DAY 44 预训练模型
知识点回顾：
- 预训练的概念
- 常见的分类预训练模型
- 图像预训练模型的发展史
- 预训练的策略
- 预训练代码实战：resnet18
作业：
- 尝试在cifar10对比如下其他的预训练模型，观察差异，尽可能和他人选择的不同
- 尝试通过ctrl进入resnet的内部，观察残差究竟是什么
python
import torch
import torch.nn as nn
import torch.optim as optim
from torchvision import datasets, transforms
from torch.utils.data import DataLoader
import matplotlib.pyplot as plt
# 设置中文字体支持
plt.rcParams["font.family"] = ["SimHei"]
plt.rcParams['axes.unicode_minus'] = False # 解决负号显示问题
# 检查GPU是否可用
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"使用设备: {device}")
# 1. 数据预处理（训练集增强，测试集标准化）
train_transform = transforms.Compose([
transforms.RandomCrop(32, padding=4),
transforms.RandomHorizontalFlip(),
transforms.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2, hue=0.1),
transforms.RandomRotation(15),
transforms.ToTensor(),
transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010))
])
test_transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010))
])
# 2. 加载CIFAR-10数据集
train_dataset = datasets.CIFAR10(
root='./data',
train=True,
download=True,
transform=train_transform
)
test_dataset = datasets.CIFAR10(
root='./data',
train=False,
transform=test_transform
)
# 3. 创建数据加载器（可调整batch_size）
batch_size = 64
train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True)
test_loader = DataLoader(test_dataset, batch_size=batch_size, shuffle=False)
# 4. 训练函数（支持学习率调度器）
def train(model, train_loader, test_loader, criterion, optimizer, scheduler, device, epochs):
model.train() # 设置为训练模式
train_loss_history = []
test_loss_history = []
train_acc_history = []
test_acc_history = []
all_iter_losses = []
iter_indices = []
for epoch in range(epochs):
running_loss = 0.0
correct_train = 0
total_train = 0
for batch_idx, (data, target) in enumerate(train_loader):
data, target = data.to(device), target.to(device)
optimizer.zero_grad()
output = model(data)
loss = criterion(output, target)
loss.backward()
optimizer.step()
# 记录Iteration损失
iter_loss = loss.item()
all_iter_losses.append(iter_loss)
iter_indices.append(epoch * len(train_loader) + batch_idx + 1)
# 统计训练指标
running_loss += iter_loss
_, predicted = output.max(1)
total_train += target.size(0)
correct_train += predicted.eq(target).sum().item()
# 每100批次打印进度
if (batch_idx + 1) % 100 == 0:
print(f"Epoch {epoch+1}/{epochs} | Batch {batch_idx+1}/{len(train_loader)} "
f"| 单Batch损失: {iter_loss:.4f}")
# 计算 epoch 级指标
epoch_train_loss = running_loss / len(train_loader)
epoch_train_acc = 100. * correct_train / total_train
# 测试阶段
model.eval()
correct_test = 0
total_test = 0
test_loss = 0.0
with torch.no_grad():
for data, target in test_loader:
data, target = data.to(device), target.to(device)
output = model(data)
test_loss += criterion(output, target).item()
_, predicted = output.max(1)
total_test += target.size(0)
correct_test += predicted.eq(target).sum().item()
epoch_test_loss = test_loss / len(test_loader)
epoch_test_acc = 100. * correct_test / total_test
# 记录历史数据
train_loss_history.append(epoch_train_loss)
test_loss_history.append(epoch_test_loss)
train_acc_history.append(epoch_train_acc)
test_acc_history.append(epoch_test_acc)
# 更新学习率调度器
if scheduler is not None:
scheduler.step(epoch_test_loss)
# 打印 epoch 结果
print(f"Epoch {epoch+1} 完成 | 训练损失: {epoch_train_loss:.4f} "
f"| 训练准确率: {epoch_train_acc:.2f}% | 测试准确率: {epoch_test_acc:.2f}%")
# 绘制损失和准确率曲线
plot_iter_losses(all_iter_losses, iter_indices)
plot_epoch_metrics(train_acc_history, test_acc_history, train_loss_history, test_loss_history)
return epoch_test_acc # 返回最终测试准确率
# 5. 绘制Iteration损失曲线
def plot_iter_losses(losses, indices):
plt.figure(figsize=(10, 4))
plt.plot(indices, losses, 'b-', alpha=0.7)
plt.xlabel('Iteration（Batch序号）')
plt.ylabel('损失值')
plt.title('训练过程中的Iteration损失变化')
plt.grid(True)
plt.show()
# 6. 绘制Epoch级指标曲线
def plot_epoch_metrics(train_acc, test_acc, train_loss, test_loss):
epochs = range(1, len(train_acc) + 1)
plt.figure(figsize=(12, 5))
# 准确率曲线
plt.subplot(1, 2, 1)
plt.plot(epochs, train_acc, 'b-', label='训练准确率')
plt.plot(epochs, test_acc, 'r-', label='测试准确率')
plt.xlabel('Epoch')
plt.ylabel('准确率 (%)')
plt.title('准确率随Epoch变化')
plt.legend()
plt.grid(True)
# 损失曲线
plt.subplot(1, 2, 2)
plt.plot(epochs, train_loss, 'b-', label='训练损失')
plt.plot(epochs, test_loss, 'r-', label='测试损失')
plt.xlabel('Epoch')
plt.ylabel('损失值')
plt.title('损失值随Epoch变化')
plt.legend()
plt.grid(True)
plt.tight_layout()
plt.show()
# 导入ResNet模型
from torchvision.models import resnet18
# 定义ResNet18模型（支持预训练权重加载）
def create_resnet18(pretrained=True, num_classes=10):
# 加载预训练模型（ImageNet权重）
model = resnet18(pretrained=pretrained)
# 修改最后一层全连接层，适配CIFAR-10的10分类任务
in_features = model.fc.in_features
model.fc = nn.Linear(in_features, num_classes)
# 将模型转移到指定设备（CPU/GPU）
model = model.to(device)
return model
# 创建ResNet18模型（加载ImageNet预训练权重，不进行微调）
model = create_resnet18(pretrained=True, num_classes=10)
model.eval() # 设置为推理模式
# 测试单张图片（示例）
from torchvision import utils
# 从测试数据集中获取一张图片
dataiter = iter(test_loader)
images, labels = dataiter.next()
images = images[:1].to(device) # 取第1张图片
# 前向传播
with torch.no_grad():
outputs = model(images)
_, predicted = torch.max(outputs.data, 1)
# 显示图片和预测结果
plt.imshow(utils.make_grid(images.cpu(), normalize=True).permute(1, 2, 0))
plt.title(f"预测类别: {predicted.item()}")
plt.axis('off')
plt.show()
import torch
import torch.nn as nn
import torch.optim as optim
from torchvision import datasets, transforms, models
from torch.utils.data import DataLoader
import matplotlib.pyplot as plt
import os
# 设置中文字体支持
plt.rcParams["font.family"] = ["SimHei"]
plt.rcParams['axes.unicode_minus'] = False # 解决负号显示问题
# 检查GPU是否可用
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"使用设备: {device}")
# 1. 数据预处理（训练集增强，测试集标准化）
train_transform = transforms.Compose([
transforms.RandomCrop(32, padding=4),
transforms.RandomHorizontalFlip(),
transforms.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2, hue=0.1),
transforms.RandomRotation(15),
transforms.ToTensor(),
transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010))
])
test_transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010))
])
# 2. 加载CIFAR-10数据集
train_dataset = datasets.CIFAR10(
root='./data',
train=True,
download=True,
transform=train_transform
)
test_dataset = datasets.CIFAR10(
root='./data',
train=False,
transform=test_transform
)
# 3. 创建数据加载器
batch_size = 64
train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True)
test_loader = DataLoader(test_dataset, batch_size=batch_size, shuffle=False)
# 4. 定义ResNet18模型
def create_resnet18(pretrained=True, num_classes=10):
model = models.resnet18(pretrained=pretrained)
# 修改最后一层全连接层
in_features = model.fc.in_features
model.fc = nn.Linear(in_features, num_classes)
return model.to(device)
# 5. 冻结/解冻模型层的函数
def freeze_model(model, freeze=True):
"""冻结或解冻模型的卷积层参数"""
# 冻结/解冻除fc层外的所有参数
for name, param in model.named_parameters():
if 'fc' not in name:
param.requires_grad = not freeze
# 打印冻结状态
frozen_params = sum(p.numel() for p in model.parameters() if not p.requires_grad)
total_params = sum(p.numel() for p in model.parameters())
if freeze:
print(f"已冻结模型卷积层参数 ({frozen_params}/{total_params} 参数)")
else:
print(f"已解冻模型所有参数 ({total_params}/{total_params} 参数可训练)")
return model
# 6. 训练函数（支持阶段式训练）
def train_with_freeze_schedule(model, train_loader, test_loader, criterion, optimizer, scheduler, device, epochs, freeze_epochs=5):
"""
前freeze_epochs轮冻结卷积层，之后解冻所有层进行训练
"""
train_loss_history = []
test_loss_history = []
train_acc_history = []
test_acc_history = []
all_iter_losses = []
iter_indices = []
# 初始冻结卷积层
if freeze_epochs > 0:
model = freeze_model(model, freeze=True)
for epoch in range(epochs):
# 解冻控制：在指定轮次后解冻所有层
if epoch == freeze_epochs:
model = freeze_model(model, freeze=False)
# 解冻后调整优化器（可选）
optimizer.param_groups[0]['lr'] = 1e-4 # 降低学习率防止过拟合
model.train() # 设置为训练模式
running_loss = 0.0
correct_train = 0
total_train = 0
for batch_idx, (data, target) in enumerate(train_loader):
data, target = data.to(device), target.to(device)
optimizer.zero_grad()
output = model(data)
loss = criterion(output, target)
loss.backward()
optimizer.step()
# 记录Iteration损失
iter_loss = loss.item()
all_iter_losses.append(iter_loss)
iter_indices.append(epoch * len(train_loader) + batch_idx + 1)
# 统计训练指标
running_loss += iter_loss
_, predicted = output.max(1)
total_train += target.size(0)
correct_train += predicted.eq(target).sum().item()
# 每100批次打印进度
if (batch_idx + 1) % 100 == 0:
print(f"Epoch {epoch+1}/{epochs} | Batch {batch_idx+1}/{len(train_loader)} "
f"| 单Batch损失: {iter_loss:.4f}")
# 计算 epoch 级指标
epoch_train_loss = running_loss / len(train_loader)
epoch_train_acc = 100. * correct_train / total_train
# 测试阶段
model.eval()
correct_test = 0
total_test = 0
test_loss = 0.0
with torch.no_grad():
for data, target in test_loader:
data, target = data.to(device), target.to(device)
output = model(data)
test_loss += criterion(output, target).item()
_, predicted = output.max(1)
total_test += target.size(0)
correct_test += predicted.eq(target).sum().item()
epoch_test_loss = test_loss / len(test_loader)
epoch_test_acc = 100. * correct_test / total_test
# 记录历史数据
train_loss_history.append(epoch_train_loss)
test_loss_history.append(epoch_test_loss)
train_acc_history.append(epoch_train_acc)
test_acc_history.append(epoch_test_acc)
# 更新学习率调度器
if scheduler is not None:
scheduler.step(epoch_test_loss)
# 打印 epoch 结果
print(f"Epoch {epoch+1} 完成 | 训练损失: {epoch_train_loss:.4f} "
f"| 训练准确率: {epoch_train_acc:.2f}% | 测试准确率: {epoch_test_acc:.2f}%")
# 绘制损失和准确率曲线
plot_iter_losses(all_iter_losses, iter_indices)
plot_epoch_metrics(train_acc_history, test_acc_history, train_loss_history, test_loss_history)
return epoch_test_acc # 返回最终测试准确率
# 7. 绘制Iteration损失曲线
def plot_iter_losses(losses, indices):
plt.figure(figsize=(10, 4))
plt.plot(indices, losses, 'b-', alpha=0.7)
plt.xlabel('Iteration（Batch序号）')
plt.ylabel('损失值')
plt.title('训练过程中的Iteration损失变化')
plt.grid(True)
plt.show()
# 8. 绘制Epoch级指标曲线
def plot_epoch_metrics(train_acc, test_acc, train_loss, test_loss):
epochs = range(1, len(train_acc) + 1)
plt.figure(figsize=(12, 5))
# 准确率曲线
plt.subplot(1, 2, 1)
plt.plot(epochs, train_acc, 'b-', label='训练准确率')
plt.plot(epochs, test_acc, 'r-', label='测试准确率')
plt.xlabel('Epoch')
plt.ylabel('准确率 (%)')
plt.title('准确率随Epoch变化')
plt.legend()
plt.grid(True)
# 损失曲线
plt.subplot(1, 2, 2)
plt.plot(epochs, train_loss, 'b-', label='训练损失')
plt.plot(epochs, test_loss, 'r-', label='测试损失')
plt.xlabel('Epoch')
plt.ylabel('损失值')
plt.title('损失值随Epoch变化')
plt.legend()
plt.grid(True)
plt.tight_layout()
plt.show()
# 主函数：训练模型
def main():
# 参数设置
epochs = 40 # 总训练轮次
freeze_epochs = 5 # 冻结卷积层的轮次
learning_rate = 1e-3 # 初始学习率
weight_decay = 1e-4 # 权重衰减
# 创建ResNet18模型（加载预训练权重）
model = create_resnet18(pretrained=True, num_classes=10)
# 定义优化器和损失函数
optimizer = optim.Adam(model.parameters(), lr=learning_rate, weight_decay=weight_decay)
criterion = nn.CrossEntropyLoss()
# 定义学习率调度器
scheduler = optim.lr_scheduler.ReduceLROnPlateau(
optimizer, mode='min', factor=0.5, patience=2, verbose=True
)
# 开始训练（前5轮冻结卷积层，之后解冻）
final_accuracy = train_with_freeze_schedule(
model=model,
train_loader=train_loader,
test_loader=test_loader,
criterion=criterion,
optimizer=optimizer,
scheduler=scheduler,
device=device,
epochs=epochs,
freeze_epochs=freeze_epochs
)
print(f"训练完成！最终测试准确率: {final_accuracy:.2f}%")
# # 保存模型
# torch.save(model.state_dict(), 'resnet18_cifar10_finetuned.pth')
# print("模型已保存至: resnet18_cifar10_finetuned.pth")
if __name__ == "__main__":
main()
几个明显的现象
-
解冻后几个epoch即可达到之前cnn训练20轮的效果，这是预训练的优势
-
由于训练集用了 RandomCrop（随机裁剪）、RandomHorizontalFlip（随机水平翻转）、ColorJitter（颜色抖动）等数据增强操作，这会让训练时模型看到的图片有更多 "干扰" 或变形。比如一张汽车图片，训练时可能被裁剪成只显示局部、颜色也有变化，模型学习难度更高；而测试集是标准的、没增强的图片，模型预测相对轻松，就可能出现训练集准确率暂时低于测试集的情况，尤其在训练前期增强对模型影响更明显。随着训练推进，模型适应增强后会缓解。
-
最后收敛后的效果超过非预训练模型的80%,大幅提升