deployment and pickle

[kegl]

If we want to be able to deploy trained models semi-automatically (e.g., but hand-selecting the model or models), we will need to solve the pickling issue. The first problem is that pickle doesn't like imp.load_source; @mehdidc found a hack by setting class.module = 'main'. But ultimately, dumping and loading depends on the model the user submitted (e.g., keras models have their own routines), so the routines have to be overridden in the user-submitted models. I think we should have super-classes that implement the most common routines, so users can inherit from them, and make it also possible to implement their own load and dump functions. This is by no means urgent or required for the RAMP (prototyping) to work, only if we want to deploy models, e.g. though an API, into production.