cifarnet

CIFARNet contains 200K images sampled from ImageNet-21K (Winter 2019 release), resized to 64x64, using coarse-grained labels that roughly match those of CIFAR-10. The exact ImageNet synsets used were:
```
{
"n02691156": 0, # airplane
"n02958343": 1, # automobile
"n01503061": 2, # bird
"n02121620": 3, # cat
"n02430045": 4, # deer
"n02083346": 5, # dog
"n01639765": 6, # frog
"n02374451": 7, # horse
"n04194289": 8, # ship
"n04490091": 9, # truck
}
```
The classes are balanced, and the dataset is pre-split into a training set of 190K images and a validation set of 10K images.

CIFARNet 数据集共包含20万张从2019冬季发布版ImageNet-21K中采样得到的图像，所有图像均被调整至64×64分辨率，其采用的粗粒度标签与CIFAR-10的标签大致匹配。本次所使用的具体ImageNet同义词集如下：

{
"n02691156": 0, # 飞机
"n02958343": 1, # 汽车
"n01503061": 2, # 鸟类
"n02121620": 3, # 猫科动物
"n02430045": 4, # 鹿科动物
"n02083346": 5, # 犬科动物
"n01639765": 6, # 蛙类
"n02374451": 7, # 马科动物
"n04194289": 8, # 船舶
"n04490091": 9, # 卡车
}

该数据集的类别分布均衡，且已预先划分为包含19万张图像的训练集与包含1万张图像的验证集。