You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@spark.apache.org by "Joseph K. Bradley (JIRA)" <ji...@apache.org> on 2015/06/24 22:10:04 UTC
[jira] [Commented] (SPARK-3382) GradientDescent convergence
tolerance
[ https://issues.apache.org/jira/browse/SPARK-3382?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14600058#comment-14600058 ]
Joseph K. Bradley commented on SPARK-3382:
------------------------------------------
[~lewuathe] I'm sorry for the iterative design, but I think this should be the final request. Could you please modify your PR to following the discussion in [SPARK-1503] for TFOCS? In summary:
* If the norm of the new weight vector is > 1, use relative tolerance (normalizing by the norm of the new weight vector).
* If the norm of the new weight vector is <= 1, use absolute tolerance (not normalizing).
If you prefer, I could also merge your PR and then send a follow-up modification myself. I appreciate your patience, and I'm glad we're taking the time to design this well!
> GradientDescent convergence tolerance
> -------------------------------------
>
> Key: SPARK-3382
> URL: https://issues.apache.org/jira/browse/SPARK-3382
> Project: Spark
> Issue Type: Improvement
> Components: MLlib
> Affects Versions: 1.1.0
> Reporter: Joseph K. Bradley
> Priority: Minor
>
> GradientDescent should support a convergence tolerance setting. In general, for optimization, convergence tolerance should be preferred over a limit on the number of iterations since it is a somewhat data-adaptive or data-specific convergence criterion.
--
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