1) three separate datasubsets (keys in the hdf5 file), one for train, one for validation, one for testing
2) one large hdf5 file but three subslice keys (e.g. lists of indices in to the big data file, each stored as an hdf5 array), one for train, one for validation, one for testing
Two options for doing this:
1) three separate datasubsets (keys in the hdf5 file), one for train, one for validation, one for testing
2) one large hdf5 file but three subslice keys (e.g. lists of indices in to the big data file, each stored as an hdf5 array), one for train, one for validation, one for testing