In XGBoost, how to change eval function and keeping same objective?

Question

I would like to keep the objective as "reg:linear" and eval_metric as customized RMSE as follows:

def customised_rmse(preds, dtrain):
    N = len(preds)
    preds = np.array(preds)
    actual = np.array(dtrain.get_label())
    crmse = np.sqrt(np.sum(np.power((preds+1)*1.0/(actual+1) - 1.1, 2))/N)
    return "custom-rmse", crmse

When I ran, train function as follows ->

model = xgb.train(param_list, xgb_train, num_rounds, watchlist, None, customised_rmse, early_stopping_rounds=30)

Output I am getting is this ->

[0] train-rmse:15.1904  val-rmse:15.2102    train-custom-rmse:0.607681  val-custom-rmse:0.610993
Multiple eval metrics have been passed: 'val-custom-rmse' will be used for early stopping.
Will train until val-custom-rmse hasn't improved in 30 rounds.
[1] train-rmse:14.4936  val-rmse:14.5103    train-custom-rmse:0.588831  val-custom-rmse:0.589902

and so on...

My question is, it is getting optimized using "rmse" or "custom-rase" or both? What I have to do to remove "rmse" as it comes by default with "reg:linear"?

Answer 1

It is easier to understand if you use keyword arguments:

model = xgb.train(params=param_list, dtrain=xgb_train, 
        num_boost_round=num_rounds, evals=watchlist, obj=None, 
        feval=customised_rmse, early_stopping_rounds=30)

obj can be the objective function. feval can be the customized evaluation function.

rmse is being used to minimize error on the trainning data. However custom-rmse on the validation dataset will define when the model stop training.

Change obj to change the defult.