python打卡day43

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

找了个街头食物图像分类的数据集Popular Street Foods（其实写代码的时候就开始后悔了），原因在于：

1、如果是比较规整的图像分类数据集，自己划分了train和test目录，就可以直接用pytorch内置的ImageFolder类，不用自己再辛辛苦苦定义数据集类了，但很遗憾这个数据集不规整，图片在类别下面，标签用一个csv文件存储，所以不仅要自己定义数据类，训练测试的时候还要自己划分数据集

2、图片尺寸非常不统一，需要自己预先处理，遍历了一遍这个数据集，最后决定全部统一成140x140的

from PIL import Image
import os
from collections import defaultdict# 替换为你的数据集路径
dataset_path = "/kaggle/input/popular-street-foods/popular_street_foods/dataset"size_dict = defaultdict(int)for root, _, files in os.walk(dataset_path):for file in files:if file.lower().endswith(('png', 'jpg', 'jpeg')):try:with Image.open(os.path.join(root, file)) as img:size_dict[img.size] += 1  # (width, height)except:print(f"Error reading: {file}")# 打印尺寸统计
print("Top 10 common sizes:")
for size, count in sorted(size_dict.items(), key=lambda x: x[1], reverse=True)[:10]:print(f"{size}: {count} images")# -----------------------------------------------------
Top 10 common sizes:
(140, 140): 851 images
(93, 140): 574 images
(162, 108): 548 images
(162, 121): 308 images
(162, 91): 295 images
(105, 140): 119 images
(112, 140): 90 images
(100, 140): 30 images
(154, 140): 28 images
(162, 85): 26 images

完成代码：

import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
from torchvision import datasets, transforms
from torchvision.datasets import ImageFolder
from torch.utils.data import Dataset, DataLoader, random_split
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd# 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
# 图像尺寸统一成140x140
class PhotoResizer:def __init__(self, target_size=140, fill_color=114): # target_size: 目标正方形尺寸，fill_color: 填充使用的灰度值self.target_size = target_sizeself.fill_color = fill_color# 预定义转换方法self.to_tensor = transforms.ToTensor()self.normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],std=[0.229, 0.224, 0.225])def __call__(self, img):"""智能处理流程：1. 对小图像进行填充，对大图像进行智能裁剪2. 保持长宽比的情况下进行保护性处理"""w, h = img.sizeif w == h == self.target_size: # 情况1：已经是目标尺寸pass  # 无需处理elif min(w, h) < self.target_size: # 情况2：至少有一个维度小于目标尺寸（需要填充）img = self.padding_resize(img)else:                           # 情况3：两个维度都大于目标尺寸（智能裁剪）img = self.crop_resize(img)# 最终统一转换return self.normalize(self.to_tensor(img))def padding_resize(self, img): # 等比缩放后居中填充不足部分w, h = img.sizescale = self.target_size / min(w, h)new_w, new_h = int(w * scale), int(h * scale)img = img.resize((new_w, new_h), Image.BILINEAR)# 等比缩放 + 居中填充# 计算需要填充的像素数（4个值：左、上、右、下）pad_left = (self.target_size - new_w) // 2pad_top = (self.target_size - new_h) // 2pad_right = self.target_size - new_w - pad_leftpad_bottom = self.target_size - new_h - pad_topreturn transforms.functional.pad(img, [pad_left, pad_top, pad_right, pad_bottom], self.fill_color)def crop_resize(self, img): # 等比缩放后中心裁剪w, h = img.sizeratio = w / h# 计算新尺寸（保护长边）if ratio < 0.9:  # 竖图new_size = (self.target_size, int(h * self.target_size / w))elif ratio > 1.1:  # 横图new_size = (int(w * self.target_size / h), self.target_size)else:  # 近似正方形new_size = (self.target_size, self.target_size)img = img.resize(new_size, Image.BILINEAR)return transforms.functional.center_crop(img, self.target_size)# 训练集测试集预处理
train_transform = transforms.Compose([PhotoResizer(target_size=140),  # 自动处理所有情况transforms.RandomHorizontalFlip(), # 随机水平翻转图像（概率0.5）transforms.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2, hue=0.1), # 随机颜色抖动：亮度、对比度、饱和度和色调随机变化transforms.RandomRotation(15), # 随机旋转图像（最大角度15度）
])test_transform = transforms.Compose([PhotoResizer(target_size=140)
])# 2. 创建dataset和dataloader实例
class StreetFoodDataset(Dataset):def __init__(self, root_dir, transform=None):self.root_dir = root_dirself.transform = transformself.image_paths = []self.labels = []self.class_to_idx = {}# 遍历目录获取类别映射classes = sorted(entry.name for entry in os.scandir(root_dir) if entry.is_dir())self.class_to_idx = {cls_name: i for i, cls_name in enumerate(classes)}# 收集图像路径和标签for class_name in classes:class_dir = os.path.join(root_dir, class_name)for img_name in os.listdir(class_dir):if img_name.lower().endswith(('.png', '.jpg', '.jpeg')):self.image_paths.append(os.path.join(class_dir, img_name))self.labels.append(self.class_to_idx[class_name])def __len__(self):return len(self.image_paths)def __getitem__(self, idx):img_path = self.image_paths[idx]image = Image.open(img_path).convert('RGB')label = self.labels[idx]if self.transform:image = self.transform(image)return image, label# 数据集路径（Kaggle路径示例）
dataset_path = '/kaggle/input/popular-street-foods/popular_street_foods/dataset'# 创建数据集实例
# 先创建基础数据集
full_dataset = StreetFoodDataset(root_dir=dataset_path)# 分割数据集
train_size = int(0.8 * len(full_dataset))
test_size = len(full_dataset) - train_size
train_dataset, test_dataset = random_split(full_dataset, [train_size, test_size])
train_dataset.dataset.transform = train_transform
test_dataset.dataset.transform = test_transform# 创建数据加载器
train_loader = DataLoader(train_dataset,batch_size=32,shuffle=True,num_workers=2,pin_memory=True
)test_loader = DataLoader(test_dataset,batch_size=32,shuffle=False,num_workers=2,pin_memory=True
)# 3. 定义CNN模型
class CNN(nn.Module):def __init__(self, num_classes=20): super(CNN, self).__init__()# 第一个卷积块self.conv1 = nn.Conv2d(3, 32, kernel_size=3, padding=1)self.bn1 = nn.BatchNorm2d(32)self.relu1 = nn.ReLU()self.pool1 = nn.MaxPool2d(kernel_size=2, stride=2)  # 140x140 → 70x70# 第二个卷积块self.conv2 = nn.Conv2d(32, 64, kernel_size=3, padding=1)self.bn2 = nn.BatchNorm2d(64)self.relu2 = nn.ReLU()self.pool2 = nn.MaxPool2d(kernel_size=2)  # 70x70 → 35x35# 第三个卷积块self.conv3 = nn.Conv2d(64, 128, kernel_size=3, padding=1)self.bn3 = nn.BatchNorm2d(128)self.relu3 = nn.ReLU()self.pool3 = nn.MaxPool2d(kernel_size=2)  # 35x35 → 17x17（下采样时尺寸向下取整）# 第四个卷积块（新增，处理更大尺寸）self.conv4 = nn.Conv2d(128, 256, kernel_size=3, padding=1)self.bn4 = nn.BatchNorm2d(256)self.relu4 = nn.ReLU()self.pool4 = nn.MaxPool2d(kernel_size=2)  # 17x17 → 8x8# 全连接层（分类器）# 修改：计算展平后的特征维度 256通道 × 8x8尺寸 = 16384self.fc1 = nn.Linear(256 * 8 * 8, 512)self.dropout = nn.Dropout(p=0.5)self.fc2 = nn.Linear(512, num_classes)  # 使用num_classes参数def forward(self, x):# 输入尺寸：[batch_size, 3, 140, 140]# 卷积块1处理x = self.conv1(x)       # [batch_size, 32, 140, 140]x = self.bn1(x)x = self.relu1(x)x = self.pool1(x)       # [batch_size, 32, 70, 70]# 卷积块2处理x = self.conv2(x)       # [batch_size, 64, 70, 70]x = self.bn2(x)x = self.relu2(x)x = self.pool2(x)       # [batch_size, 64, 35, 35]# 卷积块3处理x = self.conv3(x)       # [batch_size, 128, 35, 35]x = self.bn3(x)x = self.relu3(x)x = self.pool3(x)       # [batch_size, 128, 17, 17]# 卷积块4处理（新增）x = self.conv4(x)       # [batch_size, 256, 17, 17]x = self.bn4(x)x = self.relu4(x)x = self.pool4(x)       # [batch_size, 256, 8, 8]# 展平与全连接层x = x.view(-1, 256 * 8 * 8)  # [batch_size, 16384]x = self.fc1(x)           # [batch_size, 512]x = self.relu3(x)         # 复用relu3x = self.dropout(x)x = self.fc2(x)           # [batch_size, num_classes]return x    # 初始化模型
model = CNN()
model = model.to(device)  # 将模型移至GPU（如果可用）# 4. 训练测试
# 定义损失函数、优化器、调度器
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=0.001)
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, 'max', patience=5, factor=0.5)# 训练过程封装
def train_epoch(model, loader, criterion, optimizer):model.train()running_loss = 0.0correct = 0total = 0for images, labels in loader:images, labels = images.to(device), labels.to(device)optimizer.zero_grad()outputs = model(images)loss = criterion(outputs, labels)loss.backward()optimizer.step()running_loss += loss.item()_, predicted = outputs.max(1)total += labels.size(0)correct += predicted.eq(labels).sum().item()return running_loss/len(loader), 100.*correct/total# 测试过程封装
def test(model, loader, criterion):model.eval()running_loss = 0.0correct = 0total = 0with torch.no_grad():for images, labels in loader:images, labels = images.to(device), labels.to(device)outputs = model(images)loss = criterion(outputs, labels)running_loss += loss.item()_, predicted = outputs.max(1)total += labels.size(0)correct += predicted.eq(labels).sum().item()return running_loss/len(loader), 100.*correct/total# 训练循环
epochs = 1000
best_acc = 0.0
patience = 10  # 早停耐心值
no_improve = 0  # 没有提升的epoch计数# 创建保存目录
os.makedirs("checkpoints", exist_ok=True)for epoch in range(epochs):train_loss, train_acc = train_epoch(model, train_loader, criterion, optimizer)test_loss, test_acc = test(model, test_loader, criterion)scheduler.step(test_acc)  # 根据测试准确率调整学习率# 每200个epoch打印一次信息if (epoch + 1) % 200 == 0 or epoch == 0 or (epoch + 1) == epochs:print(f"\nEpoch {epoch+1}/{epochs}")print(f"Train Loss: {train_loss:.4f} | Acc: {train_acc:.2f}%")print(f"Test Loss: {test_loss:.4f} | Acc: {test_acc:.2f}%")print(f"Current LR: {optimizer.param_groups[0]['lr']:.6f}")print("-"*50)# 保存最佳模型if test_acc > best_acc:best_acc = test_acctorch.save(model.state_dict(), "checkpoints/best_model.pth")no_improve = 0else:no_improve += 1# 定期保存检查点if (epoch + 1) % 200 == 0:torch.save({'epoch': epoch,'model_state_dict': model.state_dict(),'optimizer_state_dict': optimizer.state_dict(),'best_acc': best_acc,}, f"checkpoints/checkpoint_{epoch+1}.pth")# 早停检查if no_improve >= patience:print(f"\nEarly stopping at epoch {epoch+1}, no improvement for {patience} epochs")print(f"Best test accuracy: {best_acc:.2f}%")breakprint(f"\n训练完成，最佳测试准确率: {best_acc:.2f}%")# 5. Grad-CAM实现
model.eval()# Grad-CAM实现
class GradCAM:def __init__(self, model, target_layer):self.model = modelself.target_layer = target_layerself.gradients = Noneself.activations = None# 注册钩子，用于获取目标层的前向传播输出和反向传播梯度self.register_hooks()def register_hooks(self):# 前向钩子函数，在目标层前向传播后被调用，保存目标层的输出（激活值）def forward_hook(module, input, output):self.activations = output.detach()# 反向钩子函数，在目标层反向传播后被调用，保存目标层的梯度def backward_hook(module, grad_input, grad_output):self.gradients = grad_output[0].detach()# 在目标层注册前向钩子和反向钩子self.target_layer.register_forward_hook(forward_hook)self.target_layer.register_backward_hook(backward_hook)def generate_cam(self, input_image, target_class=None):# 前向传播，得到模型输出model_output = self.model(input_image)if target_class is None:# 如果未指定目标类别，则取模型预测概率最大的类别作为目标类别target_class = torch.argmax(model_output, dim=1).item()# 清除模型梯度，避免之前的梯度影响self.model.zero_grad()# 反向传播，构造one-hot向量，使得目标类别对应的梯度为1，其余为0，然后进行反向传播计算梯度one_hot = torch.zeros_like(model_output)one_hot[0, target_class] = 1model_output.backward(gradient=one_hot)# 获取之前保存的目标层的梯度和激活值gradients = self.gradientsactivations = self.activations# 对梯度进行全局平均池化，得到每个通道的权重，用于衡量每个通道的重要性weights = torch.mean(gradients, dim=(2, 3), keepdim=True)# 加权激活映射，将权重与激活值相乘并求和，得到类激活映射的初步结果cam = torch.sum(weights * activations, dim=1, keepdim=True)# ReLU激活，只保留对目标类别有正贡献的区域，去除负贡献的影响cam = F.relu(cam)# 调整大小并归一化，将类激活映射调整为与输入图像相同的尺寸（140x140），并归一化到[0, 1]范围cam = F.interpolate(cam, size=(140, 140), mode='bilinear', align_corners=False)cam = cam - cam.min()cam = cam / cam.max() if cam.max() > 0 else camreturn cam.cpu().squeeze().numpy(), target_class# 选择一个随机图像
idx = np.random.randint(len(test_dataset))
image, label = test_dataset[idx]
classes = sorted(os.listdir('/kaggle/input/popular-street-foods/popular_street_foods/dataset'))
print(f"选择的图像类别: {classes[label]}")# 转换图像以便可视化
def tensor_to_np(tensor):img = tensor.cpu().numpy().transpose(1, 2, 0)mean = np.array([0.485, 0.456, 0.406])std = np.array([0.229, 0.224, 0.225])img = std * img + meanimg = np.clip(img, 0, 1)return img# 添加批次维度并移动到设备
input_tensor = image.unsqueeze(0).to(device)# 初始化Grad-CAM（选择最后一个卷积层）
grad_cam = GradCAM(model, model.conv4 )# 生成热力图
heatmap, pred_class = grad_cam.generate_cam(input_tensor)# 可视化
plt.figure(figsize=(12, 4))# 原始图像
plt.subplot(1, 3, 1)
plt.imshow(tensor_to_np(image))
plt.title(f"原始图像: {classes[label]}")
plt.axis('off')# 热力图
plt.subplot(1, 3, 2)
plt.imshow(heatmap, cmap='jet')
plt.title(f"Grad-CAM热力图: {classes[pred_class]}")
plt.axis('off')# 叠加的图像
plt.subplot(1, 3, 3)
img = tensor_to_np(image)
heatmap_resized = np.uint8(255 * heatmap)
heatmap_colored = plt.cm.jet(heatmap_resized)[:, :, :3]
superimposed_img = heatmap_colored * 0.4 + img * 0.6
plt.imshow(superimposed_img)
plt.title("叠加热力图")
plt.axis('off')plt.tight_layout()
plt.savefig('grad_cam_result.png')
plt.show()

收获心得：

1、就是为了训练有更好的效果，所以图片尺寸处理的时候那么费劲（不然完全可以全部裁剪成一个更小的尺寸），但是最后训练出来的准确率还是不尽人意，可能是cnn结构的问题？下次再看看怎么调整

2、之前的数据集创建dataset和dataloader的时候都很轻松，到了这个数据集就成了比较困难的一个点了，首先是文件路径的问题，其次就是在数据集类的定义里对图像和标签的处理

3、训练的时候没有画损失曲线，所以过程比较简单，就想着加上早停策略，然后一直报错，debug有点de麻了，最后也不知道怎么就解决了，管他的能跑就行（

4、Grag-CAM部分的改动不大，之前的拿过来用，最后不想拆分文件了，比较懒

@浙大疏锦行