【Python】XGBoost の使い方

train_df と test_df を用意して、関数に入れる。

#-------------------------
#  関数定義
#-------------------------
import xgboost as xgb

def run_xgb_kfold(train_df,test_df,features,target,folds,params):
    oof_preds = np.zeros(train_df.shape[0])
    sub_preds = np.zeros(test_df.shape[0])

    cv_list = []
    feature_imps = pd.DataFrame()
    for n_fold, (train_idx, valid_idx) in enumerate(folds.split(train_df[features], train_df['manufacturer'])):
        print ('FOLD:' + str(n_fold))
        
        train_x, train_y = train_df[features].iloc[train_idx], train_df[target].iloc[train_idx]
        valid_x, valid_y = train_df[features].iloc[valid_idx], train_df[target].iloc[valid_idx]
 
        dtrain = xgb.DMatrix(train_x, label=train_y)
        dval = xgb.DMatrix(valid_x, label=valid_y) 
        
        num_round = 30000
       #訓練データはdtrain、評価用のテストデータはdvalidと設定
        watchlist = [(dtrain, 'train'), (dval, 'eval')] 
        
        model = xgb.train(params,
                    dtrain,#訓練データ
                    num_round,#設定した学習回数
                    early_stopping_rounds=20,
                    evals=watchlist,
                    )
        
        oof_preds[valid_idx] = model.predict(dval)
        oof_cv = rmse(valid_y,  oof_preds[valid_idx])
        cv_list.append(oof_cv)
        print (cv_list)
        sub_preds += model.predict(xgb.DMatrix(test_df[features])) / folds.n_splits
 
        feature_imp = pd.DataFrame(sorted(zip(model.get_score(importance_type='gain'),features)), columns=['Value','Feature'])
        feature_imp['fold'] = n_fold
        feature_imps = pd.concat([feature_imps,feature_imp],axis=0)
        
    cv = rmse(train_df[target],  oof_preds)
    print('Full OOF RMSE %.6f' % cv)  
    
    train_df['prediction'] = oof_preds
    test_df['prediction'] = sub_preds
    
    return train_df,test_df,feature_imps

#-------------------------
#  XGBoost への適用
#-------------------------
train_df = df[df['flag']=='train']
train_df['target_col'] = np.log1p(train_df['target_col'])
test_df = df[df['flag']=='test']

target = 'target_col'
drop_features = ['id', 'target_col', 'description', 'flag'] + drop_feature_list
features = [f for f in train_df.columns if f not in drop_features]
print ('features:', len(features),features)

n_splits = 5
seed = 817
folds = StratifiedKFold(n_splits=n_splits, shuffle=True, random_state=seed)
params = {
        'objective': 'reg:squarederror',
        'silent':1, 
        'random_state':1234, 
        # 学習用の指標 (RMSE)
        'eval_metric': 'rmse',
    }

train_xgb,test_xgb,feature_imps = run_xgb_kfold(train_df,test_df,features,target,folds,params)