[GitHub] jeremiedb opened a new issue #5488: [WIP] RNN with bucketing and mask - R package

[gi...@git.apache.org]

jeremiedb opened a new issue #5488: [WIP] RNN with bucketing and mask - R package
URL: https://github.com/apache/incubator-mxnet/issues/5488
 
 
   Since there wasn't an API yet for bucketing and masking in the R-package, I started building some functionnalities to address this. Inference from iterator is also provided. Further work remains to render the interface more flexible to different structures (seq-to-seq vs. seq-to-single used to demo). I've tried to stay as close as possible to the current `model.R` for standard feedforward training. 
   
   I've detailed the approach and put the code here: 
   https://github.com/jeremiedb/mxnet_R_bucketing
   
   I was wondering if there were already plans to develop bucketing utilities in the R package. If not, would you see this approach as a decent one it and worth buildind further on it?
   
   There are some specificities with how I dealt with the iterarators, it assumes that a pre-processing is performed to put the data and label arrays in lists for each bucket. Also noticed that shared module options seems not supported in the `symbol_bind` operator, not sure if this may result in memory management issues. 
   
   In any case I can turn it into an example to add to the docs. CNN examples reaches 89% accuracy on IMDB sentiment in few minutes on laptop CPU. 
 
----------------------------------------------------------------
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