Dataloader注意RandomCrop会导致每次遍历的时候输入都不一样

def load_videos(train_seq_path ,Batch_size):
    train_transforms = transforms.Compose(
        [transforms.ToTensor(), transforms.CenterCrop((256, 256))] # 这里改成CenterCrop则每次遍历train_dataloader会保持数据一致
    )
    train_dataset = VideoFolder(
        train_seq_path,
        rnd_interval=False,
        rnd_temp_order=False,
        split="train",
        transform=train_transforms,
        max_frames = 2, #若max_frames = 4，则把每个batch的图片限定在了4
    )
    train_dataset = torch.utils.data.Subset(train_dataset, indices=range(4))
    train_dataloader = DataLoader(
        train_dataset,
        batch_size=Batch_size,
        num_workers=8,
        shuffle=False,
        # drop_last=True, #修改，丢弃最后一组
        pin_memory=(device == "cuda"),
    )
    total_data = len(train_dataloader.dataset)
    return train_dataloader, total_data