best_params, trials = optimize_lgbm_hyperopt(
train_folds,
feature_set,
average_precision_score,
seed,
hyperopt_trials=args.lgbm_hyperopt_trials)
ap_score, models_best, oof_preds = fit_lgb(best_params, train_folds,
feature_set,
average_precision_score,
True, True)
print('CV score:', feature_set, ap_score)
test_res = tools.eval_models(metadata_test,
feature_set,
models_best,
mean_std_best=None,
impute=False,
model_type='lgbm')
features_suffix = '_'.join(feature_set)
y_test = metadata_test.Progressor.values.copy() > 0
ids = metadata_test.ID.values
sides = metadata_test.Side.values
results[f'preds_MOST_BL_all_{features_suffix}'] = (ids, sides, y_test,
test_res)
with open(os.path.join(args.save_dir, 'results_baselines_lgbm.pkl'),
'wb') as f:
pickle.dump(results, f)
results = {}
for feature_set in [['AGE', 'SEX', 'BMI'], [
'KL',
], ['AGE', 'SEX', 'BMI', 'SURG', 'INJ', 'WOMAC'],
['AGE', 'SEX', 'BMI', 'KL'],
['AGE', 'SEX', 'BMI', 'KL', 'SURG', 'INJ', 'WOMAC']]:
for regularize, model_type in zip([False, True],
['statsmodels', 'sklearn']):
models_best, mean_std_best, folds_gt, folds_preds = baselines.build_logreg_model(
train_folds, feature_set, seed, args.n_vals_c,
average_precision_score, regularize)
test_res = tools.eval_models(metadata_test,
feature_set,
models_best,
mean_std_best,
model_type=model_type)
features_suffix = '_'.join(feature_set)
y_test = metadata_test.Progressor.values.copy() > 0
ids = metadata_test.ID.values
sides = metadata_test.Side.values
print(
f'[{model_type} | {feature_set}] CV AUC {roc_auc_score(folds_gt, folds_preds):.5} |'
f' AP {average_precision_score(folds_gt, folds_preds):.5}')
print(
f'[{model_type} | {feature_set}] Test AUC {roc_auc_score(y_test, test_res):.5} |'