Dataloader call getitem as many times as indices in the current batch.
In case datasets support a list of indices in one call, or a native python slice object, add a getbatch (optional) to pytorch dataset class.
If dataloader sees the dataset has such a method implemented, make it fetch the batch corresping to the list of indices with one getbatch call rather than many getitem calls.
This saves the trouble of calling getitem many times in cases where for example a disk file has to be opened then closed for every call and the underlying data structure supports a slicing or lists of indices for example hdf5 files.