'layers': 3,
# 샘플을 위해 빨리 끝나도록 설정
'nb_epoch': 5, # 1000
'patience': 10,
'dropout': 0.5,
'units': 512,
}
# 특징 지정
features = [f'feat_{i}' for i in range(1, 94)]
# xgboost에 의한 학습 및 예측
runner = Runner('xgb1', ModelXGB, features, params_xgb)
runner.run_train_cv()
runner.run_predict_cv()
Submission.create_submission('xgb1')
# 신경망에 의한 학습 예측
runner = Runner('nn1', ModelNN, features, params_nn)
runner.run_train_cv()
runner.run_predict_cv()
Submission.create_submission('nn1')
'''
# (참고）xgboost를 통한 학습 및 예측 - 학습 데이터 전체를 사용하는 경우
runner = Runner('xgb1-train-all', ModelXGB, features, params_xgb_all)
runner.run_train_all()
runner.run_test_all()
Submission.create_submission('xgb1-train-all')
'''
use_feature_name = runner.get_feature_name() # 今回の学習で使用する特徴量名を取得
# モデルのconfigをjsonで保存
value_list = [features, use_feature_name, model_params, cv, setting]
save_model_config(key_list, value_list, dir_name, run_name)
# runner.visualize_corr() # 相関係数を可視化して保存
if cv.get('method') == 'None':
runner.run_train_all() # 全データで学習
runner.run_predict_all() # 推論
else:
runner.run_train_cv() # 学習
ModelLGB.calc_feature_importance(dir_name, run_name, use_feature_name) # feature_importanceを計算
runner.run_predict_cv() # 推論
Submission.create_submission(run_name, dir_name, setting.get('target')) # submit作成
# ######################################################
# 学習・推論 xgboost ###################################
# run nameの設定
run_name = 'xgb'
run_name = run_name + suffix
dir_name = MODEL_DIR_NAME + run_name + '/'
# exist_check(MODEL_DIR_NAME, run_name) # 実行可否確認
my_makedirs(dir_name) # runディレクトリの作成。ここにlogなどが吐かれる
# 諸々の設定
setting = {
def main(mode='prd', create_features=True, model_type='lgb') -> str:
# confirm('***** mode:{}, create_feature:{} '.format(str(mode), str(create_features)))
now = datetime.datetime.now()
suffix = now.strftime("_%m%d%H%M")
# CVの設定.methodは[KFold, StratifiedKFold ,GroupKFold, StratifiedGroupKFold, CustomTimeSeriesSplitter, TrainTestSplit]から選択可能
# CVしない場合（全データで学習させる場合）はmethodに'None'を設定
# StratifiedKFold or GroupKFold or StratifiedGroupKFold の場合はcv_target_gr, cv_target_sfに対象カラム名を設定する
# TrainTestSplitの場合はtest_sizeにtest setの割合を設定する
cv = {
'method': 'StratifiedKFold',
'n_splits': 5,
'random_state': 42,
'shuffle': True,
'cv_target_gr': 'chip_id',
'cv_target_sf': 'target'
}
# run nameの設定
run_name = 'cat'
run_name = run_name + '_' + cv.get('method') + suffix + 'depth5_pl'
dir_name = MODEL_DIR_NAME + run_name + '/'
exist_check(MODEL_DIR_NAME, run_name)
my_makedirs(dir_name)
# 諸々の設定
setting = {
'run_name': run_name, # run名
'feature_directory': FEATURE_DIR_NAME, # 特徴量の読み込み先ディレクトリ
'target': 'target', # 目的変数
'calc_shap': False, # shap値を計算するか否か
'save_train_pred':
True # trainデータでの推論値を保存するか否か（trainデータでの推論値を特徴量として加えたい場合はTrueに設定する）
}
# モデルのパラメータ
model_params = {
'loss_function': 'Logloss',
'eval_metric': 'AUC',
'learning_rate': 0.05,
'iterations': 4000,
'early_stopping_rounds': 400,
'random_seed': 42,
'verbose_eval': 100,
'depth': 5
}
features = [
# 'chip_id',
'exc_wl',
# 'layout_a',
# 'layout_x',
# 'layout_y',
'distance',
'distance_x',
'distance_y',
'pos_x',
'params0',
'params1',
'params2',
'params3',
'params4',
'params5',
'params6',
# --- TODO: paramsの交互作用特徴量
'params0_multi_rms',
'params0_divid_rms',
# 'params0_plus_rms',
# 'params0_minus_rms',
'params0_multi_beta',
'params0_divid_beta',
# 'params0_plus_beta',
# 'params0_minus_beta',
'params1_multi_rms',
'params1_divid_rms',
# 'params1_plus_rms',
# 'params1_minus_rms',
'params1_multi_beta',
'params1_divid_beta',
# 'params1_plus_beta',
# 'params1_minus_beta',
'params2_multi_rms',
'params2_divid_rms',
# 'params2_plus_rms',
# 'params2_minus_rms',
'params2_multi_beta',
'params2_divid_beta',
# 'params2_plus_beta',
# 'params2_minus_beta',
'params3_multi_rms',
'params3_divid_rms',
# 'params3_plus_rms',
# 'params3_minus_rms',
'params3_multi_beta',
'params3_divid_beta',
# 'params3_plus_beta',
# 'params3_minus_beta',
'params4_multi_rms',
'params4_divid_rms',
# 'params4_plus_rms',
# 'params4_minus_rms',
'params4_multi_beta',
'params4_divid_beta',
# 'params4_plus_beta',
# 'params4_minus_beta',
'params5_multi_rms',
'params5_divid_rms',
# 'params5_plus_rms',
# 'params5_minus_rms',
'params5_multi_beta',
'params5_divid_beta',
# 'params5_plus_beta',
# 'params5_minus_beta',
'params6_multi_rms',
'params6_divid_rms',
# 'params6_plus_rms',
# 'params6_minus_rms',
'params6_multi_beta',
'params6_divid_beta',
# 'params6_plus_beta',
# 'params6_minus_beta',
'params0_multi_params1',
'params0_divid_params1',
# 'params0_plus_params1',
# 'params0_minus_params1',
'params0_multi_params2',
'params0_divid_params2',
# 'params0_plus_params2',
# 'params0_minus_params2',
'params0_multi_params3',
'params0_divid_params3',
# 'params0_plus_params3',
# 'params0_minus_params3',
'params0_multi_params4',
'params0_divid_params4',
# 'params0_plus_params4',
# 'params0_minus_params4',
'params0_multi_params5',
'params0_divid_params5',
# 'params0_plus_params5',
# 'params0_minus_params5',
'params0_multi_params6',
'params0_divid_params6',
# 'params0_plus_params6',
# 'params0_minus_params6',
'params1_multi_params0',
'params1_divid_params0',
# 'params1_plus_params0',
# 'params1_minus_params0',
'params1_multi_params2',
'params1_divid_params2',
# 'params1_plus_params2',
# 'params1_minus_params2',
'params1_multi_params3',
'params1_divid_params3',
# 'params1_plus_params3',
# 'params1_minus_params3',
'params1_multi_params4',
'params1_divid_params4',
# 'params1_plus_params4',
# 'params1_minus_params4',
'params1_multi_params5',
'params1_divid_params5',
# 'params1_plus_params5',
# 'params1_minus_params5',
'params1_multi_params6',
'params1_divid_params6',
# 'params1_plus_params6',
# 'params1_minus_params6',
'params2_multi_params0',
'params2_divid_params0',
# 'params2_plus_params0',
# 'params2_minus_params0',
'params2_multi_params1',
'params2_divid_params1',
# 'params2_plus_params1',
# 'params2_minus_params1',
'params2_multi_params3',
'params2_divid_params3',
# 'params2_plus_params3',
# 'params2_minus_params3',
'params2_multi_params4',
'params2_divid_params4',
# 'params2_plus_params4',
# 'params2_minus_params4',
'params2_multi_params5',
'params2_divid_params5',
# 'params2_plus_params5',
# 'params2_minus_params5',
# 'params2_absminus_params5',
'params2_multi_params6',
'params2_divid_params6',
# 'params2_plus_params6',
# 'params2_minus_params6',
'params3_multi_params0',
'params3_divid_params0',
# 'params3_plus_params0',
# 'params3_minus_params0',
'params3_multi_params1',
'params3_divid_params1',
# 'params3_plus_params1',
# 'params3_minus_params1',
'params3_multi_params2',
'params3_divid_params2',
# 'params3_plus_params2',
# 'params3_minus_params2',
'params3_multi_params4',
'params3_divid_params4',
# 'params3_plus_params4',
# 'params3_minus_params4',
'params3_multi_params5',
'params3_divid_params5',
# 'params3_plus_params5',
# 'params3_minus_params5',
'params3_multi_params6',
'params3_divid_params6',
# 'params3_plus_params6',
# 'params3_minus_params6',
'params4_multi_params0',
'params4_divid_params0',
# 'params4_plus_params0',
# 'params4_minus_params0',
'params4_multi_params1',
'params4_divid_params1',
# 'params4_plus_params1',
# 'params4_minus_params1',
'params4_multi_params2',
'params4_divid_params2',
# 'params4_plus_params2',
# 'params4_minus_params2',
'params4_multi_params3',
'params4_divid_params3',
# 'params4_plus_params3',
# 'params4_minus_params3',
'params4_multi_params5',
'params4_divid_params5',
# 'params4_plus_params5',
# 'params4_minus_params5',
'params4_multi_params6',
'params4_divid_params6',
# 'params4_plus_params6',
# 'params4_minus_params6',
'params5_multi_params0',
'params5_divid_params0',
# 'params5_plus_params0',
# 'params5_minus_params0',
'params5_multi_params1',
'params5_divid_params1',
# 'params5_plus_params1',
# 'params5_minus_params1',
'params5_multi_params2',
'params5_divid_params2',
# 'params5_plus_params2',
# 'params5_minus_params2',
'params5_multi_params3',
'params5_divid_params3',
# 'params5_plus_params3',
# 'params5_minus_params3',
'params5_multi_params4',
'params5_divid_params4',
# 'params5_plus_params4',
# 'params5_minus_params4',
'params5_multi_params6',
'params5_divid_params6',
# 'params5_plus_params6',
# 'params5_minus_params6',
'params6_multi_params0',
'params6_divid_params0',
# 'params6_plus_params0',
# 'params6_minus_params0',
'params6_multi_params1',
'params6_divid_params1',
# 'params6_plus_params1',
# 'params6_minus_params1',
'params6_multi_params2',
'params6_divid_params2',
# 'params6_plus_params2',
# 'params6_minus_params2',
'params6_multi_params3',
'params6_divid_params3',
# 'params6_plus_params3',
# 'params6_minus_params3',
'params6_multi_params4',
'params6_divid_params4',
# 'params6_plus_params4',
# 'params6_minus_params4',
'params6_multi_params5',
'params6_divid_params5',
# 'params6_plus_params5',
# 'params6_minus_params5',
'rms',
'beta',
# --- TODO: スペクトル波長データ
'intensity_max',
'intensity_min',
'intensity_mean',
'intensity_std',
'intensity_sum',
'intensity_median',
'intensity_amplitude_v',
# 'intensity_max_log1p',
# 'intensity_mean_log1p',
# 'intensity_std_log1p',
# 'intensity_sum_log1p',
# 'intensity_median_log1p',
# 'intensity_amplitude_v_log1p',
'intensity_q10',
'intensity_q25',
'intensity_q50',
'intensity_q75',
'intensity_q80',
'intensity_q85',
'intensity_q90',
'intensity_max_minus_q90',
'intensity_max_minus_q85',
'intensity_max_minus_q80',
'intensity_max_minus_q75',
'intensity_max_minus_q50',
'intensity_max_multi_q90',
'intensity_max_multi_q85',
'intensity_max_multi_q80',
'intensity_max_multi_q75',
'intensity_max_multi_q50',
'intensity_max_divid_q90',
'intensity_max_divid_q85',
'intensity_max_divid_q80',
'intensity_max_divid_q75',
'intensity_max_divid_q50',
'intensity_q85_divid_q90',
'intensity_q80_divid_q90',
'intensity_q75_divid_q90',
'intensity_q50_divid_q90',
# --- ピーク&半値幅のデータ
'fwhm_09',
# 'num_peak_09',
'fwhm_088',
# 'num_peak_088',
'fwhm_086',
# 'num_peak_086',
'fwhm_084',
# 'num_peak_084',
'fwhm_082',
# 'num_peak_082',
'fwhm_08',
# 'num_peak_08',
'fwhm_078',
# 'num_peak_078',
'fwhm_076',
# 'num_peak_076',
'fwhm_074',
# 'num_peak_074',
'fwhm_072',
# 'num_peak_072',
'fwhm_07',
# 'num_peak_07',
'fwhm_068',
# 'num_peak_068',
'fwhm_066',
# 'num_peak_066',
'fwhm_064',
# 'num_peak_064',
'fwhm_062',
# 'num_peak_062',
'fwhm_06',
# 'num_peak_06',
'fwhm_058',
# 'num_peak_058',
'fwhm_056',
# 'num_peak_056',
'fwhm_054',
# 'num_peak_054',
'fwhm_052',
# 'num_peak_052',
'fwhm_05',
# 'num_peak_05',
'fwhm_divid_num_peak_09',
'fwhm_mult_num_peak_09',
'fwhm_divid_num_peak_088',
'fwhm_mult_num_peak_088',
'fwhm_divid_num_peak_086',
'fwhm_mult_num_peak_086',
'fwhm_divid_num_peak_084',
'fwhm_mult_num_peak_084',
'fwhm_divid_num_peak_082',
'fwhm_mult_num_peak_082',
'fwhm_divid_num_peak_08',
'fwhm_mult_num_peak_08',
'fwhm_divid_num_peak_078',
'fwhm_mult_num_peak_078',
'fwhm_divid_num_peak_076',
'fwhm_mult_num_peak_076',
'fwhm_divid_num_peak_074',
'fwhm_mult_num_peak_074',
'fwhm_divid_num_peak_072',
'fwhm_mult_num_peak_072',
'fwhm_divid_num_peak_07',
'fwhm_mult_num_peak_07',
# 'fwhm_divid_num_peak_068',
# 'fwhm_mult_num_peak_068',
# 'fwhm_divid_num_peak_066',
# 'fwhm_mult_num_peak_066',
# 'fwhm_divid_num_peak_064',
# 'fwhm_mult_num_peak_064',
# 'fwhm_divid_num_peak_062',
# 'fwhm_mult_num_peak_062',
# 'fwhm_divid_num_peak_06',
# 'fwhm_mult_num_peak_06',
# 'fwhm_divid_num_peak_058',
# 'fwhm_mult_num_peak_058',
# 'fwhm_divid_num_peak_056',
# 'fwhm_mult_num_peak_056',
# 'fwhm_divid_num_peak_054',
# 'fwhm_mult_num_peak_054',
# 'fwhm_divid_num_peak_052',
# 'fwhm_mult_num_peak_052',
# 'fwhm_divid_num_peak_05',
# 'fwhm_mult_num_peak_05',
# --- TODO: 波形の圧縮データ
'dc_umap1',
'dc_umap2',
'dc_tsne1',
'dc_tsne2',
# --- TODO: カテゴリごとの平均値と自身との差分
# 'diff_chip_id_mean_params0',
# 'diff_chip_id_mean_params1',
# 'diff_chip_id_mean_params2',
# 'diff_chip_id_mean_params3',
# 'diff_chip_id_mean_params4',
# 'diff_chip_id_mean_params5',
# 'diff_chip_id_mean_params6',
# 'diff_chip_id_mean_rms',
# 'diff_chip_id_mean_beta',
# 'diff_chip_id_mean_intensity_amplitude_v',
# 'diff_chip_id_mean_fwhm_09',
# # 'diff_chip_id_mean_num_peak_09',
# 'diff_chip_id_mean_fwhm_088',
# # 'diff_chip_id_mean_num_peak_088',
# 'diff_chip_id_mean_fwhm_086',
# # 'diff_chip_id_mean_num_peak_086',
# 'diff_chip_id_mean_fwhm_084',
# # 'diff_chip_id_mean_num_peak_084',
# 'diff_chip_id_mean_fwhm_082',
# # 'diff_chip_id_mean_num_peak_082',
# 'diff_chip_id_mean_fwhm_08',
# # 'diff_chip_id_mean_num_peak_08',
# 'diff_chip_id_mean_fwhm_078',
# # 'diff_chip_id_mean_num_peak_078',
# 'diff_chip_id_mean_fwhm_076',
# # 'diff_chip_id_mean_num_peak_076',
# 'diff_chip_id_mean_fwhm_074',
# # 'diff_chip_id_mean_num_peak_074',
# 'diff_chip_id_mean_fwhm_072',
# # 'diff_chip_id_mean_num_peak_072',
# 'diff_chip_id_mean_fwhm_07',
# # 'diff_chip_id_mean_num_peak_07',
# 'diff_chip_id_mean_fwhm_068',
# # 'diff_chip_id_mean_num_peak_068',
# 'diff_chip_id_mean_fwhm_066',
# # 'diff_chip_id_mean_num_peak_066',
# 'diff_chip_id_mean_fwhm_064',
# # 'diff_chip_id_mean_num_peak_064',
# 'diff_chip_id_mean_fwhm_062',
# # 'diff_chip_id_mean_num_peak_062',
# 'diff_chip_id_mean_fwhm_06',
# # 'diff_chip_id_mean_num_peak_06',
# 'diff_chip_id_mean_fwhm_058',
# # 'diff_chip_id_mean_num_peak_058',
# 'diff_chip_id_mean_fwhm_056',
# # 'diff_chip_id_mean_num_peak_056',
# 'diff_chip_id_mean_fwhm_054',
# # 'diff_chip_id_mean_num_peak_054',
# 'diff_chip_id_mean_fwhm_052',
# # 'diff_chip_id_mean_num_peak_052',
# 'diff_chip_id_mean_fwhm_05',
# # 'diff_chip_id_mean_num_peak_05',
# # 'diff_chip_id_mean_intensity_max',
# # 'diff_chip_id_mean_intensity_min',
# # 'diff_chip_id_mean_intensity_q90',
# # 'diff_chip_id_mean_intensity_q75',
'diff_exc_wl_mean_params0',
'diff_exc_wl_mean_params1',
'diff_exc_wl_mean_params2',
'diff_exc_wl_mean_params3',
'diff_exc_wl_mean_params4',
'diff_exc_wl_mean_params5',
'diff_exc_wl_mean_params6',
'diff_exc_wl_mean_rms',
'diff_exc_wl_mean_beta',
'diff_exc_wl_mean_intensity_amplitude_v',
'diff_exc_wl_mean_fwhm_09',
# 'diff_exc_wl_mean_num_peak_09',
'diff_exc_wl_mean_fwhm_088',
# 'diff_exc_wl_mean_num_peak_088',
'diff_exc_wl_mean_fwhm_086',
# 'diff_exc_wl_mean_num_peak_086',
'diff_exc_wl_mean_fwhm_084',
# 'diff_exc_wl_mean_num_peak_084',
'diff_exc_wl_mean_fwhm_082',
# 'diff_exc_wl_mean_num_peak_082',
'diff_exc_wl_mean_fwhm_08',
# 'diff_exc_wl_mean_num_peak_08',
'diff_exc_wl_mean_fwhm_078',
# 'diff_exc_wl_mean_num_peak_078',
'diff_exc_wl_mean_fwhm_076',
# 'diff_exc_wl_mean_num_peak_076',
'diff_exc_wl_mean_fwhm_074',
# 'diff_exc_wl_mean_num_peak_074',
'diff_exc_wl_mean_fwhm_072',
# 'diff_exc_wl_mean_num_peak_072',
'diff_exc_wl_mean_fwhm_07',
# 'diff_exc_wl_mean_num_peak_07',
'diff_exc_wl_mean_fwhm_068',
# 'diff_exc_wl_mean_num_peak_068',
'diff_exc_wl_mean_fwhm_066',
# 'diff_exc_wl_mean_num_peak_066',
'diff_exc_wl_mean_fwhm_064',
# 'diff_exc_wl_mean_num_peak_064',
'diff_exc_wl_mean_fwhm_062',
# 'diff_exc_wl_mean_num_peak_062',
'diff_exc_wl_mean_fwhm_06',
# 'diff_exc_wl_mean_num_peak_06',
'diff_exc_wl_mean_fwhm_058',
# 'diff_exc_wl_mean_num_peak_058',
'diff_exc_wl_mean_fwhm_056',
# 'diff_exc_wl_mean_num_peak_056',
'diff_exc_wl_mean_fwhm_054',
# 'diff_exc_wl_mean_num_peak_054',
'diff_exc_wl_mean_fwhm_052',
# 'diff_exc_wl_mean_num_peak_052',
'diff_exc_wl_mean_fwhm_05',
# # 'diff_exc_wl_mean_num_peak_05',
# # 'diff_exc_wl_mean_intensity_max',
# # 'diff_exc_wl_mean_intensity_min',
# # 'diff_exc_wl_mean_intensity_q90',
# # 'diff_exc_wl_mean_intensity_q75',
# 'diff_layout_a_mean_params0',
# 'diff_layout_a_mean_params1',
# 'diff_layout_a_mean_params2',
# 'diff_layout_a_mean_params3',
# 'diff_layout_a_mean_params4',
# 'diff_layout_a_mean_params5',
# 'diff_layout_a_mean_params6',
# 'diff_layout_a_mean_rms',
# 'diff_layout_a_mean_beta',
# 'diff_layout_a_mean_intensity_amplitude_v',
# 'diff_layout_a_mean_fwhm_09',
# # 'diff_layout_a_mean_num_peak_09',
# 'diff_layout_a_mean_fwhm_088',
# # 'diff_layout_a_mean_num_peak_088',
# 'diff_layout_a_mean_fwhm_086',
# # 'diff_layout_a_mean_num_peak_086',
# 'diff_layout_a_mean_fwhm_084',
# # 'diff_layout_a_mean_num_peak_084',
# 'diff_layout_a_mean_fwhm_082',
# # 'diff_layout_a_mean_num_peak_082',
# 'diff_layout_a_mean_fwhm_08',
# # 'diff_layout_a_mean_num_peak_08',
# 'diff_layout_a_mean_fwhm_078',
# # 'diff_layout_a_mean_num_peak_078',
# 'diff_layout_a_mean_fwhm_076',
# # 'diff_layout_a_mean_num_peak_076',
# 'diff_layout_a_mean_fwhm_074',
# # 'diff_layout_a_mean_num_peak_074',
# 'diff_layout_a_mean_fwhm_072',
# # 'diff_layout_a_mean_num_peak_072',
# 'diff_layout_a_mean_fwhm_07',
# # 'diff_layout_a_mean_num_peak_07',
# 'diff_layout_a_mean_fwhm_068',
# # 'diff_layout_a_mean_num_peak_068',
# 'diff_layout_a_mean_fwhm_066',
# # 'diff_layout_a_mean_num_peak_066',
# 'diff_layout_a_mean_fwhm_064',
# # 'diff_layout_a_mean_num_peak_064',
# 'diff_layout_a_mean_fwhm_062',
# # 'diff_layout_a_mean_num_peak_062',
# 'diff_layout_a_mean_fwhm_06',
# # 'diff_layout_a_mean_num_peak_06',
# 'diff_layout_a_mean_fwhm_058',
# # 'diff_layout_a_mean_num_peak_058',
# 'diff_layout_a_mean_fwhm_056',
# # 'diff_layout_a_mean_num_peak_056',
# 'diff_layout_a_mean_fwhm_054',
# # 'diff_layout_a_mean_num_peak_054',
# 'diff_layout_a_mean_fwhm_052',
# # 'diff_layout_a_mean_num_peak_052',
# 'diff_layout_a_mean_fwhm_05',
# # 'diff_layout_a_mean_num_peak_05',
# # 'diff_layout_a_mean_intensity_max',
# # 'diff_layout_a_mean_intensity_min',
# # 'diff_layout_a_mean_intensity_q90',
# # 'diff_layout_a_mean_intensity_q75',
# --- TODO: tsfreshで生成した特徴
'intensity__ratio_beyond_r_sigma__r_10',
'intensity__ar_coefficient__coeff_2__k_10',
'intensity__change_quantiles__f_agg_"var"__isabs_True__qh_1.0__ql_0.8',
'intensity__change_quantiles__f_agg_"var"__isabs_True__qh_1.0__ql_0.6',
'intensity__change_quantiles__f_agg_"var"__isabs_True__qh_1.0__ql_0.4',
'intensity__change_quantiles__f_agg_"var"__isabs_False__qh_1.0__ql_0.8',
'intensity__change_quantiles__f_agg_"var"__isabs_True__qh_1.0__ql_0.2',
'intensity__change_quantiles__f_agg_"var"__isabs_True__qh_1.0__ql_0.0',
'intensity__change_quantiles__f_agg_"var"__isabs_False__qh_1.0__ql_0.6',
'intensity__change_quantiles__f_agg_"var"__isabs_False__qh_1.0__ql_0.4',
'intensity__change_quantiles__f_agg_"var"__isabs_False__qh_1.0__ql_0.0',
'intensity__cid_ce__normalize_False',
'intensity__change_quantiles__f_agg_"var"__isabs_False__qh_1.0__ql_0.2',
'intensity__fft_coefficient__attr_"abs"__coeff_55',
'intensity__fft_coefficient__attr_"abs"__coeff_56',
'intensity__large_standard_deviation__r_0.1',
'intensity__agg_linear_trend__attr_"stderr"__chunk_len_5__f_agg_"var"',
'intensity__fft_coefficient__attr_"abs"__coeff_64',
'intensity__fft_coefficient__attr_"abs"__coeff_57',
'intensity__fft_coefficient__attr_"abs"__coeff_51',
'intensity__fft_coefficient__attr_"abs"__coeff_52',
'intensity__fft_coefficient__attr_"abs"__coeff_66',
'intensity__fft_coefficient__attr_"abs"__coeff_68',
'intensity__fft_coefficient__attr_"abs"__coeff_59',
'intensity__fft_coefficient__attr_"abs"__coeff_61',
'intensity__fft_coefficient__attr_"abs"__coeff_53',
'intensity__change_quantiles__f_agg_"mean"__isabs_True__qh_1.0__ql_0.8',
'intensity__fft_coefficient__attr_"abs"__coeff_54',
'intensity__fft_coefficient__attr_"abs"__coeff_65',
'intensity__partial_autocorrelation__lag_2',
'intensity__fft_coefficient__attr_"abs"__coeff_63',
'intensity__fft_coefficient__attr_"abs"__coeff_62',
'intensity__fft_coefficient__attr_"abs"__coeff_71',
'intensity__ar_coefficient__coeff_1__k_10',
'intensity__fft_coefficient__attr_"abs"__coeff_50',
'intensity__fft_coefficient__attr_"abs"__coeff_49',
'intensity__fft_coefficient__attr_"abs"__coeff_48',
'intensity__fft_coefficient__attr_"abs"__coeff_58',
'intensity__fft_coefficient__attr_"abs"__coeff_46',
'intensity__fft_coefficient__attr_"abs"__coeff_69',
'intensity__fft_coefficient__attr_"abs"__coeff_67',
'intensity__change_quantiles__f_agg_"mean"__isabs_True__qh_1.0__ql_0.6',
'intensity__fft_coefficient__attr_"abs"__coeff_47',
'intensity__agg_linear_trend__attr_"stderr"__chunk_len_50__f_agg_"max"',
'intensity__fft_coefficient__attr_"abs"__coeff_60',
'intensity__fft_coefficient__attr_"abs"__coeff_80',
'intensity__fft_coefficient__attr_"abs"__coeff_72',
'intensity__fft_coefficient__attr_"abs"__coeff_76',
'intensity__binned_entropy__max_bins_10',
'intensity__fft_coefficient__attr_"abs"__coeff_70',
'intensity__fft_coefficient__attr_"abs"__coeff_79',
'intensity__ar_coefficient__coeff_3__k_10',
'intensity__fft_coefficient__attr_"abs"__coeff_44',
'intensity__fft_coefficient__attr_"abs"__coeff_78',
'intensity__fft_coefficient__attr_"abs"__coeff_74',
'intensity__fft_coefficient__attr_"abs"__coeff_77',
'intensity__fft_coefficient__attr_"abs"__coeff_45',
'intensity__maximum',
'intensity__fft_coefficient__attr_"abs"__coeff_81',
'intensity__fft_coefficient__attr_"abs"__coeff_75',
'intensity__agg_linear_trend__attr_"stderr"__chunk_len_10__f_agg_"var"',
'intensity__fft_coefficient__attr_"abs"__coeff_43',
'intensity__fft_coefficient__attr_"abs"__coeff_83',
'intensity__fft_coefficient__attr_"abs"__coeff_40',
'intensity__fft_coefficient__attr_"abs"__coeff_82',
'intensity__fft_coefficient__attr_"abs"__coeff_73',
'intensity__fft_coefficient__attr_"abs"__coeff_42',
'intensity__fft_coefficient__attr_"abs"__coeff_41',
'intensity__change_quantiles__f_agg_"mean"__isabs_True__qh_1.0__ql_0.4',
'intensity__agg_linear_trend__attr_"stderr"__chunk_len_10__f_agg_"max"',
'intensity__fft_coefficient__attr_"abs"__coeff_36',
'intensity__fft_coefficient__attr_"abs"__coeff_86',
'intensity__fft_coefficient__attr_"abs"__coeff_39',
'intensity__fft_coefficient__attr_"abs"__coeff_87',
'intensity__index_mass_quantile__q_0.2',
'intensity__fft_coefficient__attr_"abs"__coeff_84',
'intensity__energy_ratio_by_chunks__num_segments_10__segment_focus_2',
'intensity__mean_abs_change',
'intensity__change_quantiles__f_agg_"mean"__isabs_True__qh_1.0__ql_0.0',
'intensity__absolute_sum_of_changes',
'intensity__fft_aggregated__aggtype_"kurtosis"',
'intensity__fft_aggregated__aggtype_"variance"',
'intensity__energy_ratio_by_chunks__num_segments_10__segment_focus_1',
'intensity__fft_coefficient__attr_"abs"__coeff_38',
'intensity__index_mass_quantile__q_0.3',
'intensity__fft_coefficient__attr_"abs"__coeff_37',
'intensity__fft_coefficient__attr_"abs"__coeff_85',
'intensity__agg_linear_trend__attr_"stderr"__chunk_len_5__f_agg_"max"',
'intensity__fft_coefficient__attr_"abs"__coeff_90',
'intensity__fft_coefficient__attr_"abs"__coeff_88',
'intensity__fft_coefficient__attr_"abs"__coeff_92',
'intensity__change_quantiles__f_agg_"mean"__isabs_True__qh_1.0__ql_0.2',
'intensity__fft_coefficient__attr_"abs"__coeff_94',
'intensity__fft_coefficient__attr_"abs"__coeff_89',
'intensity__fft_coefficient__attr_"abs"__coeff_93',
'intensity__fft_coefficient__attr_"abs"__coeff_35',
'intensity__kurtosis',
'intensity__skewness',
'intensity__agg_linear_trend__attr_"slope"__chunk_len_5__f_agg_"var"',
'intensity__fft_coefficient__attr_"abs"__coeff_97',
'intensity__energy_ratio_by_chunks__num_segments_10__segment_focus_3',
'intensity__energy_ratio_by_chunks__num_segments_10__segment_focus_0',
'intensity__spkt_welch_density__coeff_2',
'intensity__fft_coefficient__attr_"abs"__coeff_34',
'intensity__index_mass_quantile__q_0.4',
'intensity__partial_autocorrelation__lag_5',
'intensity__index_mass_quantile__q_0.1',
'intensity__agg_linear_trend__attr_"stderr"__chunk_len_50__f_agg_"var"',
'intensity__ratio_beyond_r_sigma__r_0.5',
'intensity__agg_linear_trend__attr_"rvalue"__chunk_len_50__f_agg_"max"',
'intensity__fft_coefficient__attr_"abs"__coeff_95',
'intensity__partial_autocorrelation__lag_6',
'intensity__approximate_entropy__m_2__r_0.9',
'intensity__count_below_mean',
'intensity__count_above_mean',
'intensity__fft_coefficient__attr_"abs"__coeff_91',
'intensity__approximate_entropy__m_2__r_0.7',
'intensity__agg_linear_trend__attr_"slope"__chunk_len_10__f_agg_"max"',
'intensity__agg_linear_trend__attr_"slope"__chunk_len_10__f_agg_"var"',
'intensity__approximate_entropy__m_2__r_0.5',
'intensity__agg_linear_trend__attr_"slope"__chunk_len_5__f_agg_"max"',
'intensity__sample_entropy',
'intensity__approximate_entropy__m_2__r_0.3',
'intensity__fft_coefficient__attr_"abs"__coeff_96',
'intensity__fft_coefficient__attr_"imag"__coeff_1',
'intensity__agg_linear_trend__attr_"rvalue"__chunk_len_50__f_agg_"mean"',
'intensity__linear_trend__attr_"stderr"',
'intensity__linear_trend__attr_"slope"',
'intensity__fft_coefficient__attr_"abs"__coeff_4',
'intensity__agg_linear_trend__attr_"rvalue"__chunk_len_10__f_agg_"max"',
'intensity__agg_linear_trend__attr_"slope"__chunk_len_5__f_agg_"mean"',
'intensity__fft_coefficient__attr_"abs"__coeff_33',
'intensity__first_location_of_maximum',
'intensity__last_location_of_maximum',
'intensity__agg_linear_trend__attr_"rvalue"__chunk_len_50__f_agg_"var"',
'intensity__agg_linear_trend__attr_"rvalue"__chunk_len_10__f_agg_"mean"',
'intensity__fft_coefficient__attr_"abs"__coeff_3',
'intensity__agg_linear_trend__attr_"stderr"__chunk_len_50__f_agg_"mean"',
'intensity__fft_coefficient__attr_"abs"__coeff_32',
'intensity__agg_linear_trend__attr_"slope"__chunk_len_10__f_agg_"mean"',
'intensity__agg_linear_trend__attr_"rvalue"__chunk_len_5__f_agg_"mean"',
'intensity__standard_deviation',
'intensity__variance',
'intensity__agg_linear_trend__attr_"rvalue"__chunk_len_5__f_agg_"max"',
'intensity__agg_linear_trend__attr_"rvalue"__chunk_len_5__f_agg_"min"',
'intensity__linear_trend__attr_"rvalue"',
'intensity__fft_coefficient__attr_"abs"__coeff_6',
'intensity__agg_linear_trend__attr_"slope"__chunk_len_50__f_agg_"max"',
'intensity__energy_ratio_by_chunks__num_segments_10__segment_focus_9',
'intensity__agg_linear_trend__attr_"slope"__chunk_len_5__f_agg_"min"',
'intensity__agg_linear_trend__attr_"rvalue"__chunk_len_10__f_agg_"var"',
'intensity__agg_linear_trend__attr_"stderr"__chunk_len_5__f_agg_"mean"',
'intensity__fft_coefficient__attr_"abs"__coeff_2',
'intensity__fft_coefficient__attr_"abs"__coeff_99',
'intensity__augmented_dickey_fuller__attr_"usedlag"__autolag_"AIC"',
'intensity__spkt_welch_density__coeff_5',
'intensity__agg_linear_trend__attr_"slope"__chunk_len_10__f_agg_"min"',
'intensity__agg_linear_trend__attr_"rvalue"__chunk_len_10__f_agg_"min"',
'intensity__agg_linear_trend__attr_"stderr"__chunk_len_10__f_agg_"mean"',
'intensity__agg_linear_trend__attr_"rvalue"__chunk_len_5__f_agg_"var"',
'intensity__fft_coefficient__attr_"abs"__coeff_31',
'intensity__fft_coefficient__attr_"abs"__coeff_5',
'intensity__agg_linear_trend__attr_"slope"__chunk_len_50__f_agg_"mean"',
'intensity__c3__lag_1',
'intensity__fft_coefficient__attr_"abs"__coeff_7',
'intensity__fft_coefficient__attr_"abs"__coeff_8',
'intensity__fft_coefficient__attr_"abs"__coeff_30',
'intensity__fft_coefficient__attr_"abs"__coeff_9',
'intensity__fft_coefficient__attr_"abs"__coeff_98',
'intensity__ar_coefficient__coeff_7__k_10',
'intensity__ratio_beyond_r_sigma__r_7',
'intensity__ar_coefficient__coeff_6__k_10',
'intensity__fft_coefficient__attr_"abs"__coeff_29',
'intensity__agg_linear_trend__attr_"slope"__chunk_len_50__f_agg_"var"',
'intensity__fft_coefficient__attr_"abs"__coeff_1',
'intensity__variation_coefficient',
'intensity__fft_coefficient__attr_"abs"__coeff_10',
'intensity__max_langevin_fixed_point__m_3__r_30',
'intensity__c3__lag_2',
'intensity__spkt_welch_density__coeff_8',
'intensity__energy_ratio_by_chunks__num_segments_10__segment_focus_8',
'intensity__abs_energy',
'intensity__agg_linear_trend__attr_"stderr"__chunk_len_10__f_agg_"min"',
'intensity__c3__lag_3',
'intensity__fft_coefficient__attr_"abs"__coeff_11',
'intensity__ratio_beyond_r_sigma__r_1',
'intensity__fft_coefficient__attr_"abs"__coeff_13',
'intensity__agg_linear_trend__attr_"stderr"__chunk_len_5__f_agg_"min"',
'intensity__fft_coefficient__attr_"real"__coeff_1',
'intensity__index_mass_quantile__q_0.6',
'intensity__fft_coefficient__attr_"abs"__coeff_27',
'intensity__partial_autocorrelation__lag_7',
'intensity__fft_coefficient__attr_"abs"__coeff_12',
'intensity__agg_linear_trend__attr_"slope"__chunk_len_50__f_agg_"min"',
'intensity__linear_trend__attr_"pvalue"',
'intensity__fft_coefficient__attr_"abs"__coeff_28',
'intensity__agg_linear_trend__attr_"rvalue"__chunk_len_50__f_agg_"min"',
'intensity__agg_linear_trend__attr_"intercept"__chunk_len_5__f_agg_"min"',
'intensity__fft_coefficient__attr_"abs"__coeff_15',
'intensity__fft_coefficient__attr_"abs"__coeff_25',
'intensity__agg_linear_trend__attr_"intercept"__chunk_len_10__f_agg_"min"',
'intensity__agg_linear_trend__attr_"intercept"__chunk_len_50__f_agg_"min"',
'intensity__linear_trend__attr_"intercept"',
'intensity__fft_coefficient__attr_"abs"__coeff_14',
'intensity__agg_linear_trend__attr_"intercept"__chunk_len_5__f_agg_"mean"',
'intensity__fft_aggregated__aggtype_"skew"',
'intensity__fft_coefficient__attr_"abs"__coeff_16',
'intensity__agg_linear_trend__attr_"intercept"__chunk_len_10__f_agg_"mean"',
'intensity__partial_autocorrelation__lag_4',
'intensity__fft_coefficient__attr_"abs"__coeff_26',
'intensity__fft_coefficient__attr_"angle"__coeff_1',
'intensity__agg_linear_trend__attr_"intercept"__chunk_len_5__f_agg_"max"',
'intensity__fft_coefficient__attr_"abs"__coeff_17',
'intensity__index_mass_quantile__q_0.9',
'intensity__fft_coefficient__attr_"abs"__coeff_24',
'intensity__fft_coefficient__attr_"abs"__coeff_18',
'intensity__longest_strike_above_mean',
'intensity__fft_aggregated__aggtype_"centroid"',
'intensity__longest_strike_below_mean',
'intensity__approximate_entropy__m_2__r_0.1',
'intensity__fft_coefficient__attr_"abs"__coeff_22',
'intensity__agg_linear_trend__attr_"intercept"__chunk_len_10__f_agg_"max"',
'intensity__agg_linear_trend__attr_"intercept"__chunk_len_50__f_agg_"mean"',
'intensity__fft_coefficient__attr_"abs"__coeff_23',
'intensity__cwt_coefficients__coeff_5__w_5__widths_(2, 5, 10, 20)',
'intensity__cwt_coefficients__coeff_4__w_5__widths_(2, 5, 10, 20)',
'intensity__cwt_coefficients__coeff_6__w_5__widths_(2, 5, 10, 20)',
'intensity__sum_of_reoccurring_data_points',
'intensity__partial_autocorrelation__lag_1',
'intensity__autocorrelation__lag_1',
'intensity__ar_coefficient__coeff_8__k_10',
'intensity__fft_coefficient__attr_"imag"__coeff_7',
'intensity__cwt_coefficients__coeff_3__w_5__widths_(2, 5, 10, 20)',
'intensity__sum_of_reoccurring_values',
'intensity__fft_coefficient__attr_"abs"__coeff_20',
'intensity__fft_coefficient__attr_"abs"__coeff_19',
'intensity__ar_coefficient__coeff_4__k_10',
'intensity__cid_ce__normalize_True',
'intensity__fft_coefficient__attr_"abs"__coeff_21',
'intensity__cwt_coefficients__coeff_7__w_5__widths_(2, 5, 10, 20)',
'intensity__energy_ratio_by_chunks__num_segments_10__segment_focus_4',
'intensity__friedrich_coefficients__coeff_3__m_3__r_30',
'intensity__ar_coefficient__coeff_10__k_10',
'intensity__index_mass_quantile__q_0.7',
'intensity__minimum',
'intensity__large_standard_deviation__r_0.15000000000000002',
'intensity__quantile__q_0.1',
'intensity__quantile__q_0.2',
'intensity__fft_coefficient__attr_"real"__coeff_9',
'intensity__cwt_coefficients__coeff_7__w_10__widths_(2, 5, 10, 20)',
'intensity__cwt_coefficients__coeff_8__w_10__widths_(2, 5, 10, 20)',
'intensity__cwt_coefficients__coeff_2__w_5__widths_(2, 5, 10, 20)',
'intensity__quantile__q_0.3',
'intensity__cwt_coefficients__coeff_9__w_10__widths_(2, 5, 10, 20)',
'intensity__quantile__q_0.4',
'intensity__cwt_coefficients__coeff_6__w_10__widths_(2, 5, 10, 20)',
'intensity__cwt_coefficients__coeff_2__w_2__widths_(2, 5, 10, 20)',
'intensity__range_count__max_1000000000000.0__min_0',
'intensity__count_above__t_0',
'intensity__count_below__t_0',
'intensity__cwt_coefficients__coeff_8__w_5__widths_(2, 5, 10, 20)',
'intensity__cwt_coefficients__coeff_10__w_10__widths_(2, 5, 10, 20)',
'intensity__ratio_beyond_r_sigma__r_5',
'intensity__median',
'intensity__cwt_coefficients__coeff_5__w_10__widths_(2, 5, 10, 20)',
'intensity__number_crossing_m__m_0',
'intensity__number_crossing_m__m_1',
'intensity__number_crossing_m__m_-1',
'intensity__fft_coefficient__attr_"imag"__coeff_21',
'intensity__cwt_coefficients__coeff_11__w_10__widths_(2, 5, 10, 20)',
'intensity__quantile__q_0.6',
'intensity__cwt_coefficients__coeff_4__w_10__widths_(2, 5, 10, 20)',
'intensity__cwt_coefficients__coeff_3__w_2__widths_(2, 5, 10, 20)',
'intensity__ar_coefficient__coeff_5__k_10',
'intensity__cwt_coefficients__coeff_12__w_10__widths_(2, 5, 10, 20)',
'intensity__fft_coefficient__attr_"angle"__coeff_11',
'intensity__first_location_of_minimum',
'intensity__range_count__max_1__min_-1',
'intensity__last_location_of_minimum',
'intensity__quantile__q_0.7',
'intensity__fft_coefficient__attr_"angle"__coeff_21',
'intensity__cwt_coefficients__coeff_9__w_5__widths_(2, 5, 10, 20)',
'intensity__number_peaks__n_50',
'intensity__cwt_coefficients__coeff_3__w_10__widths_(2, 5, 10, 20)',
'intensity__cwt_coefficients__coeff_13__w_10__widths_(2, 5, 10, 20)',
'intensity__fft_coefficient__attr_"real"__coeff_61',
'intensity__change_quantiles__f_agg_"mean"__isabs_True__qh_0.4__ql_0.2',
'intensity__value_count__value_1',
'intensity__partial_autocorrelation__lag_9',
'intensity__fft_coefficient__attr_"real"__coeff_51',
'intensity__fft_coefficient__attr_"imag"__coeff_18',
'intensity__cwt_coefficients__coeff_0__w_20__widths_(2, 5, 10, 20)',
'intensity__ratio_beyond_r_sigma__r_1.5',
'intensity__cwt_coefficients__coeff_14__w_10__widths_(2, 5, 10, 20)',
'intensity__change_quantiles__f_agg_"var"__isabs_False__qh_0.4__ql_0.2',
'intensity__fft_coefficient__attr_"angle"__coeff_18',
'intensity__cwt_coefficients__coeff_1__w_5__widths_(2, 5, 10, 20)',
'intensity__fft_coefficient__attr_"real"__coeff_8',
'intensity__fft_coefficient__attr_"real"__coeff_22',
'intensity__fft_coefficient__attr_"imag"__coeff_2',
'intensity__fft_coefficient__attr_"imag"__coeff_22',
'intensity__friedrich_coefficients__coeff_2__m_3__r_30',
'intensity__cwt_coefficients__coeff_1__w_20__widths_(2, 5, 10, 20)',
'intensity__cwt_coefficients__coeff_1__w_2__widths_(2, 5, 10, 20)',
'intensity__value_count__value_-1',
'intensity__agg_autocorrelation__f_agg_"median"__maxlag_40',
'intensity__fft_coefficient__attr_"real"__coeff_50',
'intensity__cwt_coefficients__coeff_2__w_20__widths_(2, 5, 10, 20)',
'intensity__fft_coefficient__attr_"imag"__coeff_62',
'intensity__cwt_coefficients__coeff_3__w_20__widths_(2, 5, 10, 20)',
'intensity__fft_coefficient__attr_"real"__coeff_4',
'intensity__fft_coefficient__attr_"angle"__coeff_13',
'intensity__friedrich_coefficients__coeff_0__m_3__r_30',
'intensity__agg_linear_trend__attr_"intercept"__chunk_len_50__f_agg_"max"',
'intensity__cwt_coefficients__coeff_2__w_10__widths_(2, 5, 10, 20)',
'intensity__fft_coefficient__attr_"real"__coeff_2',
'intensity__quantile__q_0.8',
'intensity__autocorrelation__lag_9',
'intensity__fft_coefficient__attr_"real"__coeff_46',
'intensity__number_cwt_peaks__n_1',
# --- TODO: camaroさんの特徴量
# 'peak_wave_length',
# 'peak_intensity',
# 'mean',
# 'std',
# 'skew',
# 'kurtosis',
# 'skew_10',
# 'kurtosis_10',
# 'mean_10',
# 'std_10',
# 'skew_15',
# 'kurtosis_15',
# 'mean_15',
# 'std_15',
# 'skew_20',
# 'kurtosis_20',
# 'mean_20',
# 'std_20',
# 'skew_25',
# 'kurtosis_25',
# 'mean_25',
# 'std_25',
# 'skew_30',
# 'kurtosis_30',
# 'mean_30',
# 'std_30',
# 'skew_50',
# 'kurtosis_50',
# 'mean_50',
# 'std_50',
# 'skew_100',
# 'kurtosis_100',
# 'mean_100',
# 'std_100',
# 'skew_150',
# 'kurtosis_150',
# 'mean_150',
# 'std_150',
# 'skew_200',
# 'kurtosis_200',
# 'mean_200',
# 'std_200',
# 'em_ev',
# 'ex_ev',
# 'ev_diff',
# 'ev_ratio',
# 'num_peak',
# 'num_strong_peak',
# 'peak_wave_length2',
# 'peak_intensity2',
# 'peak_ratio',
# 'peak_ratio2'
]
try:
# インスタンス生成
runner = atm5_Runner(run_name, ModelCB, features, setting,
model_params, cv, FEATURE_DIR_NAME,
MODEL_DIR_NAME)
use_feature_name = runner.get_feature_name() # 今回の学習で使用する特徴量名を取得
# モデルのconfigをjsonで保存
value_list = [use_feature_name, model_params, cv, setting]
save_model_config(key_list, value_list, dir_name, run_name)
runner.run_train_cv() # 学習
runner.run_predict_cv() # 推論
Submission.create_submission(run_name, dir_name,
setting.get('target')) # submit作成
print('Done!')
except Exception as e:
print(traceback.format_exc())
print(f'ERROR:{e}')
save_model_config(key_list, value_list, dir_name, run_name)
# 学習
if cv_setting.get('method') == 'None':
runner.run_train_all() # 全データで学習
runner.run_predict_all() # 予測
else:
runner.run_train_cv() # 学習
ModelLGB.calc_feature_importance(
dir_name, run_name, use_feature_name) # feature_importanceを計算
ModelLGB.plot_learning_curve(run_name) # learning curveを描画
runner.run_predict_cv() # 予測
# submissionファイルの作成
lgb_preds = Util.load_df_pickle(dir_name + f'{run_name}-pred.pkl')
Submission.create_submission(run_name, dir_name, lgb_preds) # submit作成
# ##### ニューラルネットワーク ###########################################################
# features = features
# # CV設定の読み込み
# cv_setting = get_cv_info(random_state=53)
# # run nameの設定
# run_name = get_run_name(cv_setting, model_type="nn")
# dir_name = MODEL_DIR_NAME + run_name + '/'
# my_makedirs(dir_name) # runディレクトリの作成。ここにlogなどが吐かれる
# # ファイルの設定を読み込む
def main(model_type='lgb') -> str:
"""トレーニングのmain関数
model_typeによって学習するモデルを変更する
→ lgb, cb, xgb, nnが標準で用意されている
Args:
model_type (str, optional): どのモデルで学習させるかを指定. Defaults to 'lgb'.
Returns:
str: [description]
Examples:
>>> python hoge.py --model_type="lgb"
>>> python hoge.py lgb
"""
cv = get_cv_info() # CVの情報辞書
run_name = get_run_name(cv, model_type) # run名
dir_name = MODEL_DIR_NAME + run_name + '/' # 学習に使用するディレクトリ
setting = get_setting_info() # 諸々の設定ファイル辞書
# すでに実行済みのrun名がないかチェックし、ディレクトリを作成する
exist_check(MODEL_DIR_NAME, run_name)
my_makedirs(dir_name)
# モデルに合わせてパラメータを読み込む
model_cls = None
if model_type == 'lgb':
model_params = yml['MODEL_LGB']['PARAM']
model_cls = atm5_ModelLGB
elif model_type == 'cb':
model_params = yml['MODEL_CB']['PARAM']
model_cls = ModelCB
elif model_type == 'xgb':
pass
elif model_type == 'nn':
pass
else:
print('model_typeが不正なため終了します')
sys.exit(0)
features = [
'exc_wl',
'distance',
'distance_x',
'distance_y',
'pos_x',
'params0',
'params1',
'params2',
'params3',
'params4',
'params5',
'params6',
'params0_multi_rms',
'params0_divid_rms',
'params0_multi_beta',
'params0_divid_beta',
'params1_multi_rms',
'params1_divid_rms',
'params1_multi_beta',
'params1_divid_beta',
'params2_multi_rms',
'params2_divid_rms',
'params2_multi_beta',
'params2_divid_beta',
'params3_multi_rms',
'params3_divid_rms',
'params3_multi_beta',
'params3_divid_beta',
'params4_multi_rms',
'params4_divid_rms',
'params4_multi_beta',
'params4_divid_beta',
'params5_multi_rms',
'params5_divid_rms',
'params5_multi_beta',
'params5_divid_beta',
'params6_multi_rms',
'params6_divid_rms',
'params6_multi_beta',
'params6_divid_beta',
'params0_multi_params1',
'params0_divid_params1',
'params0_multi_params2',
'params0_divid_params2',
'params0_multi_params3',
'params0_divid_params3',
'params0_multi_params4',
'params0_divid_params4',
'params0_multi_params5',
'params0_divid_params5',
'params0_multi_params6',
'params0_divid_params6',
'params1_multi_params0',
'params1_divid_params0',
'params1_multi_params2',
'params1_divid_params2',
'params1_multi_params3',
'params1_divid_params3',
'params1_multi_params4',
'params1_divid_params4',
'params1_multi_params5',
'params1_divid_params5',
'params1_multi_params6',
'params1_divid_params6',
'params2_multi_params0',
'params2_divid_params0',
'params2_multi_params1',
'params2_divid_params1',
'params2_multi_params3',
'params2_divid_params3',
'params2_multi_params4',
'params2_divid_params4',
'params2_multi_params5',
'params2_divid_params5',
'params2_multi_params6',
'params2_divid_params6',
'params3_multi_params0',
'params3_divid_params0',
'params3_multi_params1',
'params3_divid_params1',
'params3_multi_params2',
'params3_divid_params2',
'params3_multi_params4',
'params3_divid_params4',
'params3_multi_params5',
'params3_divid_params5',
'params3_multi_params6',
'params3_divid_params6',
'params4_multi_params0',
'params4_divid_params0',
'params4_multi_params1',
'params4_divid_params1',
'params4_multi_params2',
'params4_divid_params2',
'params4_multi_params3',
'params4_divid_params3',
'params4_multi_params5',
'params4_divid_params5',
'params4_multi_params6',
'params4_divid_params6',
'params5_multi_params0',
'params5_divid_params0',
'params5_multi_params1',
'params5_divid_params1',
'params5_multi_params2',
'params5_divid_params2',
'params5_multi_params3',
'params5_divid_params3',
'params5_multi_params4',
'params5_divid_params4',
'params5_multi_params6',
'params5_divid_params6',
'params6_multi_params0',
'params6_divid_params0',
'params6_multi_params1',
'params6_divid_params1',
'params6_multi_params2',
'params6_divid_params2',
'params6_multi_params3',
'params6_divid_params3',
'params6_multi_params4',
'params6_divid_params4',
'params6_multi_params5',
'params6_divid_params5',
'rms',
'beta',
# --- TODO: スペクトル波長データ
'intensity_max',
'intensity_min',
'intensity_mean',
'intensity_std',
'intensity_sum',
'intensity_median',
'intensity_amplitude_v',
'intensity_q10',
'intensity_q25',
'intensity_q50',
'intensity_q75',
'intensity_q80',
'intensity_q85',
'intensity_q90',
'intensity_max_minus_q90',
'intensity_max_minus_q85',
'intensity_max_minus_q80',
'intensity_max_minus_q75',
'intensity_max_minus_q50',
'intensity_max_multi_q90',
'intensity_max_multi_q85',
'intensity_max_multi_q80',
'intensity_max_multi_q75',
'intensity_max_multi_q50',
'intensity_max_divid_q90',
'intensity_max_divid_q85',
'intensity_max_divid_q80',
'intensity_max_divid_q75',
'intensity_max_divid_q50',
'intensity_q85_divid_q90',
'intensity_q80_divid_q90',
'intensity_q75_divid_q90',
'intensity_q50_divid_q90',
# --- TODO: 波形の圧縮データ
'dc_umap1',
'dc_umap2',
'dc_tsne1',
'dc_tsne2',
'diff_exc_wl_mean_params0',
'diff_exc_wl_mean_params1',
'diff_exc_wl_mean_params2',
'diff_exc_wl_mean_params3',
'diff_exc_wl_mean_params4',
'diff_exc_wl_mean_params5',
'diff_exc_wl_mean_params6',
'diff_exc_wl_mean_rms',
'diff_exc_wl_mean_beta',
'diff_exc_wl_mean_intensity_amplitude_v',
'diff_exc_wl_mean_fwhm_09',
]
try:
# インスタンス生成
runner = atm5_Runner(run_name, model_cls, features, setting,
model_params, cv)
use_feature_name = runner.get_feature_name() # 今回の学習で使用する特徴量名を取得
# モデルのconfigをjsonで保存
value_list = [use_feature_name, model_params, cv, setting]
save_model_config(key_list, value_list, dir_name, run_name)
# 学習・推論
runner.run_train_cv()
runner.run_predict_cv()
# submit作成
Submission.create_submission(run_name, dir_name, setting.get('target'))
if model_type == 'lgb':
# feature_importanceを計算
atm5_ModelLGB.calc_feature_importance(dir_name,
run_name,
use_feature_name,
cv.get('n_splits'),
type='gain')
except Exception as e:
print(traceback.format_exc())
print(f'ERROR:{e}')
params_nn = {
"layers": 3,
# サンプルのため早く終わるように設定
"nb_epoch": 5, # 1000
"patience": 10,
"dropout": 0.5,
"units": 512,
}
# 特徴量の指定
features = [f"feat_{i}" for i in range(1, 94)]
# xgboostによる学習・予測
runner = Runner("xgb1", ModelXGB, features, params_xgb)
runner.run_train_cv()
runner.run_predict_cv()
Submission.create_submission("xgb1")
# ニューラルネットによる学習・予測
runner = Runner("nn1", ModelNN, features, params_nn)
runner.run_train_cv()
runner.run_predict_cv()
Submission.create_submission("nn1")
"""
# (参考）xgboostによる学習・予測 - 学習データ全体を使う場合
runner = Runner('xgb1-train-all', ModelXGB, features, params_xgb_all)
runner.run_train_all()
runner.run_test_all()
Submission.create_submission('xgb1-train-all')
"""
def main(mode='prd', create_features=True, model_type='lgb', is_kernel=False) -> str:
confirm('mode:{}, create_feature:{} '.format(str(mode), str(create_features)))
if create_features:
# データ生成
train, test, specs, train_labels, submission = read_data_all(mode)
features_train, features_test, win_code, list_of_user_activities, list_of_event_code, \
activities_labels, assess_titles, list_of_event_id, all_title_event_code, all_type_world = encode_title(train, test)
del train, test
gc.collect()
features_train = features_train.merge(specs, how='left', on='event_id', suffixes=('','_y'))
features_test = features_test.merge(specs, how='left', on='event_id', suffixes=('','_y'))
features_train, features_test = get_train_and_test(features_train, features_test,
win_code, list_of_user_activities, list_of_event_code,
activities_labels, assess_titles, list_of_event_id, all_title_event_code, all_type_world, is_kernel)
reduce_train, reduce_test, _ = preprocess(features_train, features_test, assess_titles)
# user属性情報の生成とマージ
""" スコア悪くなるので一旦コメント
train, train_session_master = add_session_order_to_train(train, train_labels)
user_profiles_train = create_user_profile_train(train)
user_profiles_test = create_user_profile_test(test)
train_session_master = train_session_master.merge(user_profiles_train, how='left', on=['installation_id', 'session_order'])
reduce_train = reduce_train.merge(train_session_master, how='left', on=['installation_id', 'game_session'])
reduce_test = reduce_test.merge(user_profiles_test, how='left', on='installation_id')
"""
del features_train, features_test, _
gc.collect()
# 不要なカラムの削除
# cols_to_drop = ['game_session', 'installation_id', 'timestamp', 'session_order', 'accuracy_group', 'timestampDate'] + ['acc_' + title for title in assess_titles] # installation_idでGroupKFoldしない場合はこちらを使用
cols_to_drop = ['game_session', 'timestamp', 'session_order', 'accuracy_group', 'timestampDate'] + ['acc_' + title for title in assess_titles]
cols_to_drop = [col for col in cols_to_drop if col in reduce_train.columns]
X_train = reduce_train.drop(cols_to_drop, axis=1)
X_train.columns = ["".join(c if c.isalnum() else "_" for c in str(x)) for x in X_train.columns] # カラム名にカンマなどが含まれており、lightgbmでエラーが出るため
y_train = reduce_train['accuracy_group']
cols_to_drop = [col for col in cols_to_drop if col in reduce_test.columns]
X_test = reduce_test.drop(cols_to_drop, axis=1)
X_test.columns = ["".join(c if c.isalnum() else "_" for c in str(x)) for x in X_test.columns] # カラム名にカンマなどが含まれており、lightgbmでエラーが出るため
# 特徴量選択
""" スコア悪くなるので一旦コメント
to_exclude, X_test = select_ajusted(X_train, X_test)
X_train = X_train.drop(to_exclude, axis=1)
X_test = X_test.drop(to_exclude, axis=1)
"""
X_train.to_pickle(FEATURE_DIR_NAME + 'X_train.pkl')
y_train.to_pickle(FEATURE_DIR_NAME + 'y_train.pkl')
X_test.to_pickle(FEATURE_DIR_NAME + 'X_test.pkl')
# CVの設定.methodは[KFold, StratifiedKFold ,GroupKFold]から選択可能
# CVしない場合（全データで学習させる場合）はmethodに'None'を設定
# StratifiedKFold or GroupKFoldの場合は、cv_targetに対象カラム名を設定する
cv = {
'method': 'GroupKFold',
'n_splits': 5,
'random_state': 42,
'shuffle': True,
'cv_target': 'installation_id'
}
if model_type == 'lgb' or model_type == 'all':
# ######################################################
# 学習・推論 LightGBM ###################################
# run nameの設定
run_name = 'lgb'
run_name = run_name + suffix
dir_name = MODEL_DIR_NAME + run_name + '/'
exist_check(MODEL_DIR_NAME, run_name)
my_makedirs(dir_name) # runディレクトリの作成。ここにlogなどが吐かれる
# 諸々の設定
setting = {
'run_name': run_name, # run名
'feature_directory': FEATURE_DIR_NAME, # 特徴量の読み込み先ディレクトリ
'target': 'accuracy_group', # 目的変数
'calc_shap': False, # shap値を計算するか否か
'save_train_pred': False # trainデータでの推論値を保存するか否か（trainデータでの推論値を特徴量として加えたい場合はTrueに設定する）
}
# モデルのパラメータ
model_params = {
'boosting_type': 'gbdt',
'objective': 'regression',
'metric': 'rmse',
'learning_rate': 0.01,
'subsample': 0.75,
'subsample_freq': 1,
'feature_fraction': 0.9,
'max_depth': 15,
'lambda_l1': 1,
'lambda_l2': 1,
'num_round': 50000,
'early_stopping_rounds': 300,
'verbose': -1,
'verbose_eval': 300,
'random_state': 999
}
if is_kernel:
runner = Runner(run_name, ModelLGB, setting, model_params, cv, FEATURE_DIR_NAME, MODEL_DIR_NAME, X_train, y_train, X_test)
else:
runner = Runner(run_name, ModelLGB, setting, model_params, cv, FEATURE_DIR_NAME, MODEL_DIR_NAME)
use_feature_name = runner.get_feature_name() # 今回の学習で使用する特徴量名を取得
# モデルのconfigをjsonで保存
value_list = [use_feature_name, model_params, cv, setting]
save_model_config(key_list, value_list, dir_name, run_name)
if cv.get('method') == 'None':
# TODO: こちらも動くように修正する
runner.run_train_all() # 全データで学習
runner.run_predict_all() # 推論
else:
runner.run_train_cv() # 学習
ModelLGB.calc_feature_importance(dir_name, run_name, use_feature_name) # feature_importanceを計算
_pred = runner.run_predict_cv(is_kernel) # 推論
if is_kernel:
# kaggleカーネル実行
if model_type == 'lgb':
# シングルモデルでのcsv作成
submission[setting.get('target')] = _pred.astype(int)
submission.to_csv('submission.csv', index=False)
else:
# ブレンドするためのcsv作成
submission_lgb = submission.copy()
submission_lgb[setting.get('target')] = _pred.astype(int)
submission_lgb.to_csv('submission_lgb.csv', index=False)
else:
# ローカルでの実行
Submission.create_submission(run_name, dir_name, setting.get('target')) # submit作成
if model_type == 'cb' or model_type == 'all':
# ######################################################
# 学習・推論 Catboost ###################################
# run nameの設定
run_name = 'cb'
run_name = run_name + suffix
dir_name = MODEL_DIR_NAME + run_name + '/'
exist_check(MODEL_DIR_NAME, run_name)
my_makedirs(dir_name) # runディレクトリの作成。ここにlogなどが吐かれる
# 諸々の設定
setting = {
'run_name': run_name, # run名
'feature_directory': FEATURE_DIR_NAME, # 特徴量の読み込み先ディレクトリ
'target': 'accuracy_group', # 目的変数
'calc_shap': False, # shap値を計算するか否か
'save_train_pred': False # trainデータでの推論値を保存するか否か（trainデータでの推論値を特徴量として加えたい場合はTrueに設定する）
}
# モデルのパラメータ
model_params = {
'loss_function': 'RMSE',
'task_type': "CPU",
'iterations': 50000,
'od_type': "Iter",
'depth': 10,
'colsample_bylevel': 0.5,
'early_stopping_rounds': 300,
'l2_leaf_reg': 18,
'random_seed': 42,
'verbose_eval': 300,
'use_best_model': True
}
if is_kernel:
runner = Runner(run_name, ModelCB, setting, model_params, cv, FEATURE_DIR_NAME, MODEL_DIR_NAME, X_train, y_train, X_test)
else:
runner = Runner(run_name, ModelCB, setting, model_params, cv, FEATURE_DIR_NAME, MODEL_DIR_NAME)
use_feature_name = runner.get_feature_name() # 今回の学習で使用する特徴量名を取得
# モデルのconfigをjsonで保存
value_list = [use_feature_name, model_params, cv, setting]
save_model_config(key_list, value_list, dir_name, run_name)
if cv.get('method') == 'None':
# TODO: こちらも動くように修正する
runner.run_train_all() # 全データで学習
runner.run_predict_all() # 推論
else:
runner.run_train_cv() # 学習
_pred = runner.run_predict_cv(is_kernel) # 推論
if is_kernel:
# kaggleカーネル実行
if model_type == 'cb':
# シングルモデルでのcsv作成
submission[setting.get('target')] = _pred.astype(int)
submission.to_csv('submission.csv', index=False)
else:
# ブレンドするためのcsv作成
submission_cb = submission.copy()
submission_cb[setting.get('target')] = _pred.astype(int)
submission_cb.to_csv('submission_cb.csv', index=False)
else:
# ローカルでの実行
Submission.create_submission(run_name, dir_name, setting.get('target')) # submit作成
if model_type == 'nn' or model_type == 'all':
# ######################################################
# 学習・推論 NN(MLP) ###################################
# run nameの設定
run_name = 'nn'
run_name = run_name + suffix
dir_name = MODEL_DIR_NAME + run_name + '/'
exist_check(MODEL_DIR_NAME, run_name)
my_makedirs(dir_name) # runディレクトリの作成。ここにlogなどが吐かれる
# 諸々の設定
setting = {
'run_name': run_name, # run名
'feature_directory': FEATURE_DIR_NAME, # 特徴量の読み込み先ディレクトリ
'target': 'accuracy_group', # 目的変数
'calc_shap': False, # shap値を計算するか否か
'save_train_pred': False # trainデータでの推論値を保存するか否か（trainデータでの推論値を特徴量として加えたい場合はTrueに設定する）
}
# モデルのパラメータ
model_params = {
'layers': 4,
'nb_epoch': 500,
'patience': 20,
'dropout': 0.3,
'units': 512,
'classes': 1
}
if is_kernel:
runner = Runner(run_name, ModelNN, setting, model_params, cv, FEATURE_DIR_NAME, MODEL_DIR_NAME, X_train, y_train, X_test)
else:
runner = Runner(run_name, ModelNN, setting, model_params, cv, FEATURE_DIR_NAME, MODEL_DIR_NAME)
use_feature_name = runner.get_feature_name() # 今回の学習で使用する特徴量名を取得
# モデルのconfigをjsonで保存
value_list = [use_feature_name, model_params, cv, setting]
save_model_config(key_list, value_list, dir_name, run_name)
# one-hot-encoding
if len(runner.categoricals) > 0:
one_hot_encoder = ce.OneHotEncoder(cols=runner.categoricals, drop_invariant=True)
one_hot_encoder.fit(runner.train_x[runner.categoricals])
ohe_path = os.path.join('.', 'one-hot-enc.pkl')
Util.dump(one_hot_encoder, ohe_path)
if cv.get('method') == 'None':
# TODO: こちらも動くように修正する
runner.run_train_all() # 全データで学習
runner.run_predict_all() # 推論
else:
runner.run_train_cv() # 学習
_pred = runner.run_predict_cv(is_kernel) # 推論
if is_kernel:
# kaggleカーネル実行
if model_type == 'nn':
# シングルモデルでのcsv作成
submission[setting.get('target')] = _pred.astype(int)
submission.to_csv('submission.csv', index=False)
else:
# ブレンドするためのcsv作成
submission_nn = submission.copy()
submission_nn[setting.get('target')] = _pred.astype(int)
submission_nn.to_csv('submission_nn.csv', index=False)
else:
# ローカルでの実行
Submission.create_submission(run_name, dir_name, setting.get('target')) # submit作成
if model_type == 'xgb' or model_type == 'all':
# ######################################################
# 学習・推論 xgboost ###################################
# run nameの設定
run_name = 'xgb'
run_name = run_name + suffix
dir_name = MODEL_DIR_NAME + run_name + '/'
exist_check(MODEL_DIR_NAME, run_name)
my_makedirs(dir_name) # runディレクトリの作成。ここにlogなどが吐かれる
# 諸々の設定
setting = {
'run_name': run_name, # run名
'feature_directory': FEATURE_DIR_NAME, # 特徴量の読み込み先ディレクトリ
'target': 'accuracy_group', # 目的変数
'calc_shap': False, # shap値を計算するか否か
'save_train_pred': False # trainデータでの推論値を保存するか否か（trainデータでの推論値を特徴量として加えたい場合はTrueに設定する）
}
# モデルのパラメータ
model_params = {
'objective':'reg:squarederror',
'eval_metric': 'rmse',
'colsample_bytree': 0.8,
'learning_rate': 0.01,
'max_depth': 10,
'subsample': 1,
'min_child_weight':3,
'gamma':0.25,
'num_round': 50000,
'early_stopping_rounds': 300,
'verbose': 300,
'random_state': 999
}
if is_kernel:
runner = Runner(run_name, ModelXGB, setting, model_params, cv, FEATURE_DIR_NAME, MODEL_DIR_NAME, X_train, y_train, X_test)
else:
runner = Runner(run_name, ModelXGB, setting, model_params, cv, FEATURE_DIR_NAME, MODEL_DIR_NAME)
use_feature_name = runner.get_feature_name() # 今回の学習で使用する特徴量名を取得
# モデルのconfigをjsonで保存
value_list = [use_feature_name, model_params, cv, setting]
save_model_config(key_list, value_list, dir_name, run_name)
if cv.get('method') == 'None':
# TODO: こちらも動くように修正する
runner.run_train_all() # 全データで学習
runner.run_predict_all() # 推論
else:
runner.run_train_cv() # 学習
ModelXGB.calc_feature_importance(dir_name, run_name, use_feature_name) # feature_importanceを計算
_pred = runner.run_predict_cv(is_kernel) # 推論
if is_kernel:
# kaggleカーネル実行
if model_type == 'xgb':
# シングルモデルでのcsv作成
submission[setting.get('target')] = _pred.astype(int)
submission.to_csv('submission.csv', index=False)
else:
# ブレンドするためのcsv作成
submission_xgb = submission.copy()
submission_xgb[setting.get('target')] = _pred.astype(int)
submission_xgb.to_csv('submission_xgb.csv', index=False)
else:
# ローカルでの実行
Submission.create_submission(run_name, dir_name, setting.get('target')) # submit作成
# 推論のブレンド
# TODO: xbgの結果も入れる
if model_type == 'all' and is_kernel:
weights = {'lgb': 0.30, 'cb': 0.40, 'nn': 0.00, 'xgb': 0.30}
blend_pred = (submission_lgb[setting.get('target')] * weights['lgb']) \
+ (submission_cb[setting.get('target')] * weights['cb']) \
+ (submission_nn[setting.get('target')] * weights['nn']) \
+ (submission_xgb[setting.get('target')] * weights['xgb'])
dist = Counter(reduce_train[setting.get('target')])
for k in dist:
dist[k] /= len(reduce_train)
acum = 0
bound = {}
for i in range(3):
acum += dist[i]
bound[i] = np.percentile(blend_pred, acum * 100)
def classify(x):
if x <= bound[0]:
return 0
elif x <= bound[1]:
return 1
elif x <= bound[2]:
return 2
else:
return 3
blend_pred = np.array(list(map(classify, blend_pred)))
submission[setting.get('target')] = blend_pred.astype(int)
submission.to_csv('submission.csv', index=False)
return 'Success!'