pytorch-seq2seq icon indicating copy to clipboard operation
pytorch-seq2seq copied to clipboard

Published 20 hours ago • verified publisher icon IBM

Copy attention based decoder

Open AtmaHou opened this issue 6 years ago • 10 comments

You guys really did an excellent work with high scalability. Better than OpenNMT in my opinion :)

Copy mechanism could greatly enhance the performance of Seq2Seq model, would you guys implement a decoder with copy mechanism?

AtmaHou avatar Jul 12 '18 11:07 AtmaHou

@AtmaHou we'll see if it can be integrated in the next release. You're welcome to submit a PR. We need all the help we can get.

pskrunner14 avatar Sep 01 '18 02:09 pskrunner14

Thanks for your reply, it will be in great need.

AtmaHou avatar Sep 01 '18 17:09 AtmaHou

Any Update on this?

varunAI avatar Sep 19 '18 05:09 varunAI

@varunAI Yup working on it. Will push to develop as soon as it's done.

pskrunner14 avatar Sep 19 '18 07:09 pskrunner14

@pskrunner14 I can work on this if you need some help.

rrkarim avatar Oct 06 '18 22:10 rrkarim

@pskrunner14 Any news on this ?

Diego999 avatar Nov 05 '18 15:11 Diego999

@pskrunner14 is this PR https://github.com/IBM/pytorch-seq2seq/pull/172 reliable for the copy mechanism ?

Diego999 avatar Nov 06 '18 13:11 Diego999

@Diego999 no, I'm facing some issues with dimensions as mentioned and there seems to be a problem with the dynamic dictionary too. I'll get to this asap. Will post any updates on this thread.

pskrunner14 avatar Nov 06 '18 13:11 pskrunner14

@pskrunner14 Thank you for your quick answer !

Diego999 avatar Nov 06 '18 13:11 Diego999

@pskrunner14 Do you have any news about it ? Thanks

Diego999 avatar Jan 31 '19 16:01 Diego999