[ 
https://issues.apache.org/jira/browse/SINGA-348?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16435225#comment-16435225
 ] 

ASF subversion and git services commented on SINGA-348:
-------------------------------------------------------

Commit f42d4d076df55527c34fd6816580e097d908710b in incubator-singa's branch 
refs/heads/master from Wang Wei
[ https://git-wip-us.apache.org/repos/asf?p=incubator-singa.git;h=f42d4d0 ]

SINGA-348 Support autograd MLP Example

1. rename tensor.Tensor's attribute singa_tensor to data and add CTensor
for singa.Tensor

2. update Operation and autograd to simplify the code. The dependency is
recorded in a simple Counter: op -> #children ops

3. update the comments of some operations and autograd

4. made some changes in design

5. add operation.requires_grad, which is helpful in avoiding unnecessary 
calculations.

6. a question left: broadcast elementwise multiply seems not support, bring 
problems when sent CTensor gradient to loss.backward


> Support autograd MLP example
> ----------------------------
>
>                 Key: SINGA-348
>                 URL: https://issues.apache.org/jira/browse/SINGA-348
>             Project: Singa
>          Issue Type: New Feature
>            Reporter: Xue Wanqi
>            Priority: Major
>
> Implement a basic autograd system which supports parts of operations.
> Realize a MLP example base on the implemented autograd system.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

Reply via email to