Github user jkbradley commented on the pull request:
https://github.com/apache/spark/pull/4819#issuecomment-76779548
@MechCoder I had intended to use this internally and to expose a public
method. (The "evaluateEachIteration" method was the public one, but feel free
to think of a better name.) Yes, the evaluator was the loss metric, which
should probably be an optional parameter (defaulting to the training metric).
* [https://issues.apache.org/jira/browse/SPARK-6025]: This is the JIRA for
the public method.
* [https://issues.apache.org/jira/browse/SPARK-5972]: This is the JIRA for
the internal optimization.
I'm Ok with combining the 2 JIRAs in 1 PR since they are closely related.
For the internal optimization, the "residual" to store is not really the
residual but rather the cumulative prediction of the ensemble; that in turn can
be used to compute both the gradient and the error. (Note it will be important
to use the cached residual for computing the gradient, not just the objective.)
That may require adding some internal API to ensembles to permit prediction
from a pre-computed sum of trees' predictions.
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at [email protected] or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
