[jira] [Commented] (SINGA-342) Support autograd

["Xue Wanqi (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/SINGA-342?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16389165#comment-16389165 ] 

Xue Wanqi commented on SINGA-342:
---------------------------------

I am working on this!

> Support autograd 
> -----------------
>
>                 Key: SINGA-342
>                 URL: https://issues.apache.org/jira/browse/SINGA-342
>             Project: Singa
>          Issue Type: New Feature
>            Reporter: wangwei
>            Priority: Major
>
> Autograd computes the partial derivatives of a complex function following chain rule (or back-propagation).
> To implement autograd, we can follow [https://stackoverflow.com/questions/32034237/how-does-numpys-transpose-method-permute-the-axes-of-an-array] and [https://github.com/HIPS/autograd.]
> In particular, we record the operation and operands of each result tensor during forward propagation. A graph is constructed based on the recorded information. Once the loss.backward() is triggered, we run backward propagation over the graph to compute the gradients of parameters.



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