案例-图像分类

网络结构: 卷积+BN+激活+池化

数据集介绍

CIFAR-10数据集5万张训练图像、1万张测试图像、10个类别、每个类别有6k个图像，图像大小32×32×3。下图列举了10个类，每一类随机展示了10张图片：

特征图计算

在卷积层和池化层结束后, 将特征图变形成一行n列数据, 计算特征图进行变化, 映射到全连接层时输入层特征为最后一层卷积层经池化后的特征图各维度相乘

具体流程-# Acc: 0.728

# 导包
import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader
from torchsummary import summary
from torchvision.datasets import CIFAR10
from torchvision.transforms import ToTensor, Compose  # Compose: 数据增强(扩充数据集)
import time
import matplotlib.pyplot as plt
​
batch_size = 16
​
​
# 创建数据集
def create_dataset():torch.manual_seed(21)train = CIFAR10(root='data',train=True,transform=Compose([ToTensor()]))test = CIFAR10(root='data',train=False,transform=Compose([ToTensor()]))return train, test
​
​
# 创建模型
class ImgCls(nn.Module):# 定义网络结构def __init__(self):super(ImgCls, self).__init__()# 定义网络层：卷积层+池化层self.conv1 = nn.Conv2d(3, 16, stride=1, kernel_size=3)self.batch_norm_layer1 = nn.BatchNorm2d(num_features=16, eps=1e-5, momentum=0.1, affine=True, track_running_stats=True)self.pool1 = nn.MaxPool2d(kernel_size=2, stride=2)
​self.conv2 = nn.Conv2d(16, 32, stride=1, kernel_size=3)self.batch_norm_layer2 = nn.BatchNorm2d(num_features=32, eps=1e-5, momentum=0.1, affine=True, track_running_stats=True)self.pool2 = nn.MaxPool2d(kernel_size=2, stride=1)
​self.conv3 = nn.Conv2d(32, 64, stride=1, kernel_size=3)self.batch_norm_layer3 = nn.BatchNorm2d(num_features=64, eps=1e-5, momentum=0.1, affine=True, track_running_stats=True)self.pool3 = nn.MaxPool2d(kernel_size=2, stride=1)
​self.conv4 = nn.Conv2d(64, 128, stride=1, kernel_size=2)self.batch_norm_layer4 = nn.BatchNorm2d(num_features=128, eps=1e-5, momentum=0.1, affine=True, track_running_stats=True)self.pool4 = nn.MaxPool2d(kernel_size=2, stride=2)
​self.conv5 = nn.Conv2d(128, 256, stride=1, kernel_size=2)self.batch_norm_layer5 = nn.BatchNorm2d(num_features=256, eps=1e-5, momentum=0.1, affine=True, track_running_stats=True)self.pool5 = nn.MaxPool2d(kernel_size=2, stride=1)
​# 全连接层self.linear1 = nn.Linear(1024, 2048)self.linear2 = nn.Linear(2048, 1024)self.linear3 = nn.Linear(1024, 512)self.linear4 = nn.Linear(512, 256)self.linear5 = nn.Linear(256, 128)self.out = nn.Linear(128, 10)
​# 定义前向传播def forward(self, x):# 第1层: 卷积+BN+激活+池化x = self.conv1(x)x = self.batch_norm_layer1(x)x = torch.rrelu(x)x = self.pool1(x)
​# 第2层: 卷积+BN+激活+池化x = self.conv2(x)x = self.batch_norm_layer2(x)x = torch.rrelu(x)x = self.pool2(x)
​# 第3层: 卷积+BN+激活+池化x = self.conv3(x)x = self.batch_norm_layer3(x)x = torch.rrelu(x)x = self.pool3(x)
​# 第4层: 卷积+BN+激活+池化x = self.conv4(x)x = self.batch_norm_layer4(x)x = torch.rrelu(x)x = self.pool4(x)
​# 第5层: 卷积+BN+激活+池化x = self.conv5(x)x = self.batch_norm_layer5(x)x = torch.rrelu(x)x = self.pool5(x)
​# 将特征图做成以为向量的形式：相当于特征向量x = x.reshape(x.size(0), -1)  # 将3维特征图转化为1维向量(1, n)
​# 全连接层x = torch.rrelu(self.linear1(x))x = torch.rrelu(self.linear2(x))x = torch.rrelu(self.linear3(x))x = torch.rrelu(self.linear4(x))x = torch.rrelu(self.linear5(x))# 返回输出结果return self.out(x)
​
​
# 训练
def train(model, train_dataset, epochs):torch.manual_seed(21)loss = nn.CrossEntropyLoss()opt = optim.Adam(model.parameters(), lr=1e-4)for epoch in range(epochs):dataloader = DataLoader(train_dataset, shuffle=True, batch_size=batch_size)loss_total = 0iter = 0stat_time = time.time()for x, y in dataloader:output = model(x.to(device))loss_value = loss(output, y.to(device))opt.zero_grad()loss_value.backward()opt.step()loss_total += loss_value.item()iter += 1print(f'epoch:{epoch + 1:4d}, loss:{loss_total / iter:6.4f}, time:{time.time() - stat_time:.2f}s')torch.save(model.state_dict(), 'model/img_cls_model.pth')
​
​
# 测试
def test(valid_dataset, model, batch_size):# 构建数据加载器dataloader = DataLoader(valid_dataset, batch_size=batch_size, shuffle=False)
​# 计算精度total_correct = 0# 遍历每个batch的数据，获取预测结果，计算精度for x, y in dataloader:output = model(x.to(device))y_pred = torch.argmax(output, dim=-1)total_correct += (y_pred == y.to(device)).sum()# 打印精度print(f'Acc: {(total_correct.item() / len(valid_dataset))}')
​
​
if __name__ == '__main__':batch_size = 16device = torch.device("cuda" if torch.cuda.is_available() else "cpu")# 获取数据集train_data, test_data = create_dataset()
​# # 查看数据集# print(f'数据集类别: {train_data.class_to_idx}')# print(f'训练集: {train_data.data.shape}')# print(f'验证集: {test_data.data.shape}')# print(f'类别数量: {len(np.unique(train_data.targets))}')# # 展示图像# plt.figure(figsize=(8, 8))# plt.imshow(train_data.data[0])# plt.title(train_data.classes[train_data.targets[0]])# plt.show()
​# 实例化模型model = ImgCls().to(device)
​# 查看网络结构summary(model, (3, 32, 32), device='cuda', batch_size=batch_size)
​# 模型训练train(model, train_data, epochs=60)# 加载训练好的模型参数model.load_state_dict(torch.load('model/img_cls_model.pth'))model.eval()# 模型评估test(test_data, model, batch_size=16)   # Acc: 0.728
​

调整网络结构

第一次调整: 训练50轮, Acc: 0.71

第二次调整: 训练30轮, Acc:0.7351

第三次调整: batch_size=8, epoch=50 => Acc: 0.7644

# 导包
import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader
from torchsummary import summary
from torchvision.datasets import CIFAR10
from torchvision.transforms import ToTensor, Compose  # Compose: 数据增强(扩充数据集)
import time
import matplotlib.pyplot as plt
​
batch_size = 16
​
​
# 创建数据集
def create_dataset():torch.manual_seed(21)train = CIFAR10(root='data',train=True,transform=Compose([ToTensor()]))test = CIFAR10(root='data',train=False,transform=Compose([ToTensor()]))return train, test
​
​
# 创建模型
class ImgCls(nn.Module):# 定义网络结构def __init__(self):super(ImgCls, self).__init__()# 定义网络层：卷积层+池化层self.conv1 = nn.Conv2d(3, 16, stride=1, kernel_size=3, padding=1)self.batch_norm_layer1 = nn.BatchNorm2d(num_features=16, eps=1e-5, momentum=0.1, affine=True, track_running_stats=True)self.pool1 = nn.MaxPool2d(kernel_size=2, stride=2)
​self.conv2 = nn.Conv2d(16, 32, stride=1, kernel_size=3, padding=1)self.batch_norm_layer2 = nn.BatchNorm2d(num_features=32, eps=1e-5, momentum=0.1, affine=True, track_running_stats=True)self.pool2 = nn.MaxPool2d(kernel_size=2, stride=2)
​self.conv3 = nn.Conv2d(32, 64, stride=1, kernel_size=3, padding=1)self.batch_norm_layer3 = nn.BatchNorm2d(num_features=64, eps=1e-5, momentum=0.1, affine=True, track_running_stats=True)self.pool3 = nn.MaxPool2d(kernel_size=2, stride=1)
​self.conv4 = nn.Conv2d(64, 128, stride=1, kernel_size=3, padding=1)self.batch_norm_layer4 = nn.BatchNorm2d(num_features=128, eps=1e-5, momentum=0.1, affine=True, track_running_stats=True)self.pool4 = nn.MaxPool2d(kernel_size=2, stride=1)
​self.conv5 = nn.Conv2d(128, 256, stride=1, kernel_size=3)self.batch_norm_layer5 = nn.BatchNorm2d(num_features=256, eps=1e-5, momentum=0.1, affine=True, track_running_stats=True)self.pool5 = nn.MaxPool2d(kernel_size=2, stride=2)
​# 全连接层self.linear1 = nn.Linear(1024, 2048)self.linear2 = nn.Linear(2048, 1024)self.linear3 = nn.Linear(1024, 512)self.linear4 = nn.Linear(512, 256)self.linear5 = nn.Linear(256, 128)self.out = nn.Linear(128, 10)
​# 定义前向传播def forward(self, x):# 第1层: 卷积+BN+激活+池化x = self.conv1(x)x = self.batch_norm_layer1(x)x = torch.relu(x)x = self.pool1(x)
​# 第2层: 卷积+BN+激活+池化x = self.conv2(x)x = self.batch_norm_layer2(x)x = torch.relu(x)x = self.pool2(x)
​# 第3层: 卷积+BN+激活+池化x = self.conv3(x)x = self.batch_norm_layer3(x)x = torch.relu(x)x = self.pool3(x)
​# 第4层: 卷积+BN+激活+池化x = self.conv4(x)x = self.batch_norm_layer4(x)x = torch.relu(x)x = self.pool4(x)
​# 第5层: 卷积+BN+激活+池化x = self.conv5(x)x = self.batch_norm_layer5(x)x = torch.rrelu(x)x = self.pool5(x)
​# 将特征图做成以为向量的形式：相当于特征向量x = x.reshape(x.size(0), -1)  # 将3维特征图转化为1维向量(1, n)
​# 全连接层x = torch.relu(self.linear1(x))x = torch.relu(self.linear2(x))x = torch.relu(self.linear3(x))x = torch.relu(self.linear4(x))x = torch.rrelu(self.linear5(x))# 返回输出结果return self.out(x)
​
​
# 训练
def train(model, train_dataset, epochs):torch.manual_seed(21)loss = nn.CrossEntropyLoss()opt = optim.Adam(model.parameters(), lr=1e-4)for epoch in range(epochs):dataloader = DataLoader(train_dataset, shuffle=True, batch_size=batch_size)loss_total = 0iter = 0stat_time = time.time()for x, y in dataloader:output = model(x.to(device))loss_value = loss(output, y.to(device))opt.zero_grad()loss_value.backward()opt.step()loss_total += loss_value.item()iter += 1print(f'epoch:{epoch + 1:4d}, loss:{loss_total / iter:6.4f}, time:{time.time() - stat_time:.2f}s')torch.save(model.state_dict(), 'model/img_cls_model1.pth')
​
​
# 测试
def test(valid_dataset, model, batch_size):# 构建数据加载器dataloader = DataLoader(valid_dataset, batch_size=batch_size, shuffle=False)
​# 计算精度total_correct = 0# 遍历每个batch的数据，获取预测结果，计算精度for x, y in dataloader:output = model(x.to(device))y_pred = torch.argmax(output, dim=-1)total_correct += (y_pred == y.to(device)).sum()# 打印精度print(f'Acc: {(total_correct.item() / len(valid_dataset))}')
​
​
if __name__ == '__main__':batch_size = 8device = torch.device("cuda" if torch.cuda.is_available() else "cpu")# 获取数据集train_data, test_data = create_dataset()
​# # 查看数据集# print(f'数据集类别: {train_data.class_to_idx}')# print(f'训练集: {train_data.data.shape}')# print(f'验证集: {test_data.data.shape}')# print(f'类别数量: {len(np.unique(train_data.targets))}')# # 展示图像# plt.figure(figsize=(8, 8))# plt.imshow(train_data.data[0])# plt.title(train_data.classes[train_data.targets[0]])# plt.show()
​# 实例化模型model = ImgCls().to(device)
​# 查看网络结构summary(model, (3, 32, 32), device='cuda', batch_size=batch_size)
​# 模型训练train(model, train_data, epochs=50)# 加载训练好的模型参数model.load_state_dict(torch.load('model/img_cls_model1.pth', weights_only=True))model.eval()# 模型评估test(test_data, model, batch_size=16)   # Acc: 0.7644
​