In this paper, we will classify images of CIFAR-10 using PyTorch. Follow the Official Tutorial with comments. Python and machine learning are super beginners.

What is CIFAR-10?

A 10-label image dataset that is widely used in the machine learning area. airplane、automobile、bird、cat、deer、dog、frog、horse、ship、truck 10 labels are available.

environment

macOS Catalina
Python 3.7.2
pip 19.1.1

Install PyTorch

Official site will issue an installation command according to each environment. Since I'm macOS, run the following to install.

pip install torch torchvision

Implement CNN

Import the required libraries

#Import NumPy, Matplotlib, PyTorch
import numpy as np
import matplotlib.pyplot as plt
import torch
import torchvision
import torchvision.transforms as transforms
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim

Download CIFAR-10

#ToTensor: Grayscale image (RGB 0)~255 to 0~Normalize to the range of 1), Normalize: Z-value (RGB mean and standard deviation to 0).Normalize with 5)
transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])

#Download training data
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)

#Download test data
testset = torchvision.datasets.CIFAR10(root='./data', train=False, download=True, transform=transform)
testloader = torch.utils.data.DataLoader(testset, batch_size=4, shuffle=True, num_workers=2)

Check out CIFAR-10

Check the data

#Training dataset: 50,000 RGB images with 32 pixels in height and width
print(trainset.data.shape)
(50000, 32, 32, 3)

#Test dataset: 10000 RGB images with 32 pixels in height and width
print(testset.data.shape)
(10000, 32, 32, 3)

#Check the class list
print(trainset.classes)
['airplane', 'automobile', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck']

#Classes are often used, so keep them separately
classes = trainset.classes

In the official documentation, ** airplane was redefined as plane ** and ** automobile was redefined as car **. Why?

Display image

#Try to display the downloaded image
def imshow(img):
    #Denormalize
    img = img / 2 + 0.5
    # torch.From tensor type to numpy.Convert to ndarray type
    print(type(img)) # <class 'torch.Tensor'>
    npimg = img.numpy()
    print(type(npimg))    
    #Convert shape from (RGB, vertical, horizontal) to (vertical, horizontal, RGB)
    print(npimg.shape)
    npimg = np.transpose(npimg, (1, 2, 0))
    print(npimg.shape)
    #Display image
    plt.imshow(npimg)
    plt.show()

dataiter = iter(trainloader)
images, labels = dataiter.next()
imshow(torchvision.utils.make_grid(images))
print(' '.join('%5s' % classes[labels[j]] for j in range(4)))

Implement the network

#Implement CNN
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

net = Net()

Define each layer with init and connect them with forward.

Define the loss function / optimizer

#Cross entropy
criterion = nn.CrossEntropyLoss()
#Stochastic gradient descent
optimizer = optim.SGD(net.parameters(), lr=0.001, momentum=0.9)

To train

#To train
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)
        #Backpropagation of error
        loss.backward()
        optimizer.step()
        train_loss = loss.item()
        running_loss += loss.item()
        if i % 2000 == 1999:
            print('[%d, %5d] loss: %.3f' % (epoch + 1, i + 1, running_loss / 2000))
            running_loss = 0.0
print('Finished Training')

[1,  2000] loss: 2.164
[1,  4000] loss: 1.863
[1,  6000] loss: 1.683
[1,  8000] loss: 1.603
[1, 10000] loss: 1.525
[1, 12000] loss: 1.470
[2,  2000] loss: 1.415
[2,  4000] loss: 1.369
[2,  6000] loss: 1.363
[2,  8000] loss: 1.333
[2, 10000] loss: 1.314
[2, 12000] loss: 1.317
Finished Training

The average value of Loss for every 2000 mini-batch is output to the log.

Save the model

#Save the model
PATH = './cifar_net.pth'
torch.save(net.state_dict(), PATH)

Save the model in the current directory with the extension pth (PyTorch).

Try using the model

#Load test data and display image and correct label
dataiter = iter(testloader)
images, labels = dataiter.next()
imshow(torchvision.utils.make_grid(images))
print('GroundTruth: ', ' '.join('%5s' % classes[labels[j]] for j in range(4)))
#Load the saved model and predict
net = Net()
net.load_state_dict(torch.load(PATH))
outputs = net(images)
_, predicted = torch.max(outputs, 1)
print('Predicted: ', ' '.join('%5s' % classes[predicted[j]] for j in range(4)))

GroundTruth: truck cat airplane ship Predicted: truck horse airplane ship

You can see that the predictions are correct except for cat.

print(outputs)
value, predicted = torch.max(outputs, 1)
print(value)
print(predicted)

tensor([[ 0.7114, -2.2724,  0.1225,  0.9470,  2.1940,  1.8655, -2.6655,  4.1646,
         -1.1001, -1.6991],
        [-2.2453, -4.1017,  1.8291,  3.2079,  1.1242,  3.6712,  1.0010,  1.0489,
         -3.2010, -1.9476],
        [-3.0669, -3.8900,  0.9312,  3.5649,  2.7791,  1.5095,  2.1216,  1.5274,
         -4.3077, -2.2234],
        [-2.0948, -3.4640,  2.4833,  2.6210,  4.0590,  1.8350,  0.4924,  0.7212,
         -3.5043, -2.4212]], grad_fn=<AddmmBackward>)
tensor([4.1646, 3.6712, 3.5649, 4.0590], grad_fn=<MaxBackward0>)
tensor([7, 5, 3, 4])

** torch.max ** returns the maximum value of outputs.

Test the model

correct = 0
total = 0
#Calculate without remembering the gradient (without learning)
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))

Accuracy of the network on the 10000 test images: 60 %

You can see that the correct answer rate for 10000 test data is 60%.

Below is a personal note for Python beginners. Not good ** (predicted == labels) .sum (). item () ** I didn't know how to write this, so I'll log it out and check it.

print(type((predicted == labels)))
print((predicted == labels).dtype)
print(type((predicted == labels).sum()))
print((predicted == labels).sum())
print((predicted == labels).sum().item())
# <class 'torch.Tensor'>
# torch.bool
# <class 'torch.Tensor'>
# tensor(2)
# 2

I see. Compare each element of the array and use sum () implemented in torch.Tensor to calculate the total value of true. After that, item () implemented in torch.Tensor is used to make the total value an int type numerical value. It was a little easier to understand when I checked it with numpy.

#Try with numpy
a = np.array([1, 2, 3, 4, 5])
b = np.array([1, 2, 0, 4, 5])
print(type((a == b)))
print((a == b))
print((a == b).sum())
print(type((a == b).sum()))
print((a == b).sum().item())
print(type((a == b).sum().item()))
# <class 'numpy.ndarray'>
# [ True  True False  True  True]
# 4
# <class 'numpy.int64'>
# 4
# <class 'int'>

Looking at Official, you can use almost the same API as ndarray, so ** sum () ** and ** item () ** Can be used. Convinced.

Let's see the correct answer rate for each label

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 airplane : 72 %
Accuracy of automobile : 66 %
Accuracy of  bird : 38 %
Accuracy of   cat : 58 %
Accuracy of  deer : 60 %
Accuracy of   dog : 29 %
Accuracy of  frog : 73 %
Accuracy of horse : 60 %
Accuracy of  ship : 69 %
Accuracy of truck : 73 %

Is it like this in a tutorial?

[PyTorch] Image classification of CIFAR-10