小样本方式加载Stanford_Cars数据集

准备数据

需要从数据集地址：https://ai.stanford.edu/~jkrause/cars/car_dataset.html下载3个压缩包，下载好之后解压

根据标签划分数据集

import scipy.io as scio
import os.path as osp

# 这个路径下有三个文件夹
dataset_path = "XXX"

train_img_path = osp.join(dataset_path, "cars_train")
test_img_path = osp.join(dataset_path, "cars_test")

label_path = osp.join(dataset_path, "car_devkit", "devkit")
train_label_path = osp.join(label_path, "cars_train_annos.mat")
test_label_path = osp.join(label_path, "cars_test_annos.mat")

img = 0

def devide_dataset_by_label(path):
    global img
    data = scio.loadmat(path)
    data = data["annotations"]
    data = data.squeeze()
    labels = np.zeros(data.shape[0])
    for number, label in enumerate(labels):
        if not osp.exists(osp.join(dataset_path, "classes", str(int(label)))):
            os.makedirs(osp.join(dataset_path, "classes", str(int(label))))
        os.system(
            f"cp {osp.abspath(osp.join(train_img_path, str(int(number+1)).zfill(5) + '.jpg'))} {osp.abspath(osp.join(dataset_path, 'classes', str(int(label))))}")
        img += 1

if __name__ == '__main__':
    devide_dataset_by_label(train_label_path)
    devide_dataset_by_label(test_label_path)
    print(f"devide {img} imgs")

一共给16185张图片分了类，和数据集官方数据一样

加载成dataloader

def get_Stanford_Cars_dataloader(mode="train", way=5, shot=2, query=10):
    if not osp.exists(osp.abspath(osp.join(dataset_path, "classes"))):
        devide_dataset_by_label(train_label_path)
        devide_dataset_by_label(test_label_path)
        print(f"devide {img} imgs")
    classes_path = osp.abspath(osp.join(dataset_path, "classes"))

    class_list = []
    for class_name in os.listdir(classes_path):
        if class_name.__contains__("DS_Store"):
            continue
        class_list.append(os.path.join(classes_path, class_name))
    class_names = []
    for i in class_list:
        if os.listdir(i).__len__() >= query + shot:
            class_names.append(i)

    train_class_lists = class_names[:int(class_names.__len__() * 0.6)]
    val_class_lists = class_names[int(class_names.__len__() * 0.6):int(class_names.__len__() * 0.8)]
    test_class_lists = class_names[int(class_names.__len__() * 0.8):]

    transforms = [partial(convert_dict, "class"),
                  partial(load_class_images, 64),
                  partial(extract_episode, shot, query)]

    transforms = compose(transforms)

    lists = []
    if mode == "train":
        lists = train_class_lists
    elif mode == "val":
        lists = val_class_lists
    elif mode == "test":
        lists = test_class_lists

    episode = int(len(lists) / way) + 1  # 加1防止有一类没学习

    ds = TransformDataset(ListDataset(lists),  # 先将list类型数的class_names处理成Dataset类的ListDataset
                          transforms)  # 然后对它使用transforms,对数据进行进一步处理

    sampler = EpisodicBatchSampler(len(ds), way, episode)  # 每个episode有10way，每个batch有30个episdoe

    dataloader = torch.utils.data.DataLoader(ds, batch_sampler=sampler, num_workers=0)

    return dataloader

全部代码

分享至码云