Several improvements needed:
- Output was mirroring WEKA file output, which includes lots of extra stuff – better to just kick out what the model says. Output should be given for one event per line, with the most probable label first followed by its probability, then the next most probable, followed by its probability, etc. This means the best label can be plucked out easily as the first label on each line, but a user has access to the full distribution if desired.
- ModelTrainer allowed perceptron option, but then wrote model to disk as a GISModel.
- ModelTrainer didn't pass options for cutoffs and number of iterations to the Perceptron trainer.
- ModelApplier usage didn't get printed when no arguments were given.