHorizontalFlip(),
     torchvision.transforms.ColorJitter(brightness=0.5, contrast=0.5, saturation=0.5, hue=0.5),
     torchvision.transforms.RandomResizedCrop((200, 200), scale=(0.1, 1), ratio=(0.5, 2))]
)
apply(img, augs)
1
2
3
4
5
6

利用图像增广的Resnet18对CIFAR10训练

1. 查看CIFAR10数据

all_imgs = torchvision.datasets.CIFAR10(
    train=True, root='../data', download=True
)
d2l.show_images(
    [all_imgs[i][0] for i in range(32)], 4, 8, scale=0.8
)
1
2
3
4
5
6

2. 只使用最简单的随机左右翻转

train_augs = torchvision.transforms.Compose(
    [torchvision.transforms.RandomHorizontalFlip(),
     torchvision.transforms.ToTensor()]
)
test_augs = torchvision.transforms.Compose(  # 验证集不需要数据增强
    [torchvision.transforms.ToTensor()]
)
1
2
3
4
5
6
7

3. 定义一个辅助函数，以便读取图像和应用图像增广

def load_cifar10(is_train, augs, batch_size):
    dataset = torchvision.datasets.CIFAR10(
        root='../data', train=is_train,
        transform=augs, download=True
    )
    dataloader = torch.utils.data.DataLoader(
        dataset, batch_size=batch_size, shuffle=is_train,num_workers=4
    )
    return dataloader
1
2
3
4
5
6
7
8
9

4. 定义一个函数，使用多gpu对模型进行训练和评估

def train_batch_ch13(net, X, y, loss, trainer, devices):
    """使用多GPU进行小批量训练"""
    if isinstance(X, list):
        X = [x.to(devices[0]) for x in X]
    else:
        X = X.to(devices[0])
    y = y.to(devices[0])
    net.train()
    trainer.zero_grad()
    pred = net(X)
    l = loss(pred, y)
    l.sum().backward()
    trainer.step()
    train_loss_sum = l.sum()
    train_acc_sum = d2l.accuracy(pred, y)
    return train_loss_sum, train_acc_sum
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16

def train_ch13(net, train_iter, test_iter, loss, trainer, num_epoces, devices=d2l.try_all_gpus()):
    """使用多GPU进行模型训练"""
    timer, num_batches = d2l.Timer(), len(train_iter)
    animator = d2l.Animator(xlabel='epoch', xlim=[1, num_epoces], ylim=[0, 1],
                            legend=['train loss', 'train acc', 'test acc'])
    net = nn.DataParallel(net, device_ids=devices).to(devices[0])   # 使用DataParallel在多个GPU上复制并行化网络
    for epoch in range(num_epoces):
        # 4个维度：存储训练损失、训练准确度、实例数、特征数
        metric = d2l.Accumulator(4)
        for i, (features, labels) in enumerate(train_iter):
            timer.start()
            l, acc = train_batch_ch13(net, features, labels, loss, trainer, devices)
            metric.add(l, acc, labels.shape[0], labels.numel())
            timer.stop()
            if (i+1) % (num_batches // 5) == 0 or i == num_batches -1:
                animator.add(epoch + (i+1) / num_batches, (metric[0]/metric[2], metric[1]/metric[3], None))
        test_acc = d2l.evaluate_accuracy_gpu(net, test_iter)
        animator.add(epoch+1, (None, None, test_acc))
    print(f'loss {metric[0]/metric[2]:.3f}, '
          f'train acc {metric[1]/metric[3]:.3f}, test acc {test_acc:.3f}')
    print(f'{metric[2]*num_epoces / timer.sum():.1f} example/sec on {str(devices)}')
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21

4. 定义train_with_data_aug函数，使用图像增广来训练模型

batch_size, devices, net = 256, d2l.try_all_gpus(), d2l.resnet18(10, 3)

def init_weights(m):
    if type(m) in [nn.Linear, nn.Conv2d]:
        nn.init.xavier_uniform_(m.weight)

net.apply(init_weights)

def train_with_data_aug(train_augs, test_augs, net, lr=0.001):
    train_iter = load_cifar10(True, train_augs, batch_size)
    test_iter = load_cifar10(False, test_augs, batch_size)
    loss = nn.CrossEntropyLoss(reduction='none')
    trainer = torch.optim.Adam(net.parameters(), lr=lr)
    train_ch13(net, train_iter, test_iter, loss, trainer, 10, devices)
1
2
3
4
5
6
7
8
9
10
11
12
13
14

5. 训练模型

import time

# 在开头设置开始时间
start = time.perf_counter()  # start = time.clock() python3.8之前可以
train_with_data_aug(train_augs, test_augs, net)

# 在程序运行结束的位置添加结束时间
end = time.perf_counter()  # end = time.clock()  python3.8之前可以

# 再将其进行打印，即可显示出程序完成的运行耗时
print(f'运行耗时{(end-start):.4f}')
1
2
3
4
5
6
7
8
9
10
11