[jira] [Created] (SINGA-142) A bug in BPTTWorker::Backward() function with layers partially unrolled in neuralnet

["Zheng Kaiping (JIRA)" <ji...@apache.org>]

Zheng Kaiping created SINGA-142:
-----------------------------------

             Summary: A bug in BPTTWorker::Backward() function with layers partially unrolled in neuralnet
                 Key: SINGA-142
                 URL: https://issues.apache.org/jira/browse/SINGA-142
             Project: Singa
          Issue Type: Bug
            Reporter: Zheng Kaiping


This is in SINGA_HOME/src/work.cc, in “void BPTTWorker::Backward (int step, NeuralNet* net) {}” function. It affects training in neuralnets with partially unrolled layers.

To be specific, there is no need to consider non-unrolled layers (i.e, without parameter sharing) when aggregating gradients for parameters.



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