Faster data loading #34
Description
One of the bottlenecks currently is the data loading. Most of the time is spent on loading the data and applying the collate function. Speedup methods should be tested like preloading, the collate function should be made more efficient.
Libraries like ANTsPy claim:
ANTsPy includes blazing-fast IO (~40% faster than nibabel for loading Nifti images and converting them to numpy arrays)"
Replacing nilearn/nibabel by ANTsPy for loading images could potentially solve one of the issues to some extent.