from sklearn.metrics import make_scorer
from sklearn.model_selection import GridSearchCV


def my_custom_loss_func(y_true, y_pred):
    return np.mean(np.abs((y_pred - y_true)/y_true))

mape = make_scorer(my_custom_loss_func, greater_is_better=False)

param_test1 = {
'n_estimators': [700,800]
}
init_param = { 
'booster':'gbtree',
'learning_rate': 0.2,
'n_estimators': 700,
'max_depth': 6,
'min_child_weight': 1,
'colsample_bytree':0.7,
'subsample': 0.8,
'reg_alpha': 0.05,
'gamma': 0.1,
'reg_lambda': 3,
'silent': 1,
'seed': 2020}
gsearch1 = GridSearchCV(
estimator=XGBRFRegressor(**init_param), 
param_grid=param_test1, cv=5, verbose=1, n_jobs=4, scoring=mape)
gsearch1.fit(new_train_X, train_y)