[jira] [Commented] (SPARK-5114) Should Evaluator be a PipelineStage

["Max Moroz (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/SPARK-5114?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15352651#comment-15352651 ] 

Max Moroz commented on SPARK-5114:
----------------------------------

+1 for Evaluator to be a PipelineStage.

What options are being considered for storing its output? Perhaps it can be piggy-backed onto the (mutable) Pipeline and PipelineModel (e.g., using some kinda of special parameter that accumulates all evaluated results)? In fact, I'm not 100% sure evaluators need to be executed during the pipeline fit(). I guess it should be an option. We always need to evaluate it in transform() of course, but sometimes we may need to see the training metrics for debugging / analysis purposes (in that case, though, fit() should continue past the last estimator, all the way to the last estimator or evaluator).


> Should Evaluator be a PipelineStage
> -----------------------------------
>
>                 Key: SPARK-5114
>                 URL: https://issues.apache.org/jira/browse/SPARK-5114
>             Project: Spark
>          Issue Type: Sub-task
>          Components: ML
>    Affects Versions: 1.2.0
>            Reporter: Joseph K. Bradley
>
> Pipelines can currently contain Estimators and Transformers.
> Question for debate: Should Pipelines be able to contain Evaluators?
> Pros:
> * Schema check: Evaluators take input datasets with particular schema, which should perhaps be checked before running a Pipeline.
> * Intermediate results:
> ** If a Transformer removes a column (which is not done by built-in Transformers currently but might be reasonable in the future), then the user can never evaluate that column.  (However, users could keep all columns around.)
> ** If users have to evaluate after running a Pipeline, then each evaluated column may have to be re-materialized.
> Cons:
> * API: Evaluators do not transform datasets.   They produce a scalar (or a few values), which makes it hard to say how they fit into a Pipeline or a PipelineModel.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@spark.apache.org
For additional commands, e-mail: issues-help@spark.apache.org