Many dataset implementations use walk_files
function which uses os.walk
which does not give us choice to make directory traversal in deterministic order.
This makes it harder to test dataset and causes issue on reproducibility. Dataset should have an option to be deterministic.