[
https://issues.apache.org/jira/browse/SINGA-54?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14698170#comment-14698170
]
ASF subversion and git services commented on SINGA-54:
------------------------------------------------------
Commit a584da62c14720768acbf9a82a6069faa48b91d2 in incubator-singa's branch
refs/heads/master from Wei Wang
[ https://git-wip-us.apache.org/repos/asf?p=incubator-singa.git;h=a584da6 ]
SINGA-54 Refactor job configuration to move fields in ModelProto out
Tested with mnist and cifar examples.
Four components are necessary for submitting a job, namely, neuralnet, alg,
updater and cluster.
The configuration is now consistent with the MM paper.
> Refactor job configuration to move fields in ModelProto out
> -----------------------------------------------------------
>
> Key: SINGA-54
> URL: https://issues.apache.org/jira/browse/SINGA-54
> Project: Singa
> Issue Type: Improvement
> Reporter: wangwei
>
> There are two components in JobProto now, namely ClusterProto and ModelProto,
> for cluster topology configuration and model configuration, respectively.
> To be consistent with the components introduced in the [ACM Multimedia Open
> Source Competition
> paper|http://www.comp.nus.edu.sg/~ooibc/singaopen-mm15.pdf], we need to
> refactor the JobProto.
> The plan is to move fields inside ModelProto out as the first level children
> of JobProto. Then the ModelProto can be removed.
> After refactoring, the JobProto will contain the following main components:
> * net for neural net configuration, which consists of a set of layers
> * train_one_batch for the TrainOneBatch algorithm
> * updater for updater configuration
> * cluster for cluster topology configuraltion.
> Other fields include display frequency, checkpoint frequency, workspace, etc.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
