Taxonomy integration

[gngdb]

We want to be able to train a hierarchical model that'll assign probabilities to superclasses and sub-classes with a whole set of Pylearn2 models. To do this, we need taxonomy.py to be integrated with the main codebase.

We need to be able to pass an option in run settings to transform the classes and image_fnames dictionary stored in the settings structure, and we have to make sure this is done in the dataset object. Unfortunately, settings is loaded before run settings. You'll have to decide how to deal with that and mention it here.

The transformation is fairly simple, but there are two types of transformation we want to be able to do:

• Create a superclass model: group examples under a superclass into just that superclass.
• Create a subclass model: pick a superclass and keep only the keys in the image_fnames dictionary that correspond to the superclass.

[gngdb]

Once this is done we're most of the way towards being able to train the models that make up the model described in this issue.

[gngdb]

Training of this model was succesful, was able to reach similar validation set levels of performance, but then overfit and only improved on the test set. Details should be in a notebook called Hierarchical Run.