[jira] [Commented] (SPARK-11136) Warm-start support for ML estimator

["Nick Pentreath (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/SPARK-11136?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15192975#comment-15192975 ] 

Nick Pentreath commented on SPARK-11136:
----------------------------------------

A question about the API design - it seems to me that it would be good to have the initial model (if it exists) set up the default params. e.g.
{code}
val model1 = new KMeans()
  .setK(10)
  .setInitSteps(5)
  .setTol(1e-3)
  .setInitMode("random")
  .fit(dataset)

val model2 = new KMeans()
  .setInitialModel(model1)
  .fit(dataset)
{code}
Here {{model2}} automatically is trained with the same {{k}}, {{tol}} and {{initMode}} as {{model1}} - but in this case the {{initSteps}} would be overridden to {{1}}. If the user wants to adjust those then they can of course set the params. Thoughts?


> Warm-start support for ML estimator
> -----------------------------------
>
>                 Key: SPARK-11136
>                 URL: https://issues.apache.org/jira/browse/SPARK-11136
>             Project: Spark
>          Issue Type: Sub-task
>          Components: ML
>            Reporter: Xusen Yin
>            Priority: Minor
>
> The current implementation of Estimator does not support warm-start fitting, i.e. estimator.fit(data, params, partialModel). But first we need to add warm-start for all ML estimators. This is an umbrella JIRA to add support for the warm-start estimator. 
> Treat model as a special parameter, passing it through ParamMap. e.g. val partialModel: Param[Option[M]] = new Param(...). In the case of model existing, we use it to warm-start, else we start the training process from the beginning.



--
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