singa-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "wangwei (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (SINGA-137) To be compatible with Caffe's data format and neural net configuration
Date Tue, 27 Sep 2016 08:30:20 GMT

    [ https://issues.apache.org/jira/browse/SINGA-137?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15525491#comment-15525491
] 

wangwei commented on SINGA-137:
-------------------------------

Yes. We would include popular models like VGG.
1. We have provided the IO component for saving snapshots, which could be used like this https://github.com/apache/incubator-singa/blob/master/examples/imagenet/alexnet.cc#L289.
Adding a save/load function into net class (https://github.com/apache/incubator-singa/blob/master/include/singa/model/feed_forward_net.h)
would be useful. You are also welcomed to help implement it.
2. There are examples of vgg/alexnet/resnet using pysinga. We have trained CPP versions of
resent and alexnet over ImageNet, and would create a webpage to share them.

BTW, we have tested the size of vgg-16 using protobuf for serialization, which is about 600MB
https://issues.apache.org/jira/browse/SINGA-252 (would be merged into master soon).




> To be compatible with Caffe's data format and neural net configuration
> ----------------------------------------------------------------------
>
>                 Key: SINGA-137
>                 URL: https://issues.apache.org/jira/browse/SINGA-137
>             Project: Singa
>          Issue Type: New Feature
>            Reporter: wangwei
>            Assignee: Xiangrui
>              Labels: Caffe
>
> Caffe has many built-in models and a large user base.
> If we can train over Caffe's data (input data prepared using Caffe's tools) and neural
net configuration, it would help users to switch from Caffe to SINGA for distributed training.

> Here are two options.
> 1. update SINGA's variable name and some data structure to be consistent with Caffe.
> 2. write scripts/tools to convert Caffe's configuration protocol into SINGA protocol.



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

Mime
View raw message