singa-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "ASF subversion and git services (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (SINGA-109) Refine bridge layers for inter-node layer communication
Date Fri, 04 Dec 2015 11:57:11 GMT

    [ https://issues.apache.org/jira/browse/SINGA-109?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15041488#comment-15041488
] 

ASF subversion and git services commented on SINGA-109:
-------------------------------------------------------

Commit bd2e3453ca01e7bf9bf6724b4717d956c6e00290 in incubator-singa's branch refs/heads/master
from WANG Sheng
[ https://git-wip-us.apache.org/repos/asf?p=incubator-singa.git;h=bd2e345 ]

SINGA-109 Refine bridge layers

re-implement bridge layers for model partition
 * move socket operations for sending/receiving blobs from worker into layers,
   so that it is transparent to users who will implment TrainOneBatch
 * when initialing worker, it will create a socket instance and pass it to all
   bridge layers using layer.MakePaired() function


> Refine bridge layers for inter-node layer communication
> -------------------------------------------------------
>
>                 Key: SINGA-109
>                 URL: https://issues.apache.org/jira/browse/SINGA-109
>             Project: Singa
>          Issue Type: Improvement
>            Reporter: Sheng Wang
>            Assignee: Sheng Wang
>
> Previously, sending data and gradient blobs between remote bridge layers is explicitly
handled by worker.
> Train One Batch functions need to manage sending/receiving data before doing real work.
> This ticket is to encapsulate bridge layer communications, and put them inside their
own compute feature/gradient functions.
> So that the worker do not need to know how the neuralnet is partitioned and communicated.



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

Mime
View raw message