import torch import torchvision import torchvision.transforms as transforms import os import numpy as np import random import yaml from torch.utils.data import TensorDataset, DataLoader # 加载数据集 def get_cifar10_dataloaders(batch_size=128, num_workers=2, local_dataset_path=None, shuffle=False): """获取CIFAR10数据集的数据加载器 Args: batch_size: 批次大小 num_workers: 数据加载的工作进程数 local_dataset_path: 本地数据集路径，如果提供则使用本地数据集，否则下载 Returns: trainloader: 训练数据加载器 testloader: 测试数据加载器 """ # 数据预处理 transform_train = transforms.Compose([ transforms.RandomCrop(32, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ]) transform_test = transforms.Compose([ transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ]) # 设置数据集路径 if local_dataset_path: print(f"使用本地数据集: {local_dataset_path}") # 检查数据集路径是否有数据集，没有的话则下载 cifar_path = os.path.join(local_dataset_path, 'cifar-10-batches-py') download = not os.path.exists(cifar_path) or not os.listdir(cifar_path) dataset_path = local_dataset_path else: print("未指定本地数据集路径，将下载数据集") download = True dataset_path = '../dataset' # 创建数据集路径 if not os.path.exists(dataset_path): os.makedirs(dataset_path) trainset = torchvision.datasets.CIFAR10( root=dataset_path, train=True, download=download, transform=transform_train) trainloader = torch.utils.data.DataLoader( trainset, batch_size=batch_size, shuffle=shuffle, num_workers=num_workers) testset = torchvision.datasets.CIFAR10( root=dataset_path, train=False, download=download, transform=transform_test) testloader = torch.utils.data.DataLoader( testset, batch_size=batch_size, shuffle=shuffle, num_workers=num_workers) return trainloader, testloader def get_noisy_cifar10_dataloaders(batch_size=128, num_workers=2, local_dataset_path=None, shuffle=False): """获取添加噪声后的CIFAR10数据集的数据加载器 Args: batch_size: 批次大小 num_workers: 数据加载的工作进程数 local_dataset_path: 本地数据集路径，如果提供则使用本地数据集，否则下载 shuffle: 是否打乱数据 Returns: noisy_trainloader: 添加噪声后的训练数据加载器 testloader: 正常测试数据加载器 """ # 加载原始数据集 trainloader, testloader = get_cifar10_dataloaders( batch_size=batch_size, num_workers=num_workers, local_dataset_path=local_dataset_path, shuffle=False ) # 设置设备 device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') print(f"使用设备: {device}") # 加载配置文件 config_path = './train.yaml' try: with open(config_path, 'r') as f: config = yaml.safe_load(f) except FileNotFoundError: print(f"找不到配置文件: {config_path}，使用默认配置") config = { 'noise_levels': { 'gaussian': [0.1, 0.3], 'salt_pepper': [0.05, 0.1], 'poisson': [1.0] } } # 加载噪声参数 noise_levels = config.get('noise_levels', {}) gaussian_level = noise_levels.get('gaussian', [0.1, 0.2]) salt_pepper_level = noise_levels.get('salt_pepper', [0.05, 0.1]) poisson_level = noise_levels.get('poisson', [1.0])[0] # 获取原始数据和标签 data_list = [] targets_list = [] for inputs, targets in trainloader: data_list.append(inputs) targets_list.append(targets) # 合并所有批次数据 all_data = torch.cat(data_list) all_targets = torch.cat(targets_list) # 创建噪声信息字典 noise_info = { 'noise_types': [], 'noise_levels': [], 'noise_indices': [] } # CIFAR10标准化参数 mean = torch.tensor([0.4914, 0.4822, 0.4465]).view(3, 1, 1).to(device) std = torch.tensor([0.2023, 0.1994, 0.2010]).view(3, 1, 1).to(device) print("开始添加噪声...") # 按标签分组进行处理 for label_value in range(10): # 找出所有具有当前标签的样本索引 indices = [i for i in range(len(all_targets)) if all_targets[i].item() == label_value] noise_type = None noise_ratio = 0.0 level = None # 根据标签决定噪声类型和强度 if label_value == 2: # 高斯噪声强 - 30%数据 noise_type = 1 # 高斯噪声 noise_ratio = 0.3 level = gaussian_level[1] if len(gaussian_level) > 1 else gaussian_level[0] elif label_value == 3: # 高斯噪声弱 - 10%数据 noise_type = 1 # 高斯噪声 noise_ratio = 0.1 level = gaussian_level[0] elif label_value == 4: # 椒盐噪声强 - 30%数据 noise_type = 2 # 椒盐噪声 noise_ratio = 0.3 level = salt_pepper_level[1] if len(salt_pepper_level) > 1 else salt_pepper_level[0] elif label_value == 5: # 椒盐噪声弱 - 10%数据 noise_type = 2 # 椒盐噪声 noise_ratio = 0.1 level = salt_pepper_level[0] elif label_value == 6: # 泊松噪声 - 30%数据 noise_type = 3 # 泊松噪声 noise_ratio = 0.3 level = poisson_level elif label_value == 7: # 泊松噪声 - 10%数据 noise_type = 3 # 泊松噪声 noise_ratio = 0.1 level = poisson_level # 如果需要添加噪声 if noise_type is not None and level is not None and noise_ratio > 0: # 计算要添加噪声的样本数量 num_samples_to_add_noise = int(len(indices) * noise_ratio) if num_samples_to_add_noise == 0 and len(indices) > 0: num_samples_to_add_noise = 1 # 至少添加一个样本 # 随机选择要添加噪声的样本索引 indices_to_add_noise = random.sample(indices, min(num_samples_to_add_noise, len(indices))) print(f"标签 {label_value}: 为 {len(indices_to_add_noise)}/{len(indices)} 个样本添加噪声类型 {noise_type}，强度 {level}") # 为选中的样本添加噪声 for i in indices_to_add_noise: # 获取当前图像 img = all_data[i].to(device) # 反标准化 img_denorm = img * std + mean # 添加噪声 if noise_type == 1: # 高斯噪声 # 转为numpy处理 img_np = img_denorm.cpu().numpy() img_np = np.transpose(img_np, (1, 2, 0)) # C x H x W -> H x W x C img_np = np.clip(img_np, 0, 1) * 255.0 # 添加高斯噪声 std_dev = level * 25 noise = np.random.normal(0, std_dev, img_np.shape) noisy_img = img_np + noise noisy_img = np.clip(noisy_img, 0, 255) # 转回tensor noisy_img = noisy_img / 255.0 noisy_img = np.transpose(noisy_img, (2, 0, 1)) # H x W x C -> C x H x W noisy_tensor = torch.from_numpy(noisy_img.astype(np.float32)).to(device) elif noise_type == 2: # 椒盐噪声 # 转为numpy处理 img_np = img_denorm.cpu().numpy() img_np = np.transpose(img_np, (1, 2, 0)) # C x H x W -> H x W x C img_np = np.clip(img_np, 0, 1) * 255.0 # 创建掩码 mask = np.random.random(img_np.shape[:2]) # 椒噪声 (黑点) img_np_copy = img_np.copy() img_np_copy[mask < level/2] = 0 # 盐噪声 (白点) img_np_copy[mask > 1 - level/2] = 255 # 转回tensor noisy_img = img_np_copy / 255.0 noisy_img = np.transpose(noisy_img, (2, 0, 1)) # H x W x C -> C x H x W noisy_tensor = torch.from_numpy(noisy_img.astype(np.float32)).to(device) elif noise_type == 3: # 泊松噪声 # 转为numpy处理 img_np = img_denorm.cpu().numpy() img_np = np.transpose(img_np, (1, 2, 0)) # C x H x W -> H x W x C img_np = np.clip(img_np, 0, 1) * 255.0 # 添加泊松噪声 lam = np.maximum(img_np / 255.0 * 10.0, 0.0001) noisy_img = np.random.poisson(lam) / 10.0 * 255.0 noisy_img = np.clip(noisy_img, 0, 255) # 转回tensor noisy_img = noisy_img / 255.0 noisy_img = np.transpose(noisy_img, (2, 0, 1)) # H x W x C -> C x H x W noisy_tensor = torch.from_numpy(noisy_img.astype(np.float32)).to(device) # 重新标准化 noisy_tensor_norm = (noisy_tensor - mean) / std # 更新数据 all_data[i] = noisy_tensor_norm # 记录噪声信息 noise_info['noise_types'].append(noise_type) noise_info['noise_levels'].append(level) noise_info['noise_indices'].append(i) # 保存添加噪声的样本索引 noise_indices = sorted(noise_info['noise_indices']) noise_index_path = os.path.join('..', 'dataset', 'noise_index.npy') os.makedirs(os.path.dirname(noise_index_path), exist_ok=True) np.save(noise_index_path, noise_indices) print(f"已保存噪声样本索引到 {noise_index_path}，共 {len(noise_indices)} 个样本") # 创建新的TensorDataset noisy_dataset = TensorDataset(all_data, all_targets) # 创建新的DataLoader noisy_trainloader = DataLoader( noisy_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=num_workers ) print(f"成功为{len(noise_info['noise_indices'])}/{len(all_data)} ({len(noise_info['noise_indices'])/len(all_data)*100:.1f}%)的样本添加噪声") return noisy_trainloader, testloader