(一)數(shù)據(jù)增強(qiáng)（增廣）

（1）為什么要做數(shù)據(jù)增強(qiáng)概行？

一個(gè)原因是可能你的數(shù)據(jù)集比較小细层，所以需要對(duì)數(shù)據(jù)進(jìn)行簡(jiǎn)單的操作，讓數(shù)據(jù)集增加倦逐。第二是有這樣的一個(gè)真實(shí)實(shí)例轿曙。有一家做智能售貨機(jī)的公司在公司內(nèi)部調(diào)試好參數(shù)訓(xùn)練好模型以后，將售貨機(jī)拿去展廳進(jìn)行測(cè)試的時(shí)候僻孝，發(fā)現(xiàn)原本準(zhǔn)確率非常高的機(jī)器忽然識(shí)別不出來(lái)了。原因是展廳的光源不一樣守谓，導(dǎo)致整個(gè)的測(cè)試數(shù)據(jù)集就和訓(xùn)練集發(fā)生了很大的變化穿铆。所以在產(chǎn)品研發(fā)的時(shí)候適當(dāng)?shù)耐ㄟ^(guò)數(shù)據(jù)增強(qiáng)技術(shù)能夠給模型增加魯棒性。

（2）數(shù)據(jù)增強(qiáng)方法有什么斋荞？

可通過(guò)在圖片中加入各種不一樣的背景噪音荞雏，改變圖片的顏色和形狀。

• 翻轉(zhuǎn)：左右翻轉(zhuǎn)平酿、上下翻轉(zhuǎn)

• 切割：在圖片中切割出一塊凤优，然后變形到固定的形狀

• 顏色：改變色調(diào)，飽和度蜈彼，明度等

（二）代碼實(shí)現(xiàn)

%matplotlib inline
import torch 
import torchvision
from torch import nn 
from d2l import torch as d2l
import matplotlib.image as img
import matplotlib.pyplot as plt

# 打開(kāi)圖片的方法
# image = img.imread('../img/cat1.jpg')
# plt.title("cat.jpg")
# plt.axis("off")
# plt.imshow(image)
# plt.show()

d2l.set_figsize()
img = d2l.Image.open('../img/cat1.jpg')
d2l.plt.imshow(img)

# 參數(shù)列表（圖片筑辨，增強(qiáng)的辦法，多少行幸逆，多少列棍辕，倍數(shù)）
def apply(img, aug, num_rows=2, num_cols=4, scale=1.5):
    y = [aug(img) for _ in range(num_rows*num_cols)]
    d2l.show_images(y, num_rows, num_cols, scale=scale)

# 左右翻轉(zhuǎn)圖片
apply(img,torchvision.transforms.RandomHorizontalFlip())

# 上下翻轉(zhuǎn)
apply(img,torchvision.transforms.RandomVerticalFlip())

# 隨即裁剪
shape_aug = torchvision.transforms.RandomResizedCrop(
    # （輸出大小，選擇的比例还绘，高寬比）
    size=(200,200),scale=(0.1,1),ratio=(0.5,2))
apply(img,shape_aug)

# 隨機(jī)更改圖片的亮度
apply(img,torchvision.transforms.ColorJitter(
    # (亮度區(qū)間楚昭，對(duì)比度，飽和度拍顷，色調(diào))
    brightness=0.5,contrast=0.5,saturation=0.5,hue=0.5
))

# 結(jié)合多種數(shù)據(jù)增強(qiáng)方法
augs = torchvision.transforms.Compose([
    torchvision.transforms.RandomHorizontalFlip(),
    shape_aug,
    torchvision.transforms.ColorJitter(brightness=0.5,contrast=0.5,saturation=0.5,hue=0.5)
])

apply(img, augs)

# 如果下載報(bào)錯(cuò)的話抚太，自己去網(wǎng)頁(yè)上下載
# https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz
all_images = torchvision.datasets.CIFAR10(
    train=True,
    root="../data",
    download=False
)

d2l.show_images([all_images[i][0] for i in range(32)],4,8,scale=0.8)
# d2l.show_images([all_images.data[i] for i in range(32)],4,8,scale=0.8)

train_augs = torchvision.transforms.Compose([
    torchvision.transforms.RandomHorizontalFlip(),
    torchvision.transforms.ToTensor()
])
test_augs = torchvision.transforms.Compose([
    torchvision.transforms.ToTensor()
])

def load_cifar10(is_train,augs, batch_size):
    dataset = torchvision.datasets.CIFAR10(
        train=is_train,
        root="../data",
        download=False,
        transform=augs,
    )
    data_loader = torch.utils.data.DataLoader(
        dataset,
        batch_size=batch_size,
        shuffle=True,
        num_workers=0
    )
    return data_loader

def train_batch_ch13(net, X, y, loss, trainer, devices):
    """用多GPU進(jìn)行小批量訓(xùn)練"""
    if isinstance(X, list):
        # 微調(diào)BERT中所需（稍后討論）
        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

#@save
def train_ch13(net, train_iter, test_iter, loss, trainer, num_epochs,
               devices=d2l.try_all_gpus()):
    """用多GPU進(jìn)行模型訓(xùn)練"""
    timer, num_batches = d2l.Timer(), len(train_iter)
    animator = d2l.Animator(xlabel='epoch', xlim=[1, num_epochs], ylim=[0, 1],
                            legend=['train loss', 'train acc', 'test acc'])
    net = nn.DataParallel(net, device_ids=devices).to(devices[0])
    for epoch in range(num_epochs):
        # 4個(gè)維度：儲(chǔ)存訓(xùn)練損失，訓(xùn)練準(zhǔn)確度昔案，實(shí)例數(shù)尿贫，特點(diǎn)數(shù)
        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}, train acc '
          f'{metric[1] / metric[3]:.3f}, test acc {test_acc:.3f}')
    print(f'{metric[2] * num_epochs / timer.sum():.1f} examples/sec on '
          f'{str(devices)}')

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)

train_with_data_aug(train_augs, test_augs, net)