PyTorch中shuffle=False时每epoch图像顺序仍变化的问题

核心问题出在你使用的SubsetRandomSampler上：

• 这个采样器的设计逻辑就是每个epoch都会随机打乱传入的索引列表，以此实现随机采样。
• 另外，PyTorch的DataLoader有个规则：当手动指定了sampler参数时，shuffle参数会被直接忽略，所以你设置的shuffle=False根本没起作用。

你尝试添加generator参数无效，是因为SubsetRandomSampler的随机逻辑不依赖这个全局生成器，它内部有自己的随机机制，就算固定种子，它还是会在每个epoch重新打乱索引，只是打乱的结果固定，而非保持顺序不变。

两种解决思路 #

1. 改用Subset创建固定子集（推荐） #

直接通过Subset类生成固定的训练/测试子集，此时shuffle=False会正常生效，每个epoch的样本顺序完全一致：

def create_loader_from_data_set(data_set, n_samples, batch_size, num_workers, test_size=0.2):
    indices = list(range(len(data_set)))
    selected_indices = random.sample(indices, n_samples)

    train_indices, test_indices = train_test_split(selected_indices, test_size=test_size, random_state=42)

    # 创建固定子集，而非用采样器动态采样
    train_subset = torch.utils.data.Subset(data_set, train_indices)
    test_subset = torch.utils.data.Subset(data_set, test_indices)

    # 此时shuffle=True/False会正常控制是否打乱顺序
    train_loader = DataLoader(train_subset, batch_size=batch_size, num_workers=num_workers, shuffle=False)
    test_loader = DataLoader(test_subset, batch_size=batch_size, num_workers=num_workers, shuffle=False)
    return train_loader, test_loader

2. 替换为SequentialSampler #

如果你坚持要用采样器模式，把SubsetRandomSampler换成SequentialSampler，它会严格按照传入的索引顺序遍历，不会打乱：

def create_loader_from_data_set(data_set, n_samples, batch_size, num_workers, test_size=0.2):
    indices = list(range(len(data_set)))
    selected_indices = random.sample(indices, n_samples)

    train_indices, test_indices = train_test_split(selected_indices, test_size=test_size, random_state=42)

    # 用SequentialSampler按顺序遍历索引
    train_sampler = torch.utils.data.SequentialSampler(train_indices)
    test_sampler = torch.utils.data.SequentialSampler(test_indices)

    train_loader = DataLoader(data_set, batch_size=batch_size, num_workers=num_workers, sampler=train_sampler)
    test_loader = DataLoader(data_set, batch_size=batch_size, num_workers=num_workers, sampler=test_sampler)
    return train_loader, test_loader

内容的提问来源于stack exchange，提问作者mor hale