Flatten features in the model instead of in pre-processing

This makes the MLP training a little more straightforward.

Also makes some error messages more useful and fixes the settings in train.py to account for newer models.

Fixes #39 (closed).