XGBoost CV and early stopping

Question

I am trying to use XGBoost.cv with early stopping, based on mlogloss:

params = {'booster': 'gbtree', 'objective': 'multi:softprob',
          'num_class': len(le.classes_), 'eta': 0.1,
          'max_depth': 10, 'subsample': 1.0,
          'scale_pos_weight': 1, 'min_child_weight': 5,
          'colsample_bytree': 0.2, 'gamma': 0, 'reg_alpha': 0,
          'reg_lambda': 1, 'eval_metric': 'mlogloss'}

res = xgb.cv(params, dm_train, nfold=5,
        seed=42, early_stopping_rounds=10, verbose_eval=True,
        metrics={'mlogloss'}, show_stdv=False)

print(res)

My understanding of early stopping is that, if my eval metric does not improve for n rounds (in this case 10), the run will terminate. When I run this code, it terminates after 10 rounds, printing the output:

test-mlogloss-mean            
0:             6.107054               
1:             5.403606                    
2:             4.910938                   
3:             4.546221                    
4:             4.274113                  
5:             4.056968                  
6:             3.876368               
7:             3.728714                       
8:             3.599812                  
9:             3.485113

Test-mlogloss is falling with each epoch, therefore, I expected the run to not terminate (as accuracy must be improving). Where am I going wrong?

Thanks.

Chris Parry Chris Parry · Accepted Answer · 2016-11-09T04:38:59

2

votes

I didn't set the num_rounds param, which defaults to 10. Simple.

XGBoost CV and early stopping

1 Answers