CIFAR10彩色图片识别

`- 🍨 本文为🔗365天深度学习训练营 中的学习记录博客

• 🍖 原作者：K同学啊`

前期准备

运行环境

• 语言环境：Python 3.10
• 编译器：Google Colab
• 深度学习环境：
• torch==2.3.1+cu121
• torchvision==0.18.1+cu121

模块导入

# Import package
import torch
import torch.nn as nn
import matplotlib.pyplot as plt
import torchvision
import numpy as np
import torch.nn.functional as F
from torchinfo import summary

如果设备上支持GPU就使用GPU,否则使用CPU。

# Training our model using GPU
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

导入数据

使用dataset下载CIFAR10数据集，并划分好训练集与测试集。

# Download data
train_ds = torchvision.datasets.CIFAR10('data', 
                                      train=True, 
                                      transform=torchvision.transforms.ToTensor(), # 将数据类型转化为Tensor
                                      download=True)

test_ds  = torchvision.datasets.CIFAR10('data', 
                                      train=False, 
                                      transform=torchvision.transforms.ToTensor(), # 将数据类型转化为Tensor
                                      download=True)

使用dataloader加载数据，并设置好基本的batch_size。

# Load data
batch_size = 32
train_dl = torch.utils.data.DataLoader(train_ds, batch_size=batch_size, shuffle=True)
test_dl  = torch.utils.data.DataLoader(test_ds, batch_size=batch_size)

取一个批次查看数据格式，数据的shape为：[batch_size, channel, height, weight]。其中batch_size为自己设定，channel，height和weight分别是图片的通道数，高度和宽度。

# Check the data format of a batch
# Data shape: [batch_size, channels, height, width]
imgs, labels = next(iter(train_dl))
imgs.shape

数据可视化

transpose((1, 2, 0))详解：

• 作用是对NumPy数组进行轴变换，transpose函数的参数是一个元组，定义了新轴的顺序。原始PyTorch张量通常是以(C, H, W)的格式存储的，其中：
  • C是通道数（例如，RGB图像有3个通道）。
  • H是图像的高度。
  • W是图像的宽度。
• transpose((1, 2, 0))将轴的顺序从(C, H, W)转换为(H, W, C)，这使得数据格式更适合可视化和处理。
• PyTorch 默认图像格式为 (C, H, W)，而 Matplotlib 的 imshow 需要 (H, W, C)，因此需要用 transpose 调整轴顺序以正确显示图像。

# Specify the image size; create a plot with a figure size of 20 inches wide and 5 inches tall.
plt.figure(figsize=(20, 5)) 
for i, imgs in enumerate(imgs[:20]):
    # Perform axis transformation
    npimg = imgs.numpy().transpose((1, 2, 0))
    # Divide the figure into 2 rows and 10 columns, and plot the (i+1)-th subplot.
    plt.subplot(2, 10, i+1)
    plt.imshow(npimg, cmap=plt.cm.binary)
    plt.axis('off')  # Turn off the axis display

构建简单的CNN网络

对于一般的CNN网络来说，都是由特征提取网络和分类网络构成，其中特征提取网络用于提取图片的特征，分类网络用于将图片进行分类。

⭐1. torch.nn.Conv2d()详解

函数原型：

torch.nn.Conv2d(in_channels, out_channels, kernel_size, stride=1, padding=0, dilation=1, groups=1, bias=True, padding_mode='zeros', device=None, dtype=None)

关键参数说明：

• in_channels ( int ) – 输入图像中的通道数。
• out_channels ( int ) – 卷积产生的通道数。
• kernel_size ( int or tuple ) – 卷积核的大小。
• stride ( int or tuple , optional ) -- 卷积的步幅，默认值：1。
• padding ( int , tuple或str , optional ) – 添加到输入的所有四个边的填充，默认值：0。
• dilation (int or tuple, optional) - 扩张操作：控制kernel点（卷积核点）的间距，默认值:1。
• groups（int，可选）：将输入通道分组成多个子组，每个子组使用一组卷积核来处理。默认值为 1，表示不进行分组卷积。
• padding_mode (字符串,可选) – 'zeros', 'reflect', 'replicate'或'circular'. 默认：'zeros'

⭐2. torch.nn.Linear()详解

函数原型：

torch.nn.Linear(in_features, out_features, bias=True, device=None, dtype=None)

关键参数说明：

• in_features：每个输入样本的大小。
• out_features：每个输出样本的大小。

⭐3. torch.nn.MaxPool2d()详解

函数原型：

torch.nn.MaxPool2d(kernel_size, stride=None, padding=0, dilation=1, return_indices=False, ceil_mode=False)

关键参数说明：

• kernel_size：最大的窗口大小
• stride：窗口的步幅，默认值为kernel_size
• padding：填充值，默认为0
• dilation：控制窗口中元素步幅的参数

⭐4. 关于卷积层、池化层的计算：

下面的网络数据shape变化过程为：

3, 32, 32（输入数据） -> 64, 30, 30（经过卷积层1）-> 64, 15, 15（经过池化层1） -> 64, 13, 13（经过卷积层2）-> 64, 6, 6（经过池化层2） -> 128, 4, 4（经过卷积层3） -> 128, 2, 2（经过池化层3） -> 512 -> 256 -> num_classes(10)

num_classes = 10  # Number of image categories/classes

class Model(nn.Module):
    def __init__(self):
        super().__init__()
        # Feature extraction network
        self.conv1 = nn.Conv2d(3, 64, kernel_size=3)   # First convolutional layer with a 3x3 kernel
        self.pool1 = nn.MaxPool2d(kernel_size=2)       # Pooling layer with a 2x2 kernel
        self.conv2 = nn.Conv2d(64, 64, kernel_size=3)  # Second convolutional layer with a 3x3 kernel
        self.pool2 = nn.MaxPool2d(kernel_size=2)       # Second pooling layer with a 2x2 kernel
        self.conv3 = nn.Conv2d(64, 128, kernel_size=3) # Third convolutional layer with a 3x3 kernel
        self.pool3 = nn.MaxPool2d(kernel_size=2)       # Third pooling layer with a 2x2 kernel
                                      
        # Classification network
        self.fc1 = nn.Linear(512, 256)                # Fully connected layer: 512 input units to 256 output units
        self.fc2 = nn.Linear(256, num_classes)        # Fully connected layer: 256 input units to `num_classes` output units

    # Forward pass
    def forward(self, x):
        x = self.pool1(F.relu(self.conv1(x)))         # Apply the first convolution, ReLU activation, and pooling
        x = self.pool2(F.relu(self.conv2(x)))         # Apply the second convolution, ReLU activation, and pooling
        x = self.pool3(F.relu(self.conv3(x)))         # Apply the third convolution, ReLU activation, and pooling
        
        x = torch.flatten(x, start_dim=1)             # Flatten the feature maps into a 1D tensor for the fully connected layers

        x = F.relu(self.fc1(x))                       # Apply the first fully connected layer with ReLU activation
        x = self.fc2(x)                               # Apply the final fully connected layer for classification
        
        return x                                      # Return the output (logits) for each class

加载并打印模型。

# Move the model to the GPU (all model computations will be performed on the GPU) 
model = Model().to(device) # Display a summary of the model architecture summary(model)

训练模型

设置超参数

loss_fn    = nn.CrossEntropyLoss()  # Create the loss function (Cross-Entropy Loss)
learn_rate = 1e-2                   # Set the learning rate
opt        = torch.optim.SGD(model.parameters(), lr=learn_rate)  # Define the optimizer (Stochastic Gradient Descent) with the specified learning rate

编写训练函数

1. optimizer.zero_grad()

函数会遍历模型的所有参数，通过内置方法截断反向传播的梯度流，再将每个参数的梯度值设为0，即上一次的梯度记录被清空。

2. loss.backward()

PyTorch的反向传播(即tensor.backward())是通过autograd包来实现的，autograd包会根据tensor进行过的数学运算来自动计算其对应的梯度。

具体来说，torch.tensor是autograd包的基础类，如果你设置tensor的requires_grads为True，就会开始跟踪这个tensor上面的所有运算，如果你做完运算后使用tensor.backward()，所有的梯度就会自动运算，tensor的梯度将会累加到它的.grad属性里面去。

更具体地说，损失函数loss是由模型的所有权重w经过一系列运算得到的，若某个w的requires_grads为True，则w的所有上层参数（后面层的权重w）的.grad_fn属性中就保存了对应的运算，然后在使用loss.backward()后，会一层层的反向传播计算每个w的梯度值，并保存到该w的.grad属性中。

如果没有进行tensor.backward()的话，梯度值将会是None，因此loss.backward()要写在optimizer.step()之前。

3. optimizer.step()

step()函数的作用是执行一次优化步骤，通过梯度下降法来更新参数的值。因为梯度下降是基于梯度的，所以在执行optimizer.step()函数前应先执行loss.backward()函数来计算梯度。

注意：optimizer只负责通过梯度下降进行优化，而不负责产生梯度，梯度是tensor.backward()方法产生的。

# Training loop
def train(dataloader, model, loss_fn, optimizer):
    size = len(dataloader.dataset)  # Total size of the training dataset
    num_batches = len(dataloader)   # Number of batches

    train_loss, train_acc = 0, 0  # Initialize training loss and accuracy
    
    for X, y in dataloader:  # Iterate over batches, retrieving images (X) and their labels (y)
        X, y = X.to(device), y.to(device)  # Move data and labels to the specified device (e.g., GPU)
        
        # Compute prediction error
        pred = model(X)          # Model output (predictions)
        loss = loss_fn(pred, y)  # Calculate the loss between the predictions and the true labels (targets)
        
        # Backpropagation
        optimizer.zero_grad()  # Reset gradients to zero
        loss.backward()        # Perform backpropagation to compute gradients
        optimizer.step()       # Update model parameters using the computed gradients
        
        # Record accuracy and loss
        train_acc  += (pred.argmax(1) == y).type(torch.float).sum().item()  # Count correct predictions
        train_loss += loss.item()  # Accumulate the loss value
            
    train_acc  /= size  # Average accuracy over the entire dataset
    train_loss /= num_batches  # Average loss over all batches

    return train_acc, train_loss  # Return training accuracy and loss

编写测试函数

测试函数和训练函数大致相同，但是由于不进行梯度下降对网络权重进行更新，所以不需要传入优化器。

def test(dataloader, model, loss_fn):
    size = len(dataloader.dataset)  # Total size of the test dataset, e.g., 10,000 images
    num_batches = len(dataloader)   # Number of batches, e.g., 313 (10,000/32, rounded up)
    test_loss, test_acc = 0, 0      # Initialize test loss and accuracy
    
    # Disable gradient calculation to save memory and computation during testing
    with torch.no_grad():
        for imgs, target in dataloader:  # Iterate over test batches, retrieving images and their labels
            imgs, target = imgs.to(device), target.to(device)  # Move data and labels to the specified device (e.g., GPU)
            
            # Compute loss
            target_pred = model(imgs)  # Model output (predictions)
            loss = loss_fn(target_pred, target)  # Calculate the loss between predictions and true labels
            
            test_loss += loss.item()  # Accumulate the loss value
            test_acc += (target_pred.argmax(1) == target).type(torch.float).sum().item()  # Count correct predictions

    test_acc /= size  # Average accuracy over the entire dataset
    test_loss /= num_batches  # Average loss over all batches

    return test_acc, test_loss  # Return test accuracy and loss

正式训练

1. model.train()

model.train()的作用是启用 Batch Normalization 和 Dropout。

如果模型中有BN层(Batch Normalization）和Dropout，需要在训练时添加model.train()。model.train()是保证BN层能够用到每一批数据的均值和方差。对于Dropout，model.train()是随机取一部分网络连接来训练更新参数。

2. model.eval()

model.eval()的作用是不启用 Batch Normalization 和 Dropout。

如果模型中有BN层(Batch Normalization）和Dropout，在测试时添加model.eval()。model.eval()是保证BN层能够用全部训练数据的均值和方差，即测试过程中要保证BN层的均值和方差不变。对于Dropout，model.eval()是利用到了所有网络连接，即不进行随机舍弃神经元。

训练完train样本后，生成的模型model要用来测试样本。在model(test)之前，需要加上model.eval()，否则的话，有输入数据，即使不训练，它也会改变权值。这是model中含有BN层和Dropout所带来的的性质。

epochs     = 10
train_loss = []
train_acc  = []
test_loss  = []
test_acc   = []

for epoch in range(epochs):
    model.train()
    epoch_train_acc, epoch_train_loss = train(train_dl, model, loss_fn, opt)
    
    model.eval()
    epoch_test_acc, epoch_test_loss = test(test_dl, model, loss_fn)
    
    train_acc.append(epoch_train_acc)
    train_loss.append(epoch_train_loss)
    test_acc.append(epoch_test_acc)
    test_loss.append(epoch_test_loss)
    
    template = ('Epoch:{:2d}, Train_acc:{:.1f}%, Train_loss:{:.3f}, Test_acc:{:.1f}%，Test_loss:{:.3f}')
    print(template.format(epoch+1, epoch_train_acc*100, epoch_train_loss, epoch_test_acc*100, epoch_test_loss))
print('Done')

结果可视化

epochs_range = range(epochs)  # Define the range of epochs for plotting

plt.figure(figsize=(12, 3))  # Set the figure size (width=12, height=3)

# Plot Training and Validation Accuracy
plt.subplot(1, 2, 1)  # Create a subplot in a 1x2 grid (1 row, 2 columns), position 1
plt.plot(epochs_range, train_acc, label='Training Accuracy')  # Plot training accuracy
plt.plot(epochs_range, test_acc, label='Test Accuracy')      # Plot test accuracy
plt.legend(loc='lower right')  # Add a legend in the lower-right corner
plt.title('Training and Validation Accuracy')  # Add a title

# Plot Training and Validation Loss
plt.subplot(1, 2, 2)  # Create a subplot in a 1x2 grid, position 2
plt.plot(epochs_range, train_loss, label='Training Loss')  # Plot training loss
plt.plot(epochs_range, test_loss, label='Test Loss')      # Plot test loss
plt.legend(loc='upper right')  # Add a legend in the upper-right corner
plt.title('Training and Validation Loss')  # Add a title

plt.show()  # Display the plots