预训练模型+CBAM模块
- resnet结构解析
- CBAM放置位置的思考
- 针对预训练模型的训练策略
- 差异化学习率
- 三阶段微调
cbam定义与预处理
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
import numpy as np
# 定义通道注意力
class ChannelAttention(nn.Module):
def __init__(self, in_channels, ratio=16):
"""
通道注意力机制初始化
参数:
in_channels: 输入特征图的通道数
ratio: 降维比例，用于减少参数量，默认为16
"""
super().__init__()
# 全局平均池化，将每个通道的特征图压缩为1x1，保留通道间的平均值信息
self.avg_pool = nn.AdaptiveAvgPool2d(1)
# 全局最大池化，将每个通道的特征图压缩为1x1，保留通道间的最显著特征
self.max_pool = nn.AdaptiveMaxPool2d(1)
# 共享全连接层，用于学习通道间的关系
# 先降维（除以ratio），再通过ReLU激活，最后升维回原始通道数
self.fc = nn.Sequential(
nn.Linear(in_channels, in_channels // ratio, bias=False), # 降维层
nn.ReLU(), # 非线性激活函数
nn.Linear(in_channels // ratio, in_channels, bias=False) # 升维层
)
# Sigmoid函数将输出映射到0-1之间，作为各通道的权重
self.sigmoid = nn.Sigmoid()
def forward(self, x):
"""
前向传播函数
参数:
x: 输入特征图，形状为 [batch_size, channels, height, width]
返回:
调整后的特征图，通道权重已应用
"""
# 获取输入特征图的维度信息，这是一种元组的解包写法
b, c, h, w = x.shape
# 对平均池化结果进行处理：展平后通过全连接网络
avg_out = self.fc(self.avg_pool(x).view(b, c))
# 对最大池化结果进行处理：展平后通过全连接网络
max_out = self.fc(self.max_pool(x).view(b, c))
# 将平均池化和最大池化的结果相加并通过sigmoid函数得到通道权重
attention = self.sigmoid(avg_out + max_out).view(b, c, 1, 1)
# 将注意力权重与原始特征相乘，增强重要通道，抑制不重要通道
return x * attention #这个运算是pytorch的广播机制
## 空间注意力模块
class SpatialAttention(nn.Module):
def __init__(self, kernel_size=7):
super().__init__()
self.conv = nn.Conv2d(2, 1, kernel_size, padding=kernel_size//2, bias=False)
self.sigmoid = nn.Sigmoid()
def forward(self, x):
# 通道维度池化
avg_out = torch.mean(x, dim=1, keepdim=True) # 平均池化：(B,1,H,W)
max_out, _ = torch.max(x, dim=1, keepdim=True) # 最大池化：(B,1,H,W)
pool_out = torch.cat([avg_out, max_out], dim=1) # 拼接：(B,2,H,W)
attention = self.conv(pool_out) # 卷积提取空间特征
return x * self.sigmoid(attention) # 特征与空间权重相乘
## CBAM模块
class CBAM(nn.Module):
def __init__(self, in_channels, ratio=16, kernel_size=7):
super().__init__()
self.channel_attn = ChannelAttention(in_channels, ratio)
self.spatial_attn = SpatialAttention(kernel_size)
def forward(self, x):
x = self.channel_attn(x)
x = self.spatial_attn(x)
return x
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
import numpy as np
# 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}")
# 数据预处理（与原代码一致）
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))
])
# 加载数据集（与原代码一致）
train_dataset = datasets.CIFAR10(root='./data', train=True, download=True, transform=train_transform)
test_dataset = datasets.CIFAR10(root='./data', train=False, transform=test_transform)
train_loader = DataLoader(train_dataset, batch_size=64, shuffle=True)
test_loader = DataLoader(test_dataset, batch_size=64, shuffle=False)
import torch
import torchvision.models as models
from torchinfo import summary #之前的内容说了，推荐用他来可视化模型结构，信息最全
# 加载 ResNet18（预训练）
model = models.resnet18(pretrained=True)
model.eval()
# 输出模型结构和参数概要
summary(model, input_size=(1, 3, 224, 224))
python
==========================================================================================
Layer (type:depth-idx) Output Shape Param #
==========================================================================================
ResNet [1, 1000] --
├─Conv2d: 1-1 [1, 64, 112, 112] 9,408
├─BatchNorm2d: 1-2 [1, 64, 112, 112] 128
├─ReLU: 1-3 [1, 64, 112, 112] --
├─MaxPool2d: 1-4 [1, 64, 56, 56] --
├─Sequential: 1-5 [1, 64, 56, 56] --
│ └─BasicBlock: 2-1 [1, 64, 56, 56] --
│ │ └─Conv2d: 3-1 [1, 64, 56, 56] 36,864
│ │ └─BatchNorm2d: 3-2 [1, 64, 56, 56] 128
│ │ └─ReLU: 3-3 [1, 64, 56, 56] --
│ │ └─Conv2d: 3-4 [1, 64, 56, 56] 36,864
│ │ └─BatchNorm2d: 3-5 [1, 64, 56, 56] 128
│ │ └─ReLU: 3-6 [1, 64, 56, 56] --
│ └─BasicBlock: 2-2 [1, 64, 56, 56] --
│ │ └─Conv2d: 3-7 [1, 64, 56, 56] 36,864
│ │ └─BatchNorm2d: 3-8 [1, 64, 56, 56] 128
│ │ └─ReLU: 3-9 [1, 64, 56, 56] --
│ │ └─Conv2d: 3-10 [1, 64, 56, 56] 36,864
│ │ └─BatchNorm2d: 3-11 [1, 64, 56, 56] 128
│ │ └─ReLU: 3-12 [1, 64, 56, 56] --
├─Sequential: 1-6 [1, 128, 28, 28] --
│ └─BasicBlock: 2-3 [1, 128, 28, 28] --
│ │ └─Conv2d: 3-13 [1, 128, 28, 28] 73,728
│ │ └─BatchNorm2d: 3-14 [1, 128, 28, 28] 256
│ │ └─ReLU: 3-15 [1, 128, 28, 28] --
│ │ └─Conv2d: 3-16 [1, 128, 28, 28] 147,456
│ │ └─BatchNorm2d: 3-17 [1, 128, 28, 28] 256
│ │ └─Sequential: 3-18 [1, 128, 28, 28] 8,448
│ │ └─ReLU: 3-19 [1, 128, 28, 28] --
│ └─BasicBlock: 2-4 [1, 128, 28, 28] --
│ │ └─Conv2d: 3-20 [1, 128, 28, 28] 147,456
│ │ └─BatchNorm2d: 3-21 [1, 128, 28, 28] 256
│ │ └─ReLU: 3-22 [1, 128, 28, 28] --
│ │ └─Conv2d: 3-23 [1, 128, 28, 28] 147,456
│ │ └─BatchNorm2d: 3-24 [1, 128, 28, 28] 256
│ │ └─ReLU: 3-25 [1, 128, 28, 28] --
├─Sequential: 1-7 [1, 256, 14, 14] --
│ └─BasicBlock: 2-5 [1, 256, 14, 14] --
│ │ └─Conv2d: 3-26 [1, 256, 14, 14] 294,912
│ │ └─BatchNorm2d: 3-27 [1, 256, 14, 14] 512
│ │ └─ReLU: 3-28 [1, 256, 14, 14] --
│ │ └─Conv2d: 3-29 [1, 256, 14, 14] 589,824
│ │ └─BatchNorm2d: 3-30 [1, 256, 14, 14] 512
│ │ └─Sequential: 3-31 [1, 256, 14, 14] 33,280
│ │ └─ReLU: 3-32 [1, 256, 14, 14] --
│ └─BasicBlock: 2-6 [1, 256, 14, 14] --
│ │ └─Conv2d: 3-33 [1, 256, 14, 14] 589,824
│ │ └─BatchNorm2d: 3-34 [1, 256, 14, 14] 512
│ │ └─ReLU: 3-35 [1, 256, 14, 14] --
│ │ └─Conv2d: 3-36 [1, 256, 14, 14] 589,824
│ │ └─BatchNorm2d: 3-37 [1, 256, 14, 14] 512
│ │ └─ReLU: 3-38 [1, 256, 14, 14] --
├─Sequential: 1-8 [1, 512, 7, 7] --
│ └─BasicBlock: 2-7 [1, 512, 7, 7] --
│ │ └─Conv2d: 3-39 [1, 512, 7, 7] 1,179,648
│ │ └─BatchNorm2d: 3-40 [1, 512, 7, 7] 1,024
│ │ └─ReLU: 3-41 [1, 512, 7, 7] --
│ │ └─Conv2d: 3-42 [1, 512, 7, 7] 2,359,296
│ │ └─BatchNorm2d: 3-43 [1, 512, 7, 7] 1,024
│ │ └─Sequential: 3-44 [1, 512, 7, 7] 132,096
│ │ └─ReLU: 3-45 [1, 512, 7, 7] --
│ └─BasicBlock: 2-8 [1, 512, 7, 7] --
│ │ └─Conv2d: 3-46 [1, 512, 7, 7] 2,359,296
│ │ └─BatchNorm2d: 3-47 [1, 512, 7, 7] 1,024
│ │ └─ReLU: 3-48 [1, 512, 7, 7] --
│ │ └─Conv2d: 3-49 [1, 512, 7, 7] 2,359,296
│ │ └─BatchNorm2d: 3-50 [1, 512, 7, 7] 1,024
│ │ └─ReLU: 3-51 [1, 512, 7, 7] --
├─AdaptiveAvgPool2d: 1-9 [1, 512, 1, 1] --
├─Linear: 1-10 [1, 1000] 513,000
==========================================================================================
Total params: 11,689,512
Trainable params: 11,689,512
Non-trainable params: 0
Total mult-adds (G): 1.81
==========================================================================================
Input size (MB): 0.60
Forward/backward pass size (MB): 39.75
Params size (MB): 46.76
Estimated Total Size (MB): 87.11
==========================================================================================所以完全可以在不破坏其核心结构的情况下，将CBAM模块无缝地"注入"到预训练的ResNet中。这样做的逻辑是：
-
保留原始结构：原始的残差块负责提取核心特征。
-
增强特征：紧随其后的CBAM模块对这些提取出的特征进行"精炼"，告诉模型应该"关注什么"（what - 通道注意力）和"在哪里关注"（where - 空间注意力）。
-
不破坏预训练权重：原始残差块的预训练权重得以完整保留，我们只是在其后增加了一个新的、需要从头学习的模块。
自定义ResNet18模型
python
import torch
import torch.nn as nn
from torchvision import models
# 自定义ResNet18模型，插入CBAM模块
class ResNet18_CBAM(nn.Module):
def __init__(self, num_classes=10, pretrained=True, cbam_ratio=16, cbam_kernel=7):
super().__init__()
# 加载预训练ResNet18
self.backbone = models.resnet18(pretrained=pretrained)
# 修改首层卷积以适应32x32输入（CIFAR10）
self.backbone.conv1 = nn.Conv2d(
in_channels=3, out_channels=64, kernel_size=3, stride=1, padding=1, bias=False
)
self.backbone.maxpool = nn.Identity() # 移除原始MaxPool层（因输入尺寸小）
# 在每个残差块组后添加CBAM模块
self.cbam_layer1 = CBAM(in_channels=64, ratio=cbam_ratio, kernel_size=cbam_kernel)
self.cbam_layer2 = CBAM(in_channels=128, ratio=cbam_ratio, kernel_size=cbam_kernel)
self.cbam_layer3 = CBAM(in_channels=256, ratio=cbam_ratio, kernel_size=cbam_kernel)
self.cbam_layer4 = CBAM(in_channels=512, ratio=cbam_ratio, kernel_size=cbam_kernel)
# 修改分类头
self.backbone.fc = nn.Linear(in_features=512, out_features=num_classes)
def forward(self, x):
# 主干特征提取
x = self.backbone.conv1(x)
x = self.backbone.bn1(x)
x = self.backbone.relu(x) # [B, 64, 32, 32]
# 第一层残差块 + CBAM
x = self.backbone.layer1(x) # [B, 64, 32, 32]
x = self.cbam_layer1(x)
# 第二层残差块 + CBAM
x = self.backbone.layer2(x) # [B, 128, 16, 16]
x = self.cbam_layer2(x)
# 第三层残差块 + CBAM
x = self.backbone.layer3(x) # [B, 256, 8, 8]
x = self.cbam_layer3(x)
# 第四层残差块 + CBAM
x = self.backbone.layer4(x) # [B, 512, 4, 4]
x = self.cbam_layer4(x)
# 全局平均池化 + 分类
x = self.backbone.avgpool(x) # [B, 512, 1, 1]
x = torch.flatten(x, 1) # [B, 512]
x = self.backbone.fc(x) # [B, 10]
return x
# 初始化模型并移至设备
model = ResNet18_CBAM().to(device)
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=0.001)
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='min', patience=3, factor=0.5)
python
===============================================================================================
Layer (type:depth-idx) Output Shape Param #
===============================================================================================
ResNet18_CBAM [1, 10] --
├─ResNet: 1-9 -- (recursive)
│ └─Conv2d: 2-1 [1, 64, 32, 32] 1,728
│ └─BatchNorm2d: 2-2 [1, 64, 32, 32] 128
│ └─ReLU: 2-3 [1, 64, 32, 32] --
│ └─Sequential: 2-4 [1, 64, 32, 32] --
│ │ └─BasicBlock: 3-1 [1, 64, 32, 32] 73,984
│ │ └─BasicBlock: 3-2 [1, 64, 32, 32] 73,984
├─CBAM: 1-2 [1, 64, 32, 32] --
│ └─ChannelAttention: 2-5 [1, 64, 32, 32] --
│ │ └─AdaptiveAvgPool2d: 3-3 [1, 64, 1, 1] --
│ │ └─Sequential: 3-4 [1, 64] 512
│ │ └─AdaptiveMaxPool2d: 3-5 [1, 64, 1, 1] --
│ │ └─Sequential: 3-6 [1, 64] (recursive)
│ │ └─Sigmoid: 3-7 [1, 64] --
│ └─SpatialAttention: 2-6 [1, 64, 32, 32] --
│ │ └─Conv2d: 3-8 [1, 1, 32, 32] 98
│ │ └─Sigmoid: 3-9 [1, 1, 32, 32] --
├─ResNet: 1-9 -- (recursive)
│ └─Sequential: 2-7 [1, 128, 16, 16] --
│ │ └─BasicBlock: 3-10 [1, 128, 16, 16] 230,144
│ │ └─BasicBlock: 3-11 [1, 128, 16, 16] 295,424
├─CBAM: 1-4 [1, 128, 16, 16] --
│ └─ChannelAttention: 2-8 [1, 128, 16, 16] --
│ │ └─AdaptiveAvgPool2d: 3-12 [1, 128, 1, 1] --
│ │ └─Sequential: 3-13 [1, 128] 2,048
│ │ └─AdaptiveMaxPool2d: 3-14 [1, 128, 1, 1] --
│ │ └─Sequential: 3-15 [1, 128] (recursive)
│ │ └─Sigmoid: 3-16 [1, 128] --
│ └─SpatialAttention: 2-9 [1, 128, 16, 16] --
│ │ └─Conv2d: 3-17 [1, 1, 16, 16] 98
│ │ └─Sigmoid: 3-18 [1, 1, 16, 16] --
├─ResNet: 1-9 -- (recursive)
│ └─Sequential: 2-10 [1, 256, 8, 8] --
│ │ └─BasicBlock: 3-19 [1, 256, 8, 8] 919,040
│ │ └─BasicBlock: 3-20 [1, 256, 8, 8] 1,180,672
├─CBAM: 1-6 [1, 256, 8, 8] --
│ └─ChannelAttention: 2-11 [1, 256, 8, 8] --
│ │ └─AdaptiveAvgPool2d: 3-21 [1, 256, 1, 1] --
│ │ └─Sequential: 3-22 [1, 256] 8,192
│ │ └─AdaptiveMaxPool2d: 3-23 [1, 256, 1, 1] --
│ │ └─Sequential: 3-24 [1, 256] (recursive)
│ │ └─Sigmoid: 3-25 [1, 256] --
│ └─SpatialAttention: 2-12 [1, 256, 8, 8] --
│ │ └─Conv2d: 3-26 [1, 1, 8, 8] 98
│ │ └─Sigmoid: 3-27 [1, 1, 8, 8] --
├─ResNet: 1-9 -- (recursive)
│ └─Sequential: 2-13 [1, 512, 4, 4] --
│ │ └─BasicBlock: 3-28 [1, 512, 4, 4] 3,673,088
│ │ └─BasicBlock: 3-29 [1, 512, 4, 4] 4,720,640
├─CBAM: 1-8 [1, 512, 4, 4] --
│ └─ChannelAttention: 2-14 [1, 512, 4, 4] --
│ │ └─AdaptiveAvgPool2d: 3-30 [1, 512, 1, 1] --
│ │ └─Sequential: 3-31 [1, 512] 32,768
│ │ └─AdaptiveMaxPool2d: 3-32 [1, 512, 1, 1] --
│ │ └─Sequential: 3-33 [1, 512] (recursive)
│ │ └─Sigmoid: 3-34 [1, 512] --
│ └─SpatialAttention: 2-15 [1, 512, 4, 4] --
│ │ └─Conv2d: 3-35 [1, 1, 4, 4] 98
│ │ └─Sigmoid: 3-36 [1, 1, 4, 4] --
├─ResNet: 1-9 -- (recursive)
│ └─AdaptiveAvgPool2d: 2-16 [1, 512, 1, 1] --
│ └─Linear: 2-17 [1, 10] 5,130
===============================================================================================
Total params: 11,217,874
Trainable params: 11,217,874
Non-trainable params: 0
Total mult-adds (M): 555.65
===============================================================================================
Input size (MB): 0.01
Forward/backward pass size (MB): 9.86
Params size (MB): 44.87
Estimated Total Size (MB): 54.74
===============================================================================================可以看到cbam加在了每个layer的后面
模型训练加评估
三阶段式解冻与微调 (Progressive Unfreezing)
- 阶段一 (Epoch 1-5): 预热"实习生"
解冻部分: 仅解冻分类头 (`fc`) 和所有 `CBAM` 模块。
冻结部分: 冻结 ResNet18 的所有主干卷积层 (`conv1`, `bn1`, `layer1` 至 `layer4`)。
目标: 先利用强大的预训练特征，让模型快速学习新任务的分类边界，同时让注意力模块找到初步的关注点。
学习率: `1e-3` (使用较高学习率加速收敛)。
- 阶段二 (Epoch 6-20): 唤醒"高层专家"
解冻部分: 在上一阶段的基础上，额外解冻高层语义相关的卷积层 (`layer3`, `layer4`)。
冻结部分: 底层特征提取层 (`conv1`, `bn1`, `layer1`, `layer2`) 仍然冻结。
目标: 释放模型的高层特征提取能力，使其适应新任务的抽象概念 (例如"鸟的轮廓"比"一条边"更抽象)。
学习率: `1e-4` (降低学习率，避免新解冻的层因梯度过大而破坏其宝贵的预训练权重)。
- 阶段三 (Epoch 21-50): 全员协同微调
解冻部分: 解冻模型的所有层，进行端到端微调。
冻结部分: 无。
目标: 让模型的底层特征 (如边缘、纹理) 也与新任务进行对齐，做最后的精细化调整，提升整体性能。
学习率: `1e-5` (使用最低的学习率，在整个模型上缓慢、稳定地进行全局优化)。
python
import time
# ======================================================================
# 4. 结合了分阶段策略和详细打印的训练函数
# ======================================================================
def set_trainable_layers(model, trainable_parts):
print(f"\n---> 解冻以下部分并设为可训练: {trainable_parts}")
for name, param in model.named_parameters():
param.requires_grad = False
for part in trainable_parts:
if part in name:
param.requires_grad = True
break
def train_staged_finetuning(model, criterion, train_loader, test_loader, device, epochs):
optimizer = None
# 初始化历史记录列表，与你的要求一致
all_iter_losses, iter_indices = [], []
train_acc_history, test_acc_history = [], []
train_loss_history, test_loss_history = [], []
for epoch in range(1, epochs + 1):
epoch_start_time = time.time()
# --- 动态调整学习率和冻结层 ---
if epoch == 1:
print("\n" + "="*50 + "\n🚀 **阶段 1：训练注意力模块和分类头**\n" + "="*50)
set_trainable_layers(model, ["cbam", "backbone.fc"])
optimizer = optim.Adam(filter(lambda p: p.requires_grad, model.parameters()), lr=1e-3)
elif epoch == 6:
print("\n" + "="*50 + "\n✈️ **阶段 2：解冻高层卷积层 (layer3, layer4)**\n" + "="*50)
set_trainable_layers(model, ["cbam", "backbone.fc", "backbone.layer3", "backbone.layer4"])
optimizer = optim.Adam(filter(lambda p: p.requires_grad, model.parameters()), lr=1e-4)
elif epoch == 21:
print("\n" + "="*50 + "\n🛰️ **阶段 3：解冻所有层，进行全局微调**\n" + "="*50)
for param in model.parameters(): param.requires_grad = True
optimizer = optim.Adam(model.parameters(), lr=1e-5)
# --- 训练循环 ---
model.train()
running_loss, correct, total = 0.0, 0, 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()
# 记录每个iteration的损失
iter_loss = loss.item()
all_iter_losses.append(iter_loss)
iter_indices.append((epoch - 1) * len(train_loader) + batch_idx + 1)
running_loss += iter_loss
_, predicted = output.max(1)
total += target.size(0)
correct += predicted.eq(target).sum().item()
# 按你的要求，每100个batch打印一次
if (batch_idx + 1) % 100 == 0:
print(f'Epoch: {epoch}/{epochs} | Batch: {batch_idx+1}/{len(train_loader)} '
f'| 单Batch损失: {iter_loss:.4f} | 累计平均损失: {running_loss/(batch_idx+1):.4f}')
epoch_train_loss = running_loss / len(train_loader)
epoch_train_acc = 100. * correct / total
train_loss_history.append(epoch_train_loss)
train_acc_history.append(epoch_train_acc)
# --- 测试循环 ---
model.eval()
test_loss, correct_test, total_test = 0, 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
test_loss_history.append(epoch_test_loss)
test_acc_history.append(epoch_test_acc)
# 打印每个epoch的最终结果
print(f'Epoch {epoch}/{epochs} 完成 | 耗时: {time.time() - epoch_start_time:.2f}s | 训练准确率: {epoch_train_acc:.2f}% | 测试准确率: {epoch_test_acc:.2f}%')
# 训练结束后调用绘图函数
print("\n训练完成! 开始绘制结果图表...")
plot_iter_losses(all_iter_losses, iter_indices)
plot_epoch_metrics(train_acc_history, test_acc_history, train_loss_history, test_loss_history)
# 返回最终的测试准确率
return epoch_test_acc
# ======================================================================
# 5. 绘图函数定义
# ======================================================================
def plot_iter_losses(losses, indices):
plt.figure(figsize=(10, 4))
plt.plot(indices, losses, 'b-', alpha=0.7, label='Iteration Loss')
plt.xlabel('Iteration（Batch序号）')
plt.ylabel('损失值')
plt.title('每个 Iteration 的训练损失')
plt.legend()
plt.grid(True)
plt.tight_layout()
plt.show()
def plot_epoch_metrics(train_acc, test_acc, train_loss, test_loss):
epochs = range(1, len(train_acc) + 1)
plt.figure(figsize=(12, 4))
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('训练和测试准确率')
plt.legend(); plt.grid(True)
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('训练和测试损失')
plt.legend(); plt.grid(True)
plt.tight_layout()
plt.show()
# ======================================================================
# 6. 执行训练
# ======================================================================
model = ResNet18_CBAM().to(device)
criterion = nn.CrossEntropyLoss()
epochs = 50
print("开始使用带分阶段微调策略的ResNet18+CBAM模型进行训练...")
final_accuracy = train_staged_finetuning(model, criterion, train_loader, test_loader, device, epochs)
print(f"训练完成！最终测试准确率: {final_accuracy:.2f}%")
# torch.save(model.state_dict(), 'resnet18_cbam_finetuned.pth')
# print("模型已保存为: resnet18_cbam_finetuned.pth")
@浙大疏锦行