You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@singa.apache.org by "ASF subversion and git services (JIRA)" <ji...@apache.org> on 2018/08/24 02:59:00 UTC
[jira] [Commented] (SINGA-383) Add Separable Convolution for
autograd
[ https://issues.apache.org/jira/browse/SINGA-383?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16591086#comment-16591086 ]
ASF subversion and git services commented on SINGA-383:
-------------------------------------------------------
Commit ca70bdf3f02412f216d10e8d4ba6c265bdd139ee in incubator-singa's branch refs/heads/master from xuewanqi
[ https://git-wip-us.apache.org/repos/asf?p=incubator-singa.git;h=ca70bdf ]
SINGA-383 Add Separable Convolution for autograd
- let Conv2d layer support 'groups' paramters, for grouped convolution.
- implement Separable Convolution layer.
- add unit test case for new developed SeparableConv2d layer.
- the implemented SeparableConv2d layer has passed both unit test and network test.
> Add Separable Convolution for autograd
> --------------------------------------
>
> Key: SINGA-383
> URL: https://issues.apache.org/jira/browse/SINGA-383
> Project: Singa
> Issue Type: New Feature
> Reporter: wangwei
> Priority: Major
>
> This type of convolution is used in [Xception model|https://arxiv.org/pdf/1610.02357.pdf] and is supported by [other libraries|[https://github.com/pytorch/pytorch/issues/1708].]
>
> To implement it in Singa, we create a new operation (separable_conv_2d) by calling a depthwise_conv_2d (normal convolution with number of output channels=1, and number of groups = number of input channels); and then calling normal convolution with number of groups=1, and kernel size=1, i.e. pointwise convolution.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)