How to train through a stream? Or a way that alleviates large in-memory requirement for `model.train()` usage?

[niftylettuce]

If you have a large dataset, of perhaps 5 GB+, then it is not efficient to do an in-memory model.train(sets, labels) approach whereas sets and labels could consist of hundreds of thousands of Arrays. What is an alternative approach to stream the sets and labels to training?