def train_1fold(self, fold, params, params_custom):
X_train, X_valid, y_train, y_valid, X_test, vdx, tdx = self.get_fold_data(fold)
if fold == 0:
X_train.dtypes.to_csv(self.models_path + "/dtypes.csv")
logger.info(f"X_train.shape = {X_train.shape} f{fold:02d}")
mprof_timestamp(f"lgb_dataset_f{fold}")
lgb_train = lgb.Dataset(X_train, y_train)
lgb_valid = lgb.Dataset(X_valid, y_valid)
feval = build_feval(params_custom)
# 学習の実行
evals = dict()
params2 = copy.deepcopy(params)
callbacks = [log_evaluation(logger, period=10)]
if params2["seed"] is not None:
params2["seed"] = params2["seed"] + fold
logger.info(f"Set lgbm train seed = {params2['seed']}")
logger.info(f"Start train f{fold:02d}")
mprof_timestamp(f"lgb_train_f{fold}")
model = lgb.train(params2, lgb_train,
valid_names=["train", "valid"], valid_sets=[lgb_train, lgb_valid],
verbose_eval=100, evals_result=evals, feval=feval, callbacks=callbacks,
keep_training_booster=True)
logger.info(f"current_iteration={model.current_iteration()}")
logger.info(f"best_iteration={model.best_iteration}")
mprof_timestamp(f"lgb_postproc_f{fold}")
model.save_model(self.models_path + f"/model-lgbm-f{fold:02d}.txt", num_iteration=model.best_iteration)
model.save_model(self.models_path + f"/model-lgbm-last-f{fold:02d}.txt", num_iteration=-1)
evals_df = pd.DataFrame({
f"logloss_train_f{fold:02d}":evals["train"]['logloss'],
f"accuracy_train_f{fold:02d}":evals["train"]['accuracy'],
f"logloss_valid_f{fold:02d}":evals['valid']['logloss'],
f"accuracy_valid_f{fold:02d}":evals['valid']['accuracy']
})
self.evals_df.append(evals_df)
# 予測値の保存
preds_valid = model.predict(X_valid, num_iteration=model.best_iteration)
self.preds_valid_all.loc[vdx, "pred"] = preds_valid
preds_train = model.predict(X_train, num_iteration=model.best_iteration)
self.preds_train_all.append(pd.DataFrame({fold:preds_train}, index=tdx))
preds_test = model.predict(X_test, num_iteration=model.best_iteration)
self.preds_test_all.append(preds_test)
# 性能指標の保存
ms = [fold, model.best_score["train"]["accuracy"], model.best_score["valid"]["accuracy"],
model.best_score["train"]["logloss"], model.best_score["valid"]["logloss"], model.best_iteration]
self.mets.append(ms)
show_mets(*ms)
for it in ["gain", "split"]:
imp = pd.Series(model.feature_importance(importance_type=it, iteration=model.best_iteration),
index=model.feature_name())
imp.name = fold
imp.index.name = "feature"
self.importance[it].append(imp)
def create_tarenc_agg_features(mode_target_persons, mode_target_cols):
logger.info(f"tarenc_agg_tp{mode_target_persons}_tc{mode_target_cols}")
mprof_timestamp(f"tarenc_agg_tp{mode_target_persons}_tc{mode_target_cols}")
_ = Parallel(n_jobs=args.n_jobs//2, verbose=args.verbose_joblib) \
([delayed(create_tarenc_agg_features_1fold)(fold, mode_target_persons, mode_target_cols, "tarenc") for fold in range(args.FOLD_NUM)])
def create_freqenc_agg_features(mode_target_persons, mode_target_cols):
logger.info(f"freqenc_agg_tp{mode_target_persons}_tc{mode_target_cols}")
mprof_timestamp(
f"freqenc_agg_tp{mode_target_persons}_tc{mode_target_cols}")
create_tarenc_agg_features_1fold(None, mode_target_persons,
mode_target_cols, "freqenc")
os.makedirs(args.features_path, exist_ok=False)
else: # 既存のdirに出力
if args.DO_TEST:
args.features_path = "../features/test"
os.makedirs(args.features_path, exist_ok=True)
logging_config.init(f"{args.features_path}/log_{timestr}.log")
logger.info(f"features_path = {args.features_path}")
logger.info("args =\n" + pp.pformat(vars(args)))
util.dump_json(vars(args), f"{args.features_path}/args_{timestr}.json")
shutil.copytree("../src", args.features_path + "/src_" + timestr)
mprof_timestamp("basic")
def create_team_agg_features_wrp(merge):
ma = merge[[
c for c in merge.columns if re.match("A[1234]-(level|rank-int)", c)
]]
mb = merge[[
c for c in merge.columns if re.match("B[1234]-(level|rank-int)", c)
]]
ma_agg = create_team_agg_features(ma, "A")
ma234_agg = create_team_agg_features(ma, "A234")
mb_agg = create_team_agg_features(mb, "B")
m_diff = create_team_agg_diff_features(ma_agg, mb_agg)
m_diff_a234b = create_team_agg_diff_features(ma234_agg, mb_agg)