[
https://issues.apache.org/jira/browse/SYSTEMML-2406?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
LI Guobao updated SYSTEMML-2406:
--------------------------------
Description: In the EPOCH update strategy, we should aggregate the
gradients generated after the local model updating and push them when each
epoch finishes. Otherwise, the intermediate gradients will be lost after each
local model updating and so finally, the global model should be updated with
the accumulated gradients of this epoch. (was: In the EPOCH update strategy,
we should aggregate the gradients generated after the local model updating and
push them when each epoch finishes. Otherwise, the intermediate gradients will
be lost after each local model updating and at last, the global model will be
updated with only the gradients of last iteration. )
> Paramserv low accuracy with EPOCH update frequency
> --------------------------------------------------
>
> Key: SYSTEMML-2406
> URL: https://issues.apache.org/jira/browse/SYSTEMML-2406
> Project: SystemML
> Issue Type: Bug
> Reporter: LI Guobao
> Assignee: LI Guobao
> Priority: Major
> Fix For: SystemML 1.2
>
>
> In the EPOCH update strategy, we should aggregate the gradients generated
> after the local model updating and push them when each epoch finishes.
> Otherwise, the intermediate gradients will be lost after each local model
> updating and so finally, the global model should be updated with the
> accumulated gradients of this epoch.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
