首頁.jpg

• 從更高水平理解Pytorch的Tensor（張量）和神經(jīng)網(wǎng)絡(luò)
• 訓(xùn)練一個小的圖像分類神經(jīng)網(wǎng)絡(luò)
  注意確定已經(jīng)安裝了torch和torchvision

訓(xùn)練一個分類器

上一節(jié)粗略地看到如何定義神經(jīng)網(wǎng)絡(luò)、計算損失、更新網(wǎng)絡(luò)權(quán)重數(shù)據(jù)，但是，數(shù)據(jù)呢？

數(shù)據(jù)

通常來說，當(dāng)你需要處理圖片、文字、音頻或視頻數(shù)據(jù)時，你可以使用標(biāo)準(zhǔn)python包加載數(shù)據(jù)到numpy中，然后將這些數(shù)據(jù)轉(zhuǎn)為torch.*Tensor。

• 圖像使用Pillow或是OpenCV
• 音頻使用scipy或是librosa
• 文本使用NLTP或是Spacy
  為了可視化，Pytorch提供一個包torchvision，它包含常用數(shù)據(jù)集（Imagenet、CIFAR10、MNIST等）的加載，同時還有轉(zhuǎn)換圖像用的工具。
  在這個教程中，使用CIFAR10數(shù)據(jù)集，包括‘飛機’‘汽車’‘鳥’‘貓’‘鹿’‘狗’‘青蛙’等分類。 其中的圖片為33232（3通道顏色，32*32像素大小）。
  
  CIFAR10example

訓(xùn)練一個圖像分類器

這里使用CNN，包括以下步驟

• 使用torchvision加載和歸一CIFAR10訓(xùn)練和測試數(shù)據(jù)集
• 定義一個CNN
• 定義一個損失函數(shù)
• 在訓(xùn)練集上訓(xùn)練網(wǎng)絡(luò)
• 在測試集上測試網(wǎng)絡(luò)
  下面一步一步來實現(xiàn)。這里要關(guān)注的時對網(wǎng)絡(luò)的定義、對數(shù)據(jù)的加載。

1. 加載和歸一CIRAR10數(shù)據(jù)

加載數(shù)據(jù)很簡單，同樣數(shù)據(jù)要從網(wǎng)上下載，和很久以前處理MNIST數(shù)據(jù)一樣，只是這次我在學(xué)校，直接下載了。速度會比較慢。

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)
trainloader = torch.utils.data.DataLoader(testset, batch_size=4, shuffle=False, num_workers=2)

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

數(shù)據(jù)下載下來了，我們可以看看是什么東東。

import matplotlib.pyplot as plt
import numpy as np

# functions to show an image
def imshow(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
    imshow(torchvision.utils.make_grid(images))
    # print labels
    print(' '.join('%5s' % classes[labels[j]] for j in range(4)))

cat ship ship plane

2. 定義一個CNN（卷積神經(jīng)網(wǎng)絡(luò)）

從前面定義神經(jīng)網(wǎng)絡(luò)部分復(fù)制神經(jīng)網(wǎng)絡(luò)并修改它以獲取3通道圖像（而不是定義的1通道圖像）。

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)))
        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

3. 定義損失函數(shù)和優(yōu)化函數(shù)

直接使用隨機梯度下降SGD

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

4. 訓(xùn)練網(wǎng)絡(luò)

忙了大半天，可以訓(xùn)練了

# 在這里只進行兩次迭代
for epoch in range(2):
    running_loss = 0.0
    for i, data in enumerate(trainloader, 0):
        inputs, labels = data
        optimizer.zero_grad()
        outputs = net(inputs)
        loss = criterion(outputs, labels)
        loss.backward()
        optimizer.step()

        running_loss += loss.item()
        if i % 2000 == 1999:
            print('[%d, %5d] loss: %.3f' % (epoch + 1, i + 1, running_loss / 2000))
            running_loss = 0.0
[1,  2000] loss: 2.249
[1,  4000] loss: 1.973
[1,  6000] loss: 1.739
[1,  8000] loss: 1.622
[1, 10000] loss: 1.565
[1, 12000] loss: 1.512
[2,  2000] loss: 1.439
[2,  4000] loss: 1.406
[2,  6000] loss: 1.375
[2,  8000] loss: 1.381
[2, 10000] loss: 1.348
[2, 12000] loss: 1.290

可以看到損失數(shù)值是在降低。

5. 測試網(wǎng)絡(luò)

    # 調(diào)用之前的方法，打印圖片
    imshow(torchvision.utils.make_grid(images))
    print('GroundTruth: ', ' '.join('%5s' % classes[labels[j]] for j in range(4)))
    outputs = net(images)
    _, predicted = torch.max(outputs, 1)

    print('Predicted: ', ' '.join('%5s' % classes[predicted[j]] for j in range(4)))
    
    
GroundTruth:    cat  ship  ship plane
Predicted:   bird   car  ship plane

Accuracy of the network on the 10000 test images: 54 %

可以看到，預(yù)測四個，只對了兩個……這個是和迭代次數(shù)太少有關(guān)系。
再看一下整體的正確率和各類型的正確率

    correct = 0
    total = 0
    with torch.no_grad():
        for data in testloader:
            images, labels = data
            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))

    class_correct = list(0. for i in range(10))
    class_total = list(0. for i in range(10))
    with torch.no_grad():
        for data in testloader:
            images, labels = data
            outputs = net(images)
            _, predicted = torch.max(outputs, 1)
            c = (predicted == labels).squeeze()
            for i in range(4):
                label = labels[i]
                class_correct[label] += c[i].item()
                class_total[label] += 1

    for i in range(10):
        print('Accuracy of %5s : %2d %%' % (
            classes[i], 100 * class_correct[i] / class_total[i]))



Accuracy of the network on the 10000 test images: 54 %
Accuracy of plane : 57 %
Accuracy of   car : 64 %
Accuracy of  bird : 40 %
Accuracy of   cat : 36 %
Accuracy of  deer : 33 %
Accuracy of   dog : 33 %
Accuracy of  frog : 76 %
Accuracy of horse : 64 %
Accuracy of  ship : 69 %
Accuracy of truck : 64 %

整體正確率為54%，各類的正確率不一樣。