any workaround to subset imagenet dataset from tfds instead of direct download it to disk?

Question

I want to train my custom model on imagenet datasets. For simplicity, I am interested 10/100 class classification task. But, direct downloading imagenet