Vectorize pre-processing

[alexdriedger]

Pre-processing takes way too long because each training example is processed by itself. This should be vectorized and this will significantly improve pre-processing speed.

This is probably a necessary step before trying to process the 57 million files from the devs