[jvm-packages] Automatically set maximize_evaluation_metrics if not explicitly given in XGBoost4J-Spark #4446
Conversation
hcho3
changed the title
|
thanks, would you please also update https://xgboost.readthedocs.io/en/latest/jvm/xgboost4j_spark_tutorial.html#early-stopping accordingly? |
|
@CodingCat Document is updated in new commits. |
|
@hcho3 jenkins says |
|
@CodingCat Looks like we need to expand the volume for the slave node |
|
It seems that 2 failures of checks are not caused by my code change but the disk volume. |
|
I’m looking at the disk space issue. |
|
@CodingCat @rongou I went ahead and increased the disk space. |
| @@ -160,7 +160,7 @@ object XGBoost extends Serializable { | |||
| .map(_.toString.toInt).getOrElse(0) | |||
| val overridedParams = if (numEarlyStoppingRounds > 0 && | |||
| !params.contains("maximize_evaluation_metrics")) { | |||
| if params.contains("custom_eval") { | |||
| if (params.contains("custom_eval")) { | |||
There was a problem hiding this comment.
when num_early_stopping_rounds is non-zero, eval_metric is a built-in metric, the custom_eval is still in params, just the value is null, so it will still throw an exception.
So the if condition should be if (params.contains("custom_eval") && params("custom_eval") != null) or remove custom_eval from params.
There was a problem hiding this comment.
Hmmm....I think I missed some part of this PR
The reality is early stopping doesn’t work with custom metrics
This pull request is aimed at resolving "[jvm-packages] false should be the default value of maximize_evaluation_metrics"(#4422).
The metric names to maximize are aligned with xgboost.callback.early_stop in the Python package.
To avoid duplicating codes in both
XGBoostClassifierandXGBoostRegressor, parametermaximize_evaluation_metricsis set in XGBoost.scala when not provided by user.closes #4422