[
https://issues.apache.org/jira/browse/SPARK-5844?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15802264#comment-15802264
]
Joseph K. Bradley commented on SPARK-5844:
------------------------------------------
Thanks for the ideas! There is a key question here: Should the logic for
caching live in Pipeline or in individual models? I hope that it can live
within Pipeline so that we can avoid implementing it separately for each model.
Here's what I'm imagining:
* Inspect the ParamGrid to figure out which Params belong with which
PipelineStage
* Fit the first stage on its set of ParamMaps, and get back the corresponding
set of Transformers.
* Cache...if it makes sense
* For each of these Transformers,
** Fit the next stage on its set of ParamMaps.
** etc., recursively
The main issue I see is that we may not want to cache between every stage. I
don't know the best way to handle this issue, but I bet we could sketch out a
reasonable design, with some Params for specifying preferences.
I'll also comment on the related JIRA [SPARK-19071].
> Optimize Pipeline.fit for ParamGrid
> -----------------------------------
>
> Key: SPARK-5844
> URL: https://issues.apache.org/jira/browse/SPARK-5844
> Project: Spark
> Issue Type: Improvement
> Components: ML
> Affects Versions: 1.3.0
> Reporter: Joseph K. Bradley
>
> This issue was brought up by [~prudenko] in [this JIRA |
> https://issues.apache.org/jira/browse/SPARK-4766].
> **Proposal**:
> When Pipeline.fit is given an array of ParamMaps, it should operate
> incrementally:
> * For each set of parameters applicable to the first PipelineStage,
> ** Fit/transform that stage using that set of parameters.
> ** For each set of parameters applicable to the second PipelineStage,
> *** etc.
> This is essentially a depth-first search on the parameters, where each
> node/level in the search tree is a PipelineStage and each node's child nodes
> correspond to the set of ParamMaps for that PipelineStage.
> This will avoid recomputing intermediate RDDs during model search.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
