论文： Meta-Learning for Semi-Supervised Few-Shot Classification 2018ICLR

tieredImageNet是小样本分类任务数据集。与miniImagenet一样，它是ILSVRC-12的子集。但是，tieredImageNet代表了ILSVRC-12的更大子集（608个类，而miniImageNet则为100个类）。类似于将字符分组为字母的Omniglot一样，tieredImageNet将类别分为与ImageNet层次结构中较高级别的节点相对应的更广泛的类别。共有34个大类别(category)，每个类别包含10到30个类别(class)。这些分为20个训练，6个验证和8个测试类别。

与miniImageNet和Vinyals等人提出的randImageNet等其他替代方法不同，这可以确保所有训练类别都与测试类别充分区分开。。例如，在miniImagenet中，“pipe organ”属于训练集，“electrc guitar”属于测试，即使它们都是乐器。这种情况在tieredImageNet中不会发生，因为“musical instrument”是一个大类别（category），因此不会在训练集和测试集之间划分。这代表了更现实的几次尝试学习场景，因为一般而言，我们不能假设训练集与测试集中的类别相似。此外，tieredImageNet的分层结构对于可以利用类之间的分层关系的快速学习方法可能很有用。 下面是训练集，验证集，测试集对应的类别。 Train Categories: n02087551 (hound, hound dog), n02092468 (terrier), n02120997 (feline, felid), n02370806 (ungulate, hoofed mammal), n02469914 (primate), n01726692 (snake, ser- pent, ophidian), n01674216 (saurian), n01524359 (passerine, passeriform bird), n01844917 (aquatic bird), n04081844 (restraint, constraint), n03574816 (instrument), n03800933 (mu- sical instrument, instrument), n03125870 (craft), n04451818 (tool), n03414162 (game equipment), n03278248 (electronic equipment), n03419014 (garment), n03297735 (estab- lishment), n02913152 (building, edifice), n04014297 (protective covering, protective cover, protection). Validation Categories: n02098550 (sporting dog, gun dog), n03257877 (durables, durable goods, consumer durables), n03405265 (furnishing), n03699975 (machine), n03738472 (mechanism), n03791235 (motor vehicle, automotive vehicle), Test Categories: n02103406 (working dog), n01473806 (aquatic vertebrate), n02159955 (in- sect), n04531098 (vessel), n03839993 (obstruction, obstructor, obstructer, impediment, imped- imenta), n09287968 (geological formation, formation), n00020090 (substance), n15046900 (solid).