def lgb_cv_tuning(
dataset_x: pd.DataFrame,
dataset_y: pd.DataFrame,
cv_index: List[Tuple[pd.RangeIndex]],
) -> Tuple[lgb.Booster, dict]:
train_dataset = lgb.Dataset(data=dataset_x, label=dataset_y)
lgbtuner = LightGBMTunerCV(
params={
"objective": "regression",
"metric": "rmse"
},
train_set=train_dataset,
folds=cv_index,
nfold=5,
verbose_eval=False,
num_boost_round=10000,
early_stopping_rounds=100,
return_cvbooster=True,
)
lgbtuner.run()
models = lgbtuner.get_best_booster().boosters
best_params = lgbtuner.best_params
print(" Params: ")
for key, value in best_params.items():
print(" {}: {}".format(key, value))
return models, best_params
def optimize(self, dtrain: lgbDataset, deval: lgbDataset):
# Define the base parameters
if self.objective == "binary":
params: Dict = {"objective": self.objective}
elif self.objective == "multiclass":
params: Dict = {
"objective": self.objective,
"metric": "multi_logloss"
}
elif self.objective == "regression":
params: Dict = {"objective": self.objective, "metric": "rmse"}
if self.verbose:
params["verbosity"] = 1
else:
params["verbosity"] = -1
if self.objective != "regression":
params["is_unbalance"] = self.is_unbalance
if self.objective == "multiclass":
params["num_class"] = self.num_class
# Reformat the data for LightGBM cross validation method
train_set = lgb.Dataset(
data=pd.concat([dtrain.data, deval.data]).reset_index(drop=True),
label=pd.concat([dtrain.label,
deval.label]).reset_index(drop=True),
categorical_feature=dtrain.categorical_feature,
free_raw_data=False,
)
train_index = range(len(dtrain.data))
valid_index = range(len(dtrain.data), len(train_set.data))
# Run the hyper-parameter tuning
self.tuner = LightGBMTunerCV(
params=params,
train_set=train_set,
folds=[(train_index, valid_index)],
verbose_eval=False,
num_boost_round=1000,
early_stopping_rounds=50,
)
self.tuner.run()
self.best = self.tuner.best_params
# since n_estimators is not among the params that Optuna optimizes we
# need to add it manually. We add a high value since it will be used
# with early_stopping_rounds
self.best["n_estimators"] = 1000 # type: ignore
class LGBOptimizerOptuna(object):
def __init__(
self,
objective: str = "binary",
verbose: bool = False,
):
self.objective = objective
self.verbose = verbose
self.best: Dict[str, Any] = {} # Best hyper-parameters
def optimize(self, dtrain: lgbDataset, deval: lgbDataset):
# Define the base parameters
params: Dict = {"objective": self.objective} # , "metric": "rmse"}
if self.verbose:
params["verbosity"] = 1
else:
params["verbosity"] = -1
# Reformat the data for LightGBM cross validation method
train_set = lgb.Dataset(
data=pd.concat([dtrain.data, deval.data]).reset_index(drop=True),
label=pd.concat([dtrain.label,
deval.label]).reset_index(drop=True),
categorical_feature=dtrain.categorical_feature,
free_raw_data=False,
)
train_index = range(len(dtrain.data))
valid_index = range(len(dtrain.data), len(train_set.data))
# Run the hyper-parameter tuning
self.tuner = LightGBMTunerCV(
params=params,
train_set=train_set,
folds=[(train_index, valid_index)],
verbose_eval=False,
num_boost_round=1000,
early_stopping_rounds=50,
)
self.tuner.run()
self.best = self.tuner.best_params
# since n_estimators is not among the params that Optuna optimizes we
# need to add it manually. We add a high value since it will be used
# with early_stopping_rounds
self.best["n_estimators"] = 1000 # type: ignore