[
https://issues.apache.org/jira/browse/SPARK-7407?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14531898#comment-14531898
]
Joseph K. Bradley commented on SPARK-7407:
------------------------------------------
I hope we can make this change without changing the user-facing API. That
seems very doable for Scala, where ParamMap is a class. It sounds harder for
Python. Should we make it a class there too?
> Use uid and param name to identify a parameter instead of the param object
> --------------------------------------------------------------------------
>
> Key: SPARK-7407
> URL: https://issues.apache.org/jira/browse/SPARK-7407
> Project: Spark
> Issue Type: Improvement
> Components: ML
> Affects Versions: 1.4.0
> Reporter: Xiangrui Meng
> Assignee: Xiangrui Meng
>
> Transferring parameter values from one to another have been the pain point in
> the ML pipeline implementation. Because we use the param object as the key in
> the param map, we have to correctly copy them when making a copy of the
> transformer, estimator, and models. This becomes complicated when
> meta-algorithms are involved. For example, in cross validation:
> {code}
> val cv = new CrossValidator()
> .setEstimator(lr)
> .setEstimatorParamMaps(epm)
> {code}
> When we make a copy of `cv` with extra params that contain estimator params,
> {code}
> cv.copy(ParamMap(cv.numFolds -> 3, lr.maxIter -> 10))
> {code}
> we need to make a copy of the `lr` object as well and map `epm` to use the
> new param keys from the old `lr`. This is quite error-prone, especially if
> the estimator itself is another meta-algorithm.
> Using uid + param name as the key in param maps and using the same uid in
> copy (and between estimator/model pairs) would simplify the implementations.
> We don't need to change the keys since the copied instance has the same id as
> the original instance. And it is easier to find models from a fitted pipeline.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
