当前位置：首页 > news >正文

python打卡训练营打卡记录day43

news 来源：原创 2025/6/4 13:11:19

复习日

作业：

kaggle找到一个图像数据集，用cnn网络进行训练并且用grad-cam做可视化

进阶：并拆分成多个文件

数据集来源：Flowers Recognition

选择该数据集原因：

中等规模：4242张图片 - 训练快速但足够展示效果
清晰类别：5类花朵（雏菊、蒲公英、玫瑰、向日葵、郁金香）
视觉特征明显：花朵在图像中的位置多变，Grad-CAM效果直观
高质量图片：分辨率适中（平均500×500像素）
简单结构：按类别分文件夹，无需复杂预处理

划分数据集

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
from PIL import Image
import os
from sklearn.model_selection import train_test_split
from shutil import copyfiledata_root = "flowers"  # 数据集根目录
classes = ["daisy", "tulip", "rose", "sunflower", "dandelion"]  for folder in ["train", "val", "test"]:os.makedirs(os.path.join(data_root, folder), exist_ok=True)# 数据集划分
for cls in classes:cls_path = os.path.join(data_root, cls)if not os.path.isdir(cls_path):raise FileNotFoundError(f"类别文件夹{cls}不存在！请检查数据集路径。")imgs = [f for f in os.listdir(cls_path) if f.lower().endswith((".jpg", ".jpeg", ".png"))]if not imgs:raise ValueError(f"类别{cls}中没有图片文件！")# 划分数据集（测试集20%，验证集20% of 剩余数据，训练集60%）train_val, test = train_test_split(imgs, test_size=0.2, random_state=42)train, val = train_test_split(train_val, test_size=0.25, random_state=42)  # 0.8*0.25=0.2（验证集占比）# 复制到train/val/test下的类别子文件夹（关键修正！）for split, imgs_list in zip(["train", "val", "test"], [train, val, test]):split_class_path = os.path.join(data_root, split, cls)  # 创建子文件夹：train/chamomile/os.makedirs(split_class_path, exist_ok=True)for img in imgs_list:src_path = os.path.join(cls_path, img)dst_path = os.path.join(split_class_path, img)copyfile(src_path, dst_path)

数据预处理

# 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}")# 训练集数据增强（彩色图像通用处理）
train_transform = transforms.Compose([transforms.Resize((224, 224)),         # 调整尺寸为224x224（匹配CNN输入）transforms.RandomCrop(224, padding=4),   # 随机裁剪并填充，增加数据多样性transforms.RandomHorizontalFlip(),      # 水平翻转（概率0.5）transforms.ColorJitter(brightness=0.2, contrast=0.2),  # 颜色抖动transforms.ToTensor(),                  # 转换为张量transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))  # ImageNet标准归一化
])# 测试集仅归一化，不增强
test_transform = transforms.Compose([transforms.Resize((224, 224)),transforms.ToTensor(),transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))
])

使用设备: cuda

加载数据集

data_root = "flowers"  # 数据集根目录，需包含5个子类别文件夹train_dataset = datasets.ImageFolder(root=os.path.join(data_root, "train"),  transform=train_transform
)val_dataset = datasets.ImageFolder(root=os.path.join(data_root, "val"),transform=test_transform
)test_dataset = datasets.ImageFolder(root=os.path.join(data_root, "test"),transform=test_transform
)# 创建数据加载器
batch_size = 32
train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True, num_workers=2)
val_loader = DataLoader(val_dataset, batch_size=batch_size, shuffle=False, num_workers=2)
test_loader = DataLoader(test_dataset, batch_size=batch_size, shuffle=False, num_workers=2)# 获取类别名称（自动从文件夹名获取）
class_names = train_dataset.classes
print(f"检测到的类别: {class_names}")  # 确保输出5个类别名称

检测到的类别: ['daisy', 'dandelion', 'rose', 'sunflower', 'tulip']

定义CNN模型

class FlowerCNN(nn.Module):def __init__(self, num_classes=5):super(FlowerCNN, self).__init__()# 卷积块1self.conv1 = nn.Conv2d(3, 32, kernel_size=3, padding=1)self.bn1 = nn.BatchNorm2d(32)self.relu1 = nn.ReLU()self.pool1 = nn.MaxPool2d(2, 2)# 卷积块2self.conv2 = nn.Conv2d(32, 64, kernel_size=3, padding=1)self.bn2 = nn.BatchNorm2d(64)self.relu2 = nn.ReLU()self.pool2 = nn.MaxPool2d(2, 2)# 卷积块3self.conv3 = nn.Conv2d(64, 128, kernel_size=3, padding=1)self.bn3 = nn.BatchNorm2d(128)self.relu3 = nn.ReLU()self.pool3 = nn.MaxPool2d(2, 2)# 全连接层self.fc1 = nn.Linear(128 * 28 * 28, 512)  # 计算方式：224->112->56->28（三次池化后尺寸）self.dropout = nn.Dropout(0.5)self.fc2 = nn.Linear(512, num_classes)  # 输出5个类别def forward(self, x):x = self.pool1(self.relu1(self.bn1(self.conv1(x))))x = self.pool2(self.relu2(self.bn2(self.conv2(x))))x = self.pool3(self.relu3(self.bn3(self.conv3(x))))x = x.view(x.size(0), -1)  # 展平特征图x = self.dropout(self.relu1(self.fc1(x)))x = self.fc2(x)return x# 初始化模型并移至设备
model = FlowerCNN(num_classes=5).to(device)
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=0.001)
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='min', patience=3)

训练模型

def train_model(model, train_loader, val_loader, epochs=10):best_val_acc = 0.0train_loss_history = []val_loss_history = []train_acc_history = []val_acc_history = []for epoch in range(epochs):model.train()running_loss = 0.0correct = 0total = 0for batch_idx, (data, target) in enumerate(train_loader):data, target = data.to(device), target.to(device)optimizer.zero_grad()outputs = model(data)loss = criterion(outputs, target)loss.backward()optimizer.step()running_loss += loss.item()_, predicted = torch.max(outputs.data, 1)total += target.size(0)correct += (predicted == target).sum().item()# 每50批次打印进度if (batch_idx+1) % 50 == 0:print(f"Epoch [{epoch+1}/{epochs}] Batch {batch_idx+1}/{len(train_loader)} "f"Loss: {loss.item():.4f} Acc: {(100*correct/total):.2f}%")# 计算 epoch 指标epoch_train_loss = running_loss / len(train_loader)epoch_train_acc = 100. * correct / total# 验证集评估model.eval()val_loss = 0.0val_correct = 0val_total = 0with torch.no_grad():for data, target in val_loader:data, target = data.to(device), target.to(device)outputs = model(data)val_loss += criterion(outputs, target).item()_, predicted = torch.max(outputs.data, 1)val_total += target.size(0)val_correct += (predicted == target).sum().item()epoch_val_loss = val_loss / len(val_loader)epoch_val_acc = 100. * val_correct / val_totalscheduler.step(epoch_val_loss)# 记录历史数据train_loss_history.append(epoch_train_loss)val_loss_history.append(epoch_val_loss)train_acc_history.append(epoch_train_acc)val_acc_history.append(epoch_val_acc)print(f"Epoch {epoch+1} 完成 | 训练损失: {epoch_train_loss:.4f} 验证准确率: {epoch_val_acc:.2f}%")# 保存最佳模型if epoch_val_acc > best_val_acc:torch.save(model.state_dict(), "best_flower_model.pth")best_val_acc = epoch_val_accprint("保存最佳模型...")# 绘制训练曲线（沿用你的绘图逻辑）plt.figure(figsize=(12, 4))# 损失曲线plt.subplot(1, 2, 1)plt.plot(train_loss_history, label='训练损失')plt.plot(val_loss_history, label='验证损失')plt.title('损失曲线')plt.xlabel('Epoch')plt.ylabel('损失值')plt.legend()# 准确率曲线plt.subplot(1, 2, 2)plt.plot(train_acc_history, label='训练准确率')plt.plot(val_acc_history, label='验证准确率')plt.title('准确率曲线')plt.xlabel('Epoch')plt.ylabel('准确率 (%)')plt.legend()plt.tight_layout()plt.show()return best_val_acc# 训练模型（可调整epochs，建议先试5-10轮）
print("开始训练...")
final_acc = train_model(model, train_loader, val_loader, epochs=15)
print(f"训练完成！最佳验证准确率: {final_acc:.2f}%")

开始训练...
Epoch [1/15] Batch 50/81 Loss: 2.3773 Acc: 38.50%
Epoch 1 完成 | 训练损失: 9.8838 验证准确率: 42.08%
保存最佳模型...
Epoch [2/15] Batch 50/81 Loss: 1.1984 Acc: 35.38%
Epoch 2 完成 | 训练损失: 1.3779 验证准确率: 46.13%
保存最佳模型...
Epoch [3/15] Batch 50/81 Loss: 1.2510 Acc: 41.69%
Epoch 3 完成 | 训练损失: 1.3224 验证准确率: 50.17%
保存最佳模型...
Epoch [4/15] Batch 50/81 Loss: 1.3492 Acc: 41.44%
Epoch 4 完成 | 训练损失: 1.3229 验证准确率: 49.13%
Epoch [5/15] Batch 50/81 Loss: 1.2703 Acc: 40.88%
Epoch 5 完成 | 训练损失: 1.2841 验证准确率: 45.20%
Epoch [6/15] Batch 50/81 Loss: 1.0690 Acc: 41.69%
Epoch 6 完成 | 训练损失: 1.2682 验证准确率: 51.68%
保存最佳模型...
Epoch [7/15] Batch 50/81 Loss: 1.3970 Acc: 42.94%
Epoch 7 完成 | 训练损失: 1.2666 验证准确率: 51.33%
Epoch [8/15] Batch 50/81 Loss: 1.4827 Acc: 42.38%
Epoch 8 完成 | 训练损失: 1.2770 验证准确率: 53.18%
保存最佳模型...
Epoch [9/15] Batch 50/81 Loss: 1.3886 Acc: 41.88%
Epoch 9 完成 | 训练损失: 1.2872 验证准确率: 53.29%
保存最佳模型...
Epoch [10/15] Batch 50/81 Loss: 1.1885 Acc: 44.56%
Epoch 10 完成 | 训练损失: 1.2610 验证准确率: 50.40%
Epoch [11/15] Batch 50/81 Loss: 1.1509 Acc: 44.81%
Epoch 11 完成 | 训练损失: 1.2681 验证准确率: 52.83%
Epoch [12/15] Batch 50/81 Loss: 1.5819 Acc: 44.62%
Epoch 12 完成 | 训练损失: 1.2612 验证准确率: 53.99%
保存最佳模型...
Epoch [13/15] Batch 50/81 Loss: 1.2540 Acc: 48.19%
Epoch 13 完成 | 训练损失: 1.2115 验证准确率: 52.60%
Epoch [14/15] Batch 50/81 Loss: 1.4898 Acc: 47.19%
Epoch 14 完成 | 训练损失: 1.2022 验证准确率: 57.23%
保存最佳模型...
Epoch [15/15] Batch 50/81 Loss: 1.1379 Acc: 48.56%
Epoch 15 完成 | 训练损失: 1.1783 验证准确率: 57.46%
保存最佳模型...

训练完成！最佳验证准确率: 57.46%

Grad-CAM可视化

from torch.nn import functional as F
import cv2 
from PIL import Image
import numpy as np
import torchvision.transforms as transformsclass GradCAM:def __init__(self, model, target_layer_name="conv3"):self.model = model.eval()  # 设置模型为评估模式self.target_layer_name = target_layer_name  # 目标卷积层名称（需与模型定义一致）self.gradients = None  # 存储梯度self.activations = None  # 存储激活值# 注册前向和反向钩子函数for name, module in model.named_modules():if name == target_layer_name:module.register_forward_hook(self.forward_hook)module.register_backward_hook(self.backward_hook)breakdef forward_hook(self, module, input, output):"""前向传播时保存激活值"""self.activations = output.detach()  # 不记录梯度的激活值def backward_hook(self, module, grad_input, grad_output):"""反向传播时保存梯度"""self.gradients = grad_output[0].detach()  # 提取梯度（去除批量维度）def generate(self, input_image, target_class=None):"""生成Grad-CAM热力图"""# 前向传播获取模型输出outputs = self.model(input_image)  # 输出形状: [batch_size, num_classes]if target_class is None:# 若未指定类别，取预测概率最高的类别target_class = torch.argmax(outputs, dim=1).item()# 反向传播计算梯度self.model.zero_grad()  # 清空梯度one_hot = torch.zeros_like(outputs)  # 创建one-hot向量one_hot[0, target_class] = 1  # 目标类别设为1outputs.backward(gradient=one_hot)  # 反向传播# 获取激活值和梯度（形状: [batch_size, channels, height, width]）gradients = self.gradients  # [1, channels, H, W]activations = self.activations  # [1, channels, H, W]# 计算通道权重（全局平均池化）weights = torch.mean(gradients, dim=(2, 3))  # 权重形状: [1, channels]# 生成类激活映射（CAM）cam = torch.sum(activations[0] * weights[0][:, None, None], dim=0)  # 加权求和cam = F.relu(cam)  # 保留正贡献区域cam = (cam - cam.min()) / (cam.max() - cam.min() + 1e-8)  # 归一化到[0, 1]cam = F.interpolate(cam.unsqueeze(0).unsqueeze(0),  # 添加批量和通道维度size=(224, 224),  # 调整尺寸与输入图像一致mode='bilinear', align_corners=False).squeeze()  # 去除批量和通道维度# 将GPU张量转换为CPU的NumPy数组（关键修正）return cam.cpu().numpy(), target_class  # 返回热力图和目标类别# 可视化函数（关键修改：增加图像尺寸统一和颜色通道转换）
def visualize_gradcam(img_path, model, class_names, alpha=0.6):"""可视化Grad-CAM结果:param img_path: 测试图像路径:param model: 训练好的模型:param class_names: 类别名称列表:param alpha: 热力图透明度（0-1）"""# 加载图像并统一尺寸为224x224（解决尺寸不匹配问题）img = Image.open(img_path).convert("RGB")img = img.resize((224, 224))  # 强制Resize到224x224img_np = np.array(img) / 255.0  # 原始图像（尺寸224x224，RGB通道）# 预处理图像（与模型输入一致）transform = transforms.Compose([transforms.ToTensor(),transforms.Normalize(mean=(0.485, 0.456, 0.406),std=(0.229, 0.224, 0.225))])input_tensor = transform(img).unsqueeze(0).to(device)  # 添加批量维度并移至设备# 生成Grad-CAM热力图grad_cam = GradCAM(model, target_layer_name="conv3")  # 确保层名与模型一致heatmap, pred_class = grad_cam.generate(input_tensor)# 热力图后处理（解决颜色通道问题）heatmap = np.uint8(255 * heatmap)  # 转为0-255像素值heatmap = cv2.applyColorMap(heatmap, cv2.COLORMAP_JET)  # 生成BGR格式热力图heatmap = heatmap / 255.0  # 归一化到[0,1]heatmap_rgb = heatmap[:, :, ::-1]  # BGR转RGB（正确显示颜色）# 叠加原始图像和热力图（尺寸和通道完全匹配）superimposed = cv2.addWeighted(img_np, 1 - alpha, heatmap, alpha, 0)# 绘制结果plt.figure(figsize=(12, 4))# 原始图像plt.subplot(1, 3, 1)plt.imshow(img_np)plt.title(f"原始图像\n真实类别: {img_path.split('/')[-2]}")plt.axis('off')# 热力图（显示为RGB格式）plt.subplot(1, 3, 2)plt.imshow(heatmap_rgb)  # 使用转换后的RGB热力图plt.title(f"Grad-CAM热力图\n预测类别: {class_names[pred_class]}")plt.axis('off')# 叠加图plt.subplot(1, 3, 3)plt.imshow(superimposed)plt.title("叠加热力图")plt.axis('off')plt.tight_layout()plt.show()# 选择测试图像（需存在且路径正确）
test_image_path = "flowers/tulip/100930342_92e8746431_n.jpg"  # 执行可视化
visualize_gradcam(test_image_path, model, class_names)

结果分析

1. 训练过程解析

损失曲线：
- 训练损失（蓝线）初期快速下降（Epoch 0-2），随后稳定在 1.2 左右，表明模型快速收敛并进入平稳学习阶段。
- 验证损失（橙线）与训练损失趋势一致，且差距极小（如 Epoch 15 时训练损失 1.178，验证损失约 1.2），说明模型未过拟合，泛化能力良好。
准确率曲线：
- 训练准确率（蓝线）和验证准确率（橙线）均呈上升趋势，验证准确率最终达57.46%（Epoch 15），训练准确率约 48%。验证准确率高于训练准确率，可能因：
  - 训练集与验证集数据分布差异（如数据增强在训练集的作用更显著）。
  - 模型对验证集的特征拟合更优（需进一步检查数据划分是否合理）。

2. Grad-CAM 可视化验证

分类正确性：
原始图像（郁金香）的预测类别与真实类别一致（tulip），模型分类正确。
注意力区域：
热力图（中间图）高亮花瓣区域（红色、黄色），叠加热力图（右图）显示模型聚焦花朵的花瓣和花蕊，符合人类对郁金香的视觉特征（花瓣是识别郁金香的关键部位）。
可解释性：
模型通过关注花瓣区域做出分类决策，验证了 Grad-CAM 的有效性，为后续模型优化（如调整卷积层关注区域）提供直观依据。

3.模型性能

验证准确率约 57%，在 5 类花卉分类中表现中等（随机猜测为 20%），但仍有提升空间。

接下来对模型进行改进以提高准确率。

改进模型

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
from PIL import Image
import os
from sklearn.model_selection import train_test_split
from shutil import copyfile
import cv2
from torch.nn import functional as F# 数据集划分（保持不变）
data_root = "flowers"
classes = ["daisy", "tulip", "rose", "sunflower", "dandelion"]
for folder in ["train", "val", "test"]:os.makedirs(os.path.join(data_root, folder), exist_ok=True)
for cls in classes:cls_path = os.path.join(data_root, cls)imgs = [f for f in os.listdir(cls_path) if f.lower().endswith((".jpg", ".jpeg", ".png"))]train_val, test = train_test_split(imgs, test_size=0.2, random_state=42)train, val = train_test_split(train_val, test_size=0.25, random_state=42)for split, imgs_list in zip(["train", "val", "test"], [train, val, test]):split_class_path = os.path.join(data_root, split, cls)os.makedirs(split_class_path, exist_ok=True)for img in imgs_list:copyfile(os.path.join(cls_path, img), os.path.join(split_class_path, img))# 数据预处理（新增旋转增强）
plt.rcParams["font.family"] = ["SimHei"]
plt.rcParams['axes.unicode_minus'] = False
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")train_transform = transforms.Compose([transforms.Resize((224, 224)),transforms.RandomCrop(224, padding=4),transforms.RandomHorizontalFlip(),transforms.RandomRotation(15),  # 新增旋转transforms.ColorJitter(brightness=0.2, contrast=0.2),transforms.ToTensor(),transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))
])test_transform = transforms.Compose([transforms.Resize((224, 224)),transforms.ToTensor(),transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))
])# 数据加载器（保持不变）
train_dataset = datasets.ImageFolder(os.path.join(data_root, "train"), transform=train_transform)
val_dataset = datasets.ImageFolder(os.path.join(data_root, "val"), transform=test_transform)
test_dataset = datasets.ImageFolder(os.path.join(data_root, "test"), transform=test_transform)batch_size = 32
train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True, num_workers=2)
val_loader = DataLoader(val_dataset, batch_size=batch_size, shuffle=False, num_workers=2)
test_loader = DataLoader(test_dataset, batch_size=batch_size, shuffle=False, num_workers=2)class_names = train_dataset.classes# 模型定义（新增第4卷积块）
class FlowerCNN(nn.Module):def __init__(self, num_classes=5):super().__init__()self.conv1 = nn.Conv2d(3, 32, 3, padding=1)self.bn1 = nn.BatchNorm2d(32)self.relu1 = nn.ReLU()self.pool1 = nn.MaxPool2d(2, 2)  # 224→112self.conv2 = nn.Conv2d(32, 64, 3, padding=1)self.bn2 = nn.BatchNorm2d(64)self.relu2 = nn.ReLU()self.pool2 = nn.MaxPool2d(2, 2)  # 112→56self.conv3 = nn.Conv2d(64, 128, 3, padding=1)self.bn3 = nn.BatchNorm2d(128)self.relu3 = nn.ReLU()self.pool3 = nn.MaxPool2d(2, 2)  # 56→28self.conv4 = nn.Conv2d(128, 256, 3, padding=1)  # 新增卷积块self.bn4 = nn.BatchNorm2d(256)self.relu4 = nn.ReLU()self.pool4 = nn.MaxPool2d(2, 2)  # 28→14self.fc1 = nn.Linear(256 * 14 * 14, 512)self.dropout = nn.Dropout(0.5)self.fc2 = nn.Linear(512, num_classes)def forward(self, x):x = self.pool1(self.relu1(self.bn1(self.conv1(x))))x = self.pool2(self.relu2(self.bn2(self.conv2(x))))x = self.pool3(self.relu3(self.bn3(self.conv3(x))))x = self.pool4(self.relu4(self.bn4(self.conv4(x))))  # 新增池化x = x.view(x.size(0), -1)x = self.dropout(self.relu1(self.fc1(x)))x = self.fc2(x)return x# 训练配置（增加轮数，使用StepLR）
model = FlowerCNN().to(device)
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=0.001)
scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=10, gamma=0.1)def train_model(epochs=30):best_val_acc = 0.0train_loss, val_loss, train_acc, val_acc = [], [], [], []for epoch in range(epochs):model.train()running_loss, correct, total = 0.0, 0, 0for data, target in train_loader:data, target = data.to(device), target.to(device)optimizer.zero_grad()outputs = model(data)loss = criterion(outputs, target)loss.backward()optimizer.step()running_loss += loss.item()_, pred = torch.max(outputs, 1)correct += (pred == target).sum().item()total += target.size(0)epoch_train_loss = running_loss / len(train_loader)epoch_train_acc = 100 * correct / totalmodel.eval()val_running_loss, val_correct, val_total = 0.0, 0, 0with torch.no_grad():for data, target in val_loader:data, target = data.to(device), target.to(device)outputs = model(data)val_running_loss += criterion(outputs, target).item()_, pred = torch.max(outputs, 1)val_correct += (pred == target).sum().item()val_total += target.size(0)epoch_val_loss = val_running_loss / len(val_loader)epoch_val_acc = 100 * val_correct / val_totalscheduler.step()train_loss.append(epoch_train_loss)val_loss.append(epoch_val_loss)train_acc.append(epoch_train_acc)val_acc.append(epoch_val_acc)print(f"Epoch {epoch+1}/{epochs} | 训练损失: {epoch_train_loss:.4f} 验证准确率: {epoch_val_acc:.2f}%")if epoch_val_acc > best_val_acc:torch.save(model.state_dict(), "best_model.pth")best_val_acc = epoch_val_acc# 绘制曲线（保持不变）plt.figure(figsize=(12, 4))plt.subplot(1, 2, 1); plt.plot(train_loss, label='训练损失'); plt.plot(val_loss, label='验证损失'); plt.legend()plt.subplot(1, 2, 2); plt.plot(train_acc, label='训练准确率'); plt.plot(val_acc, label='验证准确率'); plt.legend()plt.show()return best_val_acc# 训练与可视化（保持不变）
print("开始训练...")
train_model(epochs=30)
print("训练完成，开始可视化...")class GradCAM:def __init__(self, model, target_layer_name="conv3"):self.model = model.eval()self.target_layer_name = target_layer_nameself.gradients, self.activations = None, Nonefor name, module in model.named_modules():if name == target_layer_name:module.register_forward_hook(self.forward_hook)module.register_backward_hook(self.backward_hook)breakdef forward_hook(self, module, input, output):self.activations = output.detach()def backward_hook(self, module, grad_input, grad_output):self.gradients = grad_output[0].detach()def generate(self, input_image, target_class=None):outputs = self.model(input_image)target_class = torch.argmax(outputs, dim=1).item() if target_class is None else target_classself.model.zero_grad()one_hot = torch.zeros_like(outputs); one_hot[0, target_class] = 1outputs.backward(gradient=one_hot)weights = torch.mean(self.gradients, dim=(2, 3))cam = torch.sum(self.activations[0] * weights[0][:, None, None], dim=0)cam = F.relu(cam); cam = (cam - cam.min()) / (cam.max() - cam.min() + 1e-8) cam = F.interpolate(cam.unsqueeze(0).unsqueeze(0), size=(224, 224), mode='bilinear').squeeze()return cam.cpu().numpy(), target_classdef visualize_gradcam(img_path, model, class_names, alpha=0.6):img = Image.open(img_path).convert("RGB").resize((224, 224))img_np = np.array(img) / 255.0transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))])input_tensor = transform(img).unsqueeze(0).to(device)grad_cam = GradCAM(model, target_layer_name="conv3")heatmap, pred_class = grad_cam.generate(input_tensor)heatmap = np.uint8(255 * heatmap); heatmap = cv2.applyColorMap(heatmap, cv2.COLORMAP_JET) / 255.0; heatmap_rgb = heatmap[:, :, ::-1]superimposed = cv2.addWeighted(img_np, 1 - alpha, heatmap, alpha, 0)plt.figure(figsize=(12, 4))plt.subplot(1, 3, 1); plt.imshow(img_np); plt.title(f"原始图像\n真实类别: {img_path.split('/')[-2]}"); plt.axis('off')plt.subplot(1, 3, 2); plt.imshow(heatmap_rgb); plt.title(f"Grad-CAM热力图\n预测类别: {class_names[pred_class]}"); plt.axis('off')plt.subplot(1, 3, 3); plt.imshow(superimposed); plt.title("叠加热力图"); plt.axis('off')plt.tight_layout(); plt.show()test_image_path = "flowers/tulip/100930342_92e8746431_n.jpg"  
visualize_gradcam(test_image_path, model, class_names)

开始训练...
Epoch 1/30 | 训练损失: 5.8699 验证准确率: 47.05%
Epoch 2/30 | 训练损失: 1.3307 验证准确率: 53.76%
Epoch 3/30 | 训练损失: 1.3045 验证准确率: 52.95%
Epoch 4/30 | 训练损失: 1.2460 验证准确率: 55.38%
Epoch 5/30 | 训练损失: 1.2342 验证准确率: 49.48%
Epoch 6/30 | 训练损失: 1.2442 验证准确率: 54.10%
Epoch 7/30 | 训练损失: 1.2309 验证准确率: 50.75%
Epoch 8/30 | 训练损失: 1.2172 验证准确率: 56.65%
Epoch 9/30 | 训练损失: 1.2025 验证准确率: 56.53%
Epoch 10/30 | 训练损失: 1.1733 验证准确率: 56.53%
Epoch 11/30 | 训练损失: 1.1167 验证准确率: 61.04%
Epoch 12/30 | 训练损失: 1.0763 验证准确率: 64.28%
Epoch 13/30 | 训练损失: 1.0564 验证准确率: 63.12%
Epoch 14/30 | 训练损失: 1.0469 验证准确率: 62.31%
Epoch 15/30 | 训练损失: 1.0295 验证准确率: 65.09%
Epoch 16/30 | 训练损失: 1.0365 验证准确率: 65.78%
Epoch 17/30 | 训练损失: 1.0091 验证准确率: 66.71%
Epoch 18/30 | 训练损失: 1.0152 验证准确率: 65.32%
Epoch 19/30 | 训练损失: 0.9794 验证准确率: 65.43%
Epoch 20/30 | 训练损失: 0.9875 验证准确率: 68.90%
Epoch 21/30 | 训练损失: 0.9496 验证准确率: 69.94%
Epoch 22/30 | 训练损失: 0.9608 验证准确率: 69.71%
Epoch 23/30 | 训练损失: 0.9342 验证准确率: 69.71%
Epoch 24/30 | 训练损失: 0.9586 验证准确率: 69.25%
Epoch 25/30 | 训练损失: 0.9554 验证准确率: 69.60%
Epoch 26/30 | 训练损失: 0.9463 验证准确率: 69.83%
Epoch 27/30 | 训练损失: 0.9373 验证准确率: 69.94%
Epoch 28/30 | 训练损失: 0.9282 验证准确率: 69.48%
Epoch 29/30 | 训练损失: 0.9130 验证准确率: 69.36%
Epoch 30/30 | 训练损失: 0.9585 验证准确率: 69.94%

训练完成，开始可视化...

对比分析

1. 训练过程对比

初版代码（15 轮）：
- 训练损失初期下降快（Epoch 1-2），但后续波动较大（如验证准确率在 Epoch 5 降至 45.20%），最终验证准确率 57.46%。
- 曲线稳定性不足，反映模型对数据变化的适应性一般。
修改后代码（30 轮，优化后）：
- 训练损失更平滑（最终 0.96），验证准确率稳定在 69.94%（提升约 12.5%）。
- 学习率衰减（StepLR）和数据增强（旋转）使训练过程更稳定，模型收敛更充分，泛化能力增强。

2. 模型性能提升

准确率：修改后验证准确率提升至 69.94%，在 5 类花卉分类中表现更优（初版约 57%，随机猜测 20%）。
损失曲线：修改后训练与验证损失差距缩小（初版差距约 0.1，修改后约 0.05），说明过拟合风险降低，模型更鲁棒。

3. Grad-CAM 可视化对比

初版（图 2）：热力图聚焦花瓣（红色区域），分类正确，但热力图颜色分布较分散（蓝色背景占比大），对关键区域的突出度一般。
修改后（图 4）：热力图更清晰地突出花瓣和花蕊（颜色更鲜艳，红色 / 黄色区域集中），说明模型对花卉关键特征（如花瓣纹理、花蕊形状）的关注度提升，可解释性更强，决策依据更直观。

4. 优化策略有效性

数据增强（旋转）：新增RandomRotation(15)，使模型学习不同角度的花卉，增强对姿态变化的适应性（如郁金香的不同拍摄角度），反映在验证准确率的稳定性提升（修改后后期波动更小）。
模型深度（新增卷积块）：通过conv4增加特征提取深度，捕捉更细粒度的特征（如花瓣的细微纹理），体现在热力图的更精确聚焦，提升分类精度。
训练策略（轮数 + 学习率衰减）：延长训练轮数（30 轮）并使用StepLR，让模型充分学习（初版 15 轮可能未完全收敛），避免学习率过高导致的震荡，最终准确率显著提升。