機械学習

Tweet

ランダムフォレスト,CatBoost,LightGBMのサンプルコート

# coding: UTF-8
import os
import numpy as np
from catboost import CatBoost, CatBoostRegressor, CatBoostClassifier
from catboost import Pool
from sklearn.preprocessing import MinMaxScaler
from sklearn.model_selection import train_test_split
from sklearn.metrics import accuracy_score, roc_curve, roc_auc_score
from sklearn.metrics import mean_squared_error
from lightgbm import LGBMRegressor, LGBMClassifier
from sklearn.ensemble import RandomForestRegressor, RandomForestClassifier
from util import read_csv


def create_train_data():
    records = read_csv(r"data/train.csv")
    dataset = []
    labels = []

    for record in records:
        dataset.append(record[0:-1])
        labels.append(record[-1])
    # 学習データとテストデータを分ける
    train_dataset, eval_dataset, train_labels, eval_labels = train_test_split(dataset, labels, test_size=0.1, random_state=0)

    # 正規化する
    scaler  = MinMaxScaler(feature_range=(0, 1), copy=True)

    # 学習データの正規化
    scaler.fit(train_dataset)
    train_dataset = scaler.transform(train_dataset)

    # テストデータの正規化
    scaler.fit(eval_dataset)
    eval_dataset = scaler.transform(eval_dataset)

    return train_dataset, eval_dataset, train_labels, eval_labels


def load_test_data():
    test_data = read_csv(r"data/test.csv")
    # 正規化する
    scaler  = MinMaxScaler(feature_range=(0, 1), copy=True)

    # 学習データの正規化
    scaler.fit(test_data)
    test_data = scaler.transform(test_data)


    return test_data



def write_result_csv(file_name, y_pred):
    if not os.path.exists("results"):
        os.makedirs("results")

    with open(os.path.join("results", file_name), "w") as f:
        for i, y in enumerate(y_pred):
            if y < 0.:
                y = 0.
            if y > 1.:
                y = 1.
            f.write(f"{i},{y:.6f}\n")


def cat_boost_classifier(train_dataset, eval_dataset, train_labels, eval_labels, test_data):
    """
    catBoostによる2値分類
    """
    print("======start cat_boost_classifier======")
    # catBoost用のデータ作成
    train_pool = Pool(train_dataset, label=train_labels)
    eval_pool = Pool(eval_dataset, label=eval_labels)

    params = {
        'depth': 8,
        'early_stopping_rounds' : 500,
        'iterations' : 1000, 
        #'custom_loss' : ['Accuracy'],
        'eval_metric': 'AUC',
        'random_seed' : 42,
        #'use_best_model': True
    }

    model = CatBoostClassifier(**params)
    cab = model.fit(train_pool, eval_set=eval_pool)
    preds_train_acc = cab.predict(train_pool)

    #preds_train_acc = cab.predict_proba(train_pool)
    preds_train_acc = cab.predict(train_pool)
    preds_eval_acc = cab.predict(eval_pool)

    train_acc = accuracy_score(train_labels, preds_train_acc)
    eval_acc = accuracy_score(eval_labels, preds_eval_acc)
    print(f'train_acc: {train_acc:.3f} eval_acc: {eval_acc:.3f}')

    # 予測データの作成
    index_pred = cab.predict(test_data)
    test_prob = cab.predict_proba(test_data)[:, 1]

    write_result_csv("cat_boost_classifier_result.csv", test_prob)


def cat_boost_regressor(train_dataset, eval_dataset, train_labels, eval_labels, test_data):
    """
    catBoostによる回帰
    """
    print("======start cat_boost_regressor======")
    # catBoost用のデータ作成
    train_pool = Pool(train_dataset, label=train_labels)
    eval_pool = Pool(eval_dataset, label=eval_labels)

    params = {
        'early_stopping_rounds' : 100,
        'iterations' : 1000,
        'loss_function' :'MAE', 
        'random_seed' :42
    }

    model = CatBoostRegressor(**params)
    cab = model.fit(train_pool, eval_set=eval_pool)
    preds = cab.predict(eval_pool)

    eval_mse = mean_squared_error(eval_labels, preds)
    train_score = model.score(train_dataset, train_labels)
    eval_score = model.score(eval_dataset, eval_labels)
    print(f"MSE: {eval_mse:.3f} train_score: {train_score:.3f} eval_score:{eval_score:.3f}")

    # 予測データの作成
    test_preds = cab.predict(test_data)
    write_result_csv("cat_boost_regressor_result.csv", test_preds)


def light_gbm_classifier(train_dataset, eval_dataset, train_labels, eval_labels, test_data):
    """
    LightGBMによる2値分類
    """
    # LightGBM のハイパーパラメータ
    lgbm_params = {
        'boosting_type': 'gbdt',
        'objective': 'binary',
        'early_stopping_rounds':200,
        'num_leaves': 4,
        'learning_rate': 0.05,
        'bagging_freq': 5,   
        'lambda_l1': 2,
        'lambda_l2': 2,
        'n_estimators': 80,
        'min_data_in_leaf': 20,
        'seed': 20111019,
        'max_depth': 7
    }

    # 上記のパラメータでモデルを学習する
    model = LGBMClassifier(**lgbm_params)
    model.fit(train_dataset, train_labels, 
              eval_set=[(eval_dataset, eval_labels)])

    # 予測とスコア
    preds_train_acc = model.predict(train_dataset)
    preds_eval_acc = model.predict(eval_dataset)

    train_acc = accuracy_score(train_labels, preds_train_acc.round(0))
    eval_acc = accuracy_score(eval_labels, preds_eval_acc.round(0))
    print(f'train_acc: {train_acc:.3f} eval_acc: {eval_acc:.3f}')

    # 予測データの作成
    test_prob = model.predict_proba(test_data)[:, 1]
    write_result_csv("light_gbm_classifier_result.csv", test_prob)


def light_gbm_regressor(train_dataset, eval_dataset, train_labels, eval_labels, test_data):
    """
    LightGBMによる回帰
    """

    # LightGBM のハイパーパラメータ
    params = {
        'boosting_type': 'gbdt',
        'objective': 'regression',
        'metric': 'mse',
        'num_leaves': 64,
        'max_depth': 8,
        'min_data_in_leaf': 20,
        'early_stopping_rounds':100
    }

    # 上記のパラメータでモデルを学習する
    model = LGBMRegressor(**params)
    model.fit(train_dataset, train_labels,
              eval_set=[(eval_dataset, eval_labels)])

    # 検証用データを予測する
    preds = model.predict(eval_dataset)

    # スコアの表示
    eval_mse = mean_squared_error(eval_labels, preds)
    train_score = model.score(train_dataset, train_labels)
    eval_score = model.score(eval_dataset, eval_labels)
    print(f"MSE: {eval_mse:.3f} train_score: {train_score:.3f} eval_score:{eval_score:.3f}")

    # 予測データの作成
    test_preds = model.predict(test_data)
    write_result_csv("light_gbm_regressor_result.csv", test_preds)


def random_forest_classifier(train_dataset, eval_dataset, train_labels, eval_labels, test_data):
    """
    ランダムフォレストによる2値分類
    """
    params = {
        'max_depth': 8,
        'n_estimators': 1000
    }

    model = RandomForestClassifier(**params)
    model.fit(train_dataset, train_labels)

    # 検証用データを予測する
    preds = model.predict(eval_dataset)

    # 予測とスコア
    preds_train_acc = model.predict(train_dataset)
    preds_eval_acc = model.predict(eval_dataset)

    train_acc = accuracy_score(train_labels, preds_train_acc)
    eval_acc = accuracy_score(eval_labels, preds_eval_acc)
    print(f'train_acc: {train_acc:.3f} eval_acc: {eval_acc:.3f}')

    # AUCの計算
    preds_eval_auc = model.predict_proba(eval_dataset)[:, 1]
    auc = roc_auc_score(eval_labels, preds_eval_auc)

    print(f"val_auc: {auc:.4f}")

    # 予測データの作成
    test_prob = model.predict_proba(test_data)[:, 1]
    write_result_csv("random_forest_classifier_result.csv", test_prob)


def random_forest_regressor(train_dataset, eval_dataset, train_labels, eval_labels, test_data):
    """
    ランダムフォレストによる2値分類
    """
    params = {
        'max_depth': 8,
        'n_estimators': 1000
    }

    model = RandomForestRegressor(**params)
    model.fit(train_dataset, train_labels)

    # 検証用データを予測する
    preds = model.predict(eval_dataset)

    # スコアの表示
    eval_mse = mean_squared_error(eval_labels, preds)
    train_score = model.score(train_dataset, train_labels)
    eval_score = model.score(eval_dataset, eval_labels)
    print(f"MSE: {eval_mse:.3f} train_score: {train_score:.3f} eval_score:{eval_score:.3f}")

    # 予測データの作成
    test_preds = model.predict(test_data)
    write_result_csv("random_forest_regressor_result.csv", test_preds)

if __name__ == '__main__':

    train_dataset, eval_dataset, train_labels, eval_labels = create_train_data()
    test_data = load_test_data()
    # 使いたいモデルのコメントを外してください

    #cat_boost_classifier(train_dataset, eval_dataset, train_labels, eval_labels, test_data)
    #cat_boost_regressor(train_dataset, eval_dataset, train_labels, eval_labels, test_data)
    #light_gbm_classifier(train_dataset, eval_dataset, train_labels, eval_labels, test_data)
    #light_gbm_regressor(train_dataset, eval_dataset, train_labels, eval_labels, test_data)
    random_forest_classifier(train_dataset, eval_dataset, train_labels, eval_labels, test_data)
    #random_forest_regressor(train_dataset, eval_dataset, train_labels, eval_labels, test_data)