You are viewing a plain text version of this content. The canonical link for it is here.

Posted to commits@mxnet.apache.org by GitBox <gi...@apache.org> on 2019/01/10 09:13:37 UTC

[GitHub] jiangzhengkai opened a new issue #13829: Mxnet Training deterministic

jiangzhengkai opened a new issue #13829: Mxnet Training deterministic
URL: https://github.com/apache/incubator-mxnet/issues/13829
 
 
   Hi, I'm using mxnet for research experiments. However, the experiments results is not deterministic.
   
   I'm using np, random, mxnet.nd, so i fix the seed for all of them. besides, i add os.environ['MXNET_ENFORCE_DETERMINISM'] = '1' and disable cudnn option in convolution option.
   
   here is the log for different two times
   ![image](https://user-images.githubusercontent.com/19780166/50958079-d5089600-14fa-11e9-9dfc-3930a43608c5.png)
   ![image](https://user-images.githubusercontent.com/19780166/50958091-da65e080-14fa-11e9-9196-61d90977ee81.png)
   
   It starts different when 10 iters.
   
   here is the debugging process.
   I first confirm that the load data is totally same. when disables backward process, the forward loss is totally same. So the problem is in the backward process.
   
   
   there is any random process for bachward?
   
   anyone helps me?
   
   Thanks
   
   
   
   

----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
 
For queries about this service, please contact Infrastructure at:
users@infra.apache.org


With regards,
Apache Git Services