ZipDatasets don't have an index therefore you can't sample properly

[calebrob6]

All of our samplers take an index that we use to find places that we'd like to sample from. When we create a ZipDataset there isn't a single index that we can use to sample from. When the component datasets of the ZipDatasets don't overlap spatially this is a problem. For example, this happens when I want to combine several splits from the ChesapeakeCVPR dataset into a larger dataset to train with.

This is a specific instance of some of the stuff in #86.