def keras_classifier_bayesian_opt(model_config, input_length):
bayesian_opt = KerasClassifierBayesianOpt(
model_config, keras_classifier_bayesian_space(), input_length)
bayesian_opt.nkfolds = 3
bayesian_opt.scoring = get_scorers(["AUC", "Accuracy"])
bayesian_opt.scoring_opt = "AUC"
bayesian_opt.low_is_better = False
bayesian_opt.n_trials = 30
return bayesian_opt
def xgboost_classifier_bayesian_opt(model_config):
bayesian_opt = XGBoostClassifierBayesianOpt(
model_config, xgboost_classifier_bayesian_space())
bayesian_opt.nkfolds = 3
bayesian_opt.scoring = get_scorers(["AUC", "Accuracy"])
bayesian_opt.scoring_opt = "AUC"
bayesian_opt.low_is_better = False
bayesian_opt.n_trials = 100
return bayesian_opt
def do_gridsearch(names,
classifiers,
grid_params,
x_train,
y_train,
nkfolds,
out_dirs,
ncores=-1):
"""Hyperparameter grid search for a list of classifiers
Given a list of classifiers, do a hyperparameter grid search based on a corresponding
set of parameters
Args:
names: iteratable of classifier names
classifiers: iterable of classifiers
grid_params: iterable of parameters used to perform the grid search
x_train: feature dataframe
y_train: targets dataframe
nkfolds: int, cross-validation generator or an iterable
out_dirs: Write parameters and pickle of summary dataframe
ncores: number of cores to distribute jobs to
Returns:
lists of grid search models, the best model and scoring dataframes
"""
logger = get_logger()
for clf_name, clf, gps, out_dir in zip(names, classifiers, grid_params,
out_dirs):
if not gps:
logger.info("Nothing to be done for grid search of model %s",
clf_name)
continue
logger.info("Grid search for model %s with following parameters:",
clf_name)
print_dict(gps)
# To work for probabilities. This will call model.decision_function or
# model.predict_proba as it is done for the nominal ROC curves as well to decide on the
# performance
scoring = get_scorers(gps["scoring"])
grid_search = GridSearchCV(clf,
gps["params"],
cv=nkfolds,
refit=gps["refit"],
scoring=scoring,
n_jobs=ncores,
verbose=2,
return_train_score=True)
grid_search.fit(x_train, y_train)
cvres = grid_search.cv_results_
# Save the results as soon as we have them in case something goes wrong later
# (would be quite unfortunate to loose grid search reults...)
out_file = osjoin(out_dir, "results.pkl")
pickle.dump(pd.DataFrame(cvres), openfile(out_file, "wb"), protocol=4)
# Parameters
dump_yaml_from_dict(gps, osjoin(out_dir, "parameters.yaml"))
savemodels((clf_name, ), (grid_search.best_estimator_, ), out_dir, "")