A problem when fitting XGBRegressor with xgb_model initialized

#1

Hello! I want to continue the training of a model with parameter xgb_model assigned in the fit function as follows:

xgb_model_full_data = xgboost_model.fit(train_features, train_labels, eval_metric=custimized_error, eval_set=[(validation_features, validation_labels)], xgb_model=“test_xgb.bin”, early_stopping_rounds=1000)

The training process runs smoothly and the debug information show the final score assume 2.8 with early stopping.

When the fitting is finished, I do the following:
predictions = xgb_model_full_data.predict(validation_features)

The I calculate the score using the predictions and validation_labels, I got 3.23.

I do not understand why the scores are different since the predication is done on the same validation set.

I tried to train the regressor without initializing the xgb_model parameter, and repeat above procedure. The two scores show exactly same value.

What am I missing here?

Thanks and Happy new year.

#2

Which XGBoost version are you using?

#3

Hi hcho3,

the version is 1.2.0.

Should the training set used to continue training be exactly the same as the one used in test_xgb.bin?
Even a subset of the training set used in test_xgb.bin will lead to problem?

Some additional information:

  1. random seed is always the same for different run.