Where to implement pre-processing in PyTorch Lightning (e.g. tokenizing input text)

Question

Is there a convention to implement some kind of predict() method in PyTorch Lightning that does pre-processing before performing the actual prediction using