The preprocessing function -which takes up some times before every training step- can be trivially parallelized.