You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@systemml.apache.org by "Mike Dusenberry (JIRA)" <ji...@apache.org> on 2017/03/10 05:14:37 UTC
[jira] [Commented] (SYSTEMML-1384) Revisit the weight and bias of
fully connected layer
[ https://issues.apache.org/jira/browse/SYSTEMML-1384?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15904456#comment-15904456 ]
Mike Dusenberry commented on SYSTEMML-1384:
-------------------------------------------
Yeah this is a fair thing to investigate. Just to be clear, the {{conv}} biases are column vectors, but the {{affine}}, {{rnn}}, and {{lstm}} are all row vectors.
> Revisit the weight and bias of fully connected layer
> ----------------------------------------------------
>
> Key: SYSTEMML-1384
> URL: https://issues.apache.org/jira/browse/SYSTEMML-1384
> Project: SystemML
> Issue Type: Sub-task
> Reporter: Niketan Pansare
>
> Since all our bias are column vector (which is consistent with Keras/Caffe), whereas bias of fully connected layer is a row-vector. Similarly, the weight that is passed to caffe is transpose of weights passed to SystemML (since both store in row-major NCHW format).
> Making the dimensions consistent will simplify loading of Caffe/Keras models.
> [~mwdusenb@us.ibm.com]
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)