How to train on batches?

It is a nice implementation. But assumes all data to be in memory. This  is not possible always, specially for the case of graph classification. Can you suggest methods for enabling the training with batches?