def tune_params(
base_param: Dict,
X: pd.DataFrame,
y: pd.Series,
cv: BaseCrossValidator,
time_budget: Optional[int] = None,
) -> Dict:
train_index, test_index = next(cv.split(X, y))
dtrain = lgb.Dataset(X.iloc[train_index], y.iloc[train_index])
dvalid = lgb.Dataset(X.iloc[test_index], y.iloc[test_index])
params = copy.deepcopy(base_param)
if "early_stopping_rounds" not in params:
params["early_stopping_rounds"] = 100
best_params, tuning_history = dict(), list()
lightgbm_tuner.train(
params,
dtrain,
valid_sets=[dvalid],
verbose_eval=0,
best_params=best_params,
tuning_history=tuning_history,
time_budget=time_budget,
)
result_param = copy.deepcopy(base_param)
result_param.update(best_params)
return result_param
def run(self):
feature = (self.load("feature").query(
f"store_id == '{self.store_id}'").reset_index(level=1))
train = feature.query("date < '2016-03-28'").drop(
columns=["store_id", "date"])
valid = feature.query("'2016-03-28' <= date < '2016-04-25'").drop(
columns=["store_id", "date"])
test = feature.query("'2016-04-25' <= date").copy()
dataset_train = lgb.Dataset(train.drop(columns="demand"),
train["demand"])
dataset_valid = lgb.Dataset(valid.drop(columns="demand"),
valid["demand"])
params = {
"objective": "regression",
"seed": 110,
"learning_rate": 0.01,
"boosting_type": "gbdt",
"metric": "rmse",
"lambda_l1": 0.0,
"lambda_l2": 0.0,
"num_leaves": 131,
"feature_fraction": 0.41600000000000004,
"bagging_fraction": 1.0,
"bagging_freq": 0,
"min_data_in_leaf": 20,
"min_child_samples": 25,
}
print("hoge", self.time_budget)
if not self.time_budget:
print("tuningしないよ")
model = lgb.train(
params,
dataset_train,
num_boost_round=100000,
valid_sets=[dataset_train, dataset_valid],
early_stopping_rounds=200,
verbose_eval=100,
)
else:
print("tuningするよ")
model = lightgbm_tuner.train(
params,
dataset_train,
num_boost_round=100000,
valid_sets=[dataset_train, dataset_valid],
early_stopping_rounds=200,
verbose_eval=-1,
time_budget=self.time_budget,
)
predict = model.predict(
test.drop(columns=["date", "store_id", "demand"]))
test["demand"] = predict
test = test.reset_index().set_index(["id", "date"])[["demand"]]
result = ModelResult(model, params, test)
self.dump(result)
def get_best_params(train_x: t.Any, train_y: t.Any, num_class: int) -> t.Any:
tr_x, val_x, tr_y, val_y = train_test_split(train_x,
train_y,
test_size=0.2,
random_state=1)
lgb_train = lgb.Dataset(tr_x, tr_y)
lgb_eval = lgb.Dataset(val_x, val_y)
best_params = {}
params = {
"objective": "multiclass",
"metric": "multi_logloss",
"boosting_type": "gbdt",
"num_class": num_class,
}
best_params = {}
tuning_history = []
gbm = lightgbm_tuner.train(
params,
lgb_train,
valid_sets=lgb_eval,
num_boost_round=10000,
early_stopping_rounds=20,
verbose_eval=10,
best_params=best_params,
tuning_history=tuning_history,
)
return best_params
def get_tuned_model(train_x, train_y, valid_x, valid_y, num_class) -> t.Any:
# 学習用データセット
train_set = lgb.Dataset(train_x, train_y)
# 評価用データセット
valid_set = lgb.Dataset(valid_x, valid_y)
# チューニングしたハイパラを格納
best_params = {}
params = {
'objective': 'multiclass',
'metric': 'multi_logloss',
'boosting_type': 'gbdt',
'num_class': num_class,
'num_threads': 2
}
best_params = {}
tuning_history = []
gbm = lightgbm_tuner.train(
params,
train_set,
valid_sets=[train_set, valid_set],
num_boost_round=10000,
early_stopping_rounds=20,
verbose_eval=10,
best_params=best_params,
tuning_history=tuning_history
)
joblib.dump(gbm, f"{DATA_DIR}/lgb_model.pkl")
importance = pd.DataFrame(gbm.feature_importance(), index=train_x.columns, columns=['importance']).sort_values('importance', ascending=[False])
print(importance)
return gbm
def train_lgb(bst_params, fit_params, X, y, cv, drop_when_train=None):
models = []
if drop_when_train is None:
drop_when_train = []
for idx_fold, (idx_trn, idx_val) in enumerate(cv.split(X, y)):
print(f"\n---------- Fold: ({idx_fold + 1} / {cv.get_n_splits()}) ----------\n")
X_trn, X_val = X.iloc[idx_trn], X.iloc[idx_val]
y_trn, y_val = y.iloc[idx_trn], y.iloc[idx_val]
train_set = lgb.Dataset(X_trn.drop(drop_when_train, axis=1), label=y_trn)
val_set = lgb.Dataset(X_val.drop(drop_when_train, axis=1), label=y_val)
best_params, tuning_history = dict(), list()
print("start")
model = lightgbm_tuner.train(
bst_params,
train_set,
valid_sets=[train_set, val_set],
valid_names=["train", "valid"],
**fit_params,
#fobj = custom_asymmetric_train,
#feval = custom_asymmetric_valid,
best_params=best_params,
tuning_history=tuning_history
)
models.append(model)
print(best_params)
print(tuning_history)
del idx_trn, idx_val, X_trn, X_val, y_trn, y_val
gc.collect()
return models
def get_best_params(train_x, train_y):
tr_x, val_x, tr_y, val_y = train_test_split(train_x,
train_y,
test_size=0.2,
random_state=1)
tr_set = lgb.Dataset(tr_x, tr_y)
val_set = lgb.Dataset(val_x, val_y, reference=tr_set)
best_params = {}
params = {
"objective": "binary",
"metric": "binary_logloss",
"boosting_type": "gbdt",
}
best_params = {}
tuning_history = []
gbm = lightgbm_tuner.train(
params,
tr_set,
valid_sets=val_set,
num_boost_round=1000,
early_stopping_rounds=50,
verbose_eval=10,
best_params=best_params,
tuning_history=tuning_history,
)
return best_params
def fit(self,
tr_x,
tr_y,
va_x=None,
va_y=None,
cat_features=None,
feval=None):
validation = va_x is not None
lgb_train = lgb.Dataset(tr_x, tr_y, categorical_feature=cat_features)
if validation:
lgb_eval = lgb.Dataset(va_x,
va_y,
reference=lgb_train,
categorical_feature=cat_features)
logger = logging.getLogger('main')
best_params, tuning_history = dict(), list()
if validation:
self.model = lgb_tuner.train(self.params,
lgb_train,
num_boost_round=10000,
valid_sets=[lgb_eval],
verbose_eval=0,
early_stopping_rounds=200,
feval=feval,
best_params=best_params,
tuning_history=tuning_history)
else:
self.model = lgb_tuner.train(self.params,
lgb_train,
num_boost_round=10000,
best_params=best_params,
tuning_history=tuning_history)
logging.debug('Best Params:', best_params)
logging.debug('Tuning history:', tuning_history)
def main():
# load dataset
boston = datasets.load_boston()
X, y = boston.data, boston.target
# split train and test
X_train, X_test, y_train, y_test = train_test_split(X, y)
# create dataset for lightgbm
lgb_train = lgb.Dataset(X_train, y_train)
lgb_eval = lgb.Dataset(X_test, y_test)
# parameters
params = {
'objective': 'regression',
'metric': 'rmse' # RMSE（Root Mean Square Error ,平均二乗誤差平方根）
}
# stepwise tuning of OPTUNA
best_params = {}
tuning_history = []
# training , train_setは説明変数が入ってる。
model = lightgbm_tuner.train(params=params,
train_set=lgb_train,
valid_sets=(lgb_train, lgb_eval),
early_stopping_rounds=100,
num_boost_round=10000,
verbose_eval=50,
best_params=best_params,
tuning_history=tuning_history)
# prediction
y_pred = model.predict(data=X_test, num_iteration=model.best_iteration)
# caluculate RMSE
mse = mean_squared_error(y_test, y_pred)
rmse = np.sqrt(mse)
print(rmse)
fig = plt.figure()
ax = fig.add_subplot(1, 1, 1)
lgb.plot_importance(model, height=0.5, ax=ax, figsize=(8, 10))
plt.show()
print(best_params)
print(tuning_history)
def train(self):
lgb_train = lgb.Dataset(self.X_train, self.y_train)
lgb_eval = lgb.Dataset(self.X_test, self.y_test)
params = {
'objective': 'multiclass',
'num_class': self.ans_len,
'metric': 'multi_error'
}
best_params = {}
tuning_history = []
model = lightgbm_tuner.train(
params,
lgb_train,
valid_sets=lgb_eval,
num_boost_round=self.num_boost_round,
early_stopping_rounds=self.early_stopping_round,
verbose_eval=False,
best_params=best_params,
tuning_history=tuning_history,
)
self.y_pred_prob = model.predict(self.X_test,
num_iteration=model.best_iteration)
self.y_pred = np.argmax(self.y_pred_prob, axis=1)
df_pred = pd.DataFrame({
'target': self.y_test[0],
'target_pred': self.y_pred
})
print(df_pred)
# df_pred_prob = pd.DataFrame({'y':self.y_test[0],
# 'target0_prob':self.y_pred_prob[:,0], 'target1_prob':self.y_pred_prob[:,1], 'target2_prob':self.y_pred_prob[:,2],
# 'target3_prob':self.y_pred_prob[:,3], 'target4_prob':self.y_pred_prob[:,4], 'target5_prob':self.y_pred_prob[:,5],
# 'target6_prob':self.y_pred_prob[:,6], 'target7_prob':self.y_pred_prob[:,7], 'target8_prob':self.y_pred_prob[:,8],
# 'target9_prob':self.y_pred_prob[:,9], 'target10_prob':self.y_pred_prob[:,10], 'target11_prob':self.y_pred_prob[:,11],
# 'target12_prob':self.y_pred_prob[:,12]})
acc = accuracy_score(self.y_test, self.y_pred)
print('Acc :', acc)
def get_best_params(train_x: t.Any, train_y: t.Any, valid_x: t.Any,
valid_y: t.Any) -> t.Any:
lgb_train = lgb.Dataset(train_x, train_y)
lgb_eval = lgb.Dataset(valid_x, valid_y, reference=lgb_train)
best_params = {}
params = {
'objective': 'regression',
'metric': 'rmse',
'boosting_type': 'gbdt',
}
best_params = {}
tuning_history = []
gbm = lightgbm_tuner.train(params,
lgb_train,
valid_sets=lgb_eval,
num_boost_round=10000,
early_stopping_rounds=20,
verbose_eval=10,
best_params=best_params,
tuning_history=tuning_history)
return best_params
def tune_lgb(
X: Union[pd.DataFrame, np.array], y: Union[pd.Series, np.array], seed: int,
):
"""lightgbmを用いてfoldごとのモデルを返す。
Args:
X (Union[pd.DataFrame, np.array]): 特徴行列
y (Union[pd.Series, np.array]): 目的変数
Returns:
List[lgb.Booster]: Boosterのリスト
"""
params = {
"objective": "binary",
"metric": "binary_logloss",
"verbosity": -1,
"boosting_type": "gbdt",
}
best_params, tuning_history = dict(), list()
train_x, val_x, train_y, val_y = train_test_split(X, y, test_size=0.25)
dtrain = lgb.Dataset(train_x, label=train_y)
dval = lgb.Dataset(val_x, label=val_y)
model = lightgbm_tuner.train(
params,
dtrain,
valid_sets=[dtrain, dval],
best_params=best_params,
tuning_history=tuning_history,
num_boost_round=1000000,
verbose_eval=100,
early_stopping_rounds=100,
)
prediction = model.predict(val_x, num_iteration=model.best_iteration)
auc = roc_auc_score(val_y, prediction)
print("AUC: ", auc)
return {"best_params": best_params, "history": tuning_history}
def get_best_params(train_x: t.Any, train_y: t.Any) -> t.Any:
tr_x, val_x, tr_y, val_y = train_test_split(train_x,
train_y,
test_size=0.2,
random_state=1)
lgb_train = lgb.Dataset(tr_x, tr_y)
lgb_eval = lgb.Dataset(val_x, val_y)
best_params = {}
params = {
'objective': 'binary',
'metric': 'binary_logloss',
'boosting_type': 'gbdt',
}
best_params = {}
tuning_history = []
gbm = lightgbm_tuner.train(params,
lgb_train,
valid_sets=lgb_eval,
num_boost_round=1000,
early_stopping_rounds=20,
verbose_eval=50,
best_params=best_params,
tuning_history=tuning_history)
return best_params
def run(self):
tuning = False
train = self.load_data_frame("train")
valid = self.load_data_frame("valid")
test = self.load_data_frame("test")
# test期間の何日目のモデルかというのとdのmap
shift_day_map = {
i + 1: d
for (i, d) in enumerate(test.reset_index()["d"].unique())
}
# test のdが何曜日かというmap
dow_map = test.drop_duplicates("d").set_index(
"d")["dayofweek"].to_dict()
d = shift_day_map[self.target_day]
dow = dow_map[d]
valid = valid.query(f"dayofweek == {dow}")
test = test.query(f"d == {d}")
common_columns = [
"sell_price",
"lag_t28",
"lag_t29",
"lag_t30",
"rolling_mean_t7",
"rolling_mean_t30",
"rolling_std_t30",
"rolling_skew_t30",
"rolling_kurt_t30",
"rolling_mean_t60",
"rolling_mean_t90",
"rolling_std_t90",
"rolling_mean_t180",
"rolling_std_t180",
"price_change_t1",
"price_change_t365",
"rolling_price_std_t7",
"rolling_price_std_t30",
"snap_CA",
"snap_TX",
"snap_WI",
"wm_yr_wk",
"dayofweek",
# "ratio_by_store_t7",
# "ratio_by_store_t30",
# "ratio_by_item_t7",
# "ratio_by_item_t30",
"is_weekend",
"is_US_holiday",
"before_day_off",
"after_day_off",
# "lag_sales_mul_lag_price_dev_price",
# "item_predict",
]
id_columns = [
"item_id",
"dept_id",
"cat_id",
"store_id",
"state_id",
"event_name_1",
"event_type_1",
"event_name_2",
"event_type_2",
]
if not tuning:
try:
params: dict = json.load(
open(
f"./model_params/params_day_{str(self.target_day).zfill(2)}.json"
))
except FileNotFoundError as e:
print("not found tuned parameter!", e)
params = {
"boosting_type": "gbdt",
"metric": "rmse",
"objective": "poisson",
"n_jobs": -1,
"seed": 110,
"learning_rate": 0.05,
"bagging_fraction": 0.75,
"bagging_freq": 10,
"colsample_bytree": 0.75,
}
else:
params = {
"boosting_type": "gbdt",
"metric": "rmse",
"objective": "poisson",
"n_jobs": -1,
"seed": 110,
"learning_rate": 0.05,
}
if self.target_day < TEST_DAYS: # 28日目のときはshift_columnsがないから
feature_columns = train.columns
shift_column_names = [
f"shift_{day}" for day in range(self.target_day, TEST_DAYS)
]
# 全部使うと多すぎるので最新と7日ごとのものを使う
shift_column_names = list(
set(shift_column_names)
& set([shift_column_names[0]] +
["shift_7", "shift_14", "shift_21"]))
shift_columns = []
for col in shift_column_names:
shift_columns += feature_columns[
feature_columns.str.startswith(col)].tolist()
use_columns = common_columns + id_columns + shift_columns
elif self.target_day == TEST_DAYS: # 28日目なら
use_columns = common_columns + id_columns
else:
raise ValueError("something wrong when make use_columns")
dataset_train = lgb.Dataset(train[use_columns], train[TARGET])
dataset_valid = lgb.Dataset(valid[use_columns], valid[TARGET])
if tuning:
model = lightgbm_tuner.train(
params,
dataset_train,
num_boost_round=1000,
valid_sets=[dataset_train, dataset_valid],
early_stopping_rounds=200,
verbose_eval=100,
time_budget=3600,
)
json.dump(
model.params,
open(
f"./model_params/params_day_{str(self.target_day).zfill(2)}.json",
"w",
),
indent=4,
)
else:
model = lgb.train(
params,
dataset_train,
num_boost_round=1000,
valid_sets=[dataset_train, dataset_valid],
early_stopping_rounds=200,
verbose_eval=100,
)
valid["pred"] = model.predict(valid[use_columns])
test[TARGET] = model.predict(test[use_columns])
importance = pd.DataFrame({
"feature_name":
model.feature_name(),
"importance":
model.feature_importance("gain"),
})
result = Result(valid[["d", TARGET, "pred"]], test[["d", TARGET]],
importance)
self.dump(result)