Multiple changes to train.py

[zmek]

What needs to be done in short-term?

• 1. Add AUPRC metrics to model_metadata
• 2. Change MODEL__ED_ADMISSIONS__NAME to model_name
• 3. Create a train module, with submodules admissions (allow for submodule for discharges later)
• 4. Save class balance
• 5. Create a dedicated script for training

Note about other things we have discussed

create two scripts (one for ed-predictor to use, one for public data) that call this one, and pass in these parameters (OR turn main() into a function that is called by these external scripts:

• visits dataframe
• yet-to-arrive dataframe
• prediction_times - a list of tuples
• start_training_set
• start_validation_set
• start_test_set
• end_test_set
• grid
• ordinal_mappings
• exclude_from_training_data
• special_category_objects

And these are returned to the calling scripts

• 7 trained models (5 admissions models as a list, one specialty, one yet-to-arrive)
• model_metadata

The calling script would

• load the data
• apply training/test splits based on dates passed [possibly, although I've worked in some careful logic that makes sure any mrn is not included in more than one set, so I may retain the current approach, which is to pre-label each visit and then check at training stage that the modelling dates match the labelling in the dataset]
• retrieve special_category_objects depending on whether its uclh or other
• specify variables to be exclude_from_training_data (some columns need to be passed that are then removed, I believe, but I will check this)
• save models and metadata

[zmek]

• 1. Add AUPRC metrics to model_metadata
• 2. Change MODEL__ED_ADMISSIONS__NAME to model_name
• 3. Create a train module, with submodules admissions (allow for submodule for discharges later)
• 4. Save class balance
• 5. Create a dedicated script for training

4. has been done
5. done in commit afe908e