How to correctly define a Pytorch dataset and dataloader for data which is scattered accross disk, AND in each file?

Question

1. I have data scattered across several files, say f1, f2, f3 whose paths are p1, p2, p3.