"""
# Modeling

- run many models from [imodels](https://github.com/csinva/imodels)
- extract out stable rules: screen for high predictive acc, look at what is kept
- build stable rules model (e.g. using RuleFit or Corels)
"""

import importlib
import random

import numpy as np
import pandas as pd
from autogluon.tabular import TabularDataset, TabularPredictor

import rulevetting
import rulevetting.api.util


def fit_models(train_data: pd.DataFrame, tune_data: pd.DataFrame, interpretable: bool = True):
    """Note: don't use this function, still depends on experimental autogluon dependencies

    Parameters
    ----------
    train_datas
    tune_data
    interpretable: bool
        Whether to fit interpretable models or standard models

    Returns
    -------
    predictor
    """
    train_data = TabularDataset(train_data)
    test_data = TabularDataset(tune_data)
    predictor = TabularPredictor(label='outcome',
                                 path=rulevetting.AUTOGLUON_CACHE_PATH,
                                 eval_metric='roc_auc')
    kwargs = dict(
        verbosity=2,
        time_limit=30,
    )
    if interpretable:
        predictor.fit(train_data, presets='interpretable', **kwargs)
        print(predictor.interpretable_models_summary())
    else:
        predictor.fit(train_data, **kwargs)
    return predictor


if __name__ == '__main__':
    project_ids = rulevetting.api.util.get_project_ids()
    for project_id in project_ids:
        np.random.seed(0)
        random.seed(0)
        print('fitting on', project_id)
        project_module_name = f'rulevetting.projects.{project_id}.dataset'
        module = importlib.import_module(project_module_name)
        dset = module.Dataset()
        df_train, df_tune, df_test = dset.get_data(load_csvs=True)
        predictor = fit_models(df_train, df_tune)
        print(predictor)