Optunaを用いて RF or lightGBM を最適化

個人的な備忘録

Boston住宅価格データを使用

model.py

# !pip install optuna lightgbm
from functools import partial

import optuna
import lightgbm as lgb

from sklearn.ensemble import RandomForestRegressor
from sklearn.model_selection import KFold
from sklearn.model_selection import cross_validate
from sklearn.metrics import mean_absolute_error
from sklearn import datasets

def objective(X, y, trial, clf = 'RandomForestRegressor'):
    """最小化する目的関数"""

    if clf == 'RandomForestRegressor':
        # RF のとき
        params = {
            'n_estimators': int(trial.suggest_loguniform('n_estimators', 1e+2, 1e+3)),
            'max_depth': int(trial.suggest_loguniform('max_depth', 2, 32)),
        }
        model = RandomForestRegressor(**params)


        kf = KFold(n_splits=5, shuffle=True, random_state=42)
        scores = cross_validate(model, X=X, y=y, cv=kf, n_jobs=-1, scoring='neg_mean_absolute_error')
        score = -1*scores['test_score'].mean()


    elif clf == 'LGB':
        # LGB のとき
        params = {
            'objective': 'regression',
            'max_bin': int(trial.suggest_int('max_bin', 255, 500)),
            'learning_rate': 0.05,
            'num_leaves': int(trial.suggest_int('num_leaves', 32, 128)),
            'metrics': 'mae'
        }

        lgb_train = lgb.Dataset(X, y)
        res = lgb.cv(params, lgb_train, num_boost_round=1000, early_stopping_rounds=10, nfold=5, shuffle=True, stratified=False, seed=42)
        score = res['l1-mean'][-1]

    return score

def main(): # 例としてボストンの住宅価格データを使用
    dataset = datasets.load_boston()
    X, y = dataset.data, dataset.target
    f = partial(objective, X, y)
    study = optuna.create_study()
    study.optimize(f, n_trials=30)
    print('params:', study.best_params)


if __name__ == '__main__':
    main()