Allow usage of arbitrary regressor and selector objects from CLI

[dwysocki]

Currently we have a predefined list of supported regressor and selector objects, which can be accessed via the --regressor <regressor> and --selector <selector> CLI options. This approach restricts users to the ones we decided to include. It also means that if an object is added or deprecated in a future scikit-learn release, we will have to break compatibility between releases, or perform some gymnastics to support both.

I propose the following. We allow users to provide the name of any object to use as a regressor/selector, e.g. --regressor sklearn.linear_model.LassoCV. The following self-contained code snippet demonstrates how we might implement this.

from importlib import import_module # that's a mouthful

# this will be the `args` object that already exists in `plotypus.cli.get_args`
from argparse import Namespace
args = Namespace(regressor="sklearn.linear_model.LassoCV")

# separate module and object name
regressor_components = args.regressor.split(".")
regressor_module_name = ".".join(regressor_components[:-1])
regressor_object_name = regressor_components[-1]
# import the module and then the regressor object from it
regressor_module = import_module(regressor_module_name)
regressor_object = getattr(regressor_module, regressor_object_name)

This functionality could of course be wrapped in a utility function (perhaps plotypus.utils.import_object) for reuse on the selector and possibly other objects, but this makes for a nice demonstration.

The choice still remains to include some shortcuts like we already do, such as having LassoCV be a shortcut for sklearn.linear_model.LassoCV. This could be taken care of by a check like

if "." in args.regressor:
    regressor = utils.import_object(args.regressor)
elif args.regressor in regressor_shortcuts:
    regressor = regressor_shortcuts[args.regressor]
else:
    raise Exception("Undefined regressor: {}".format(args.regressor))

I think we should definitely go ahead with the new functionality, the only thing I'm iffy on is keeping the old functionality. Thoughts?

[dwysocki]

One caveat: our current implementation of Baart's criteria will not work here. It will have to be refactored as a legitimate selector object, as I had originally envisioned it, instead of an option to plotypus.preprocessing.Fourier. This is something I planned on doing eventually, anyway, so it's not a real issue.

EDIT: This now has its own issue, #98.

[dwysocki]

This issue also depends on #90 being completed, in order for it to work.