The Ultimate wish-list for OpenNMT-Lua


(Vincent Nguyen) #1

Maybe it could be good to pin this thread and amend / checkmark if needed.
This is mainly done in comparison to the SOA of the community.

  1. GNMT attention (cf https://github.com/tensorflow/nmt) seems to bring +1 BLEU

  2. RNN “Deep Architecture” from Rico Sennrich & co https://arxiv.org/pdf/1707.07631.pdf

  3. Label Smoothing

  4. Multi-gpu which works and brings linear speed. cf https://github.com/OpenNMT/OpenNMT-tf/pull/54 requires in-graph replication.

  5. Cuda 9 compatibility ===> DONE

  6. Token batch / long sequence OOM prevention / Automatic memory management (batch length) ===> DONE

  7. Confidence score on translation

  8. transformer …:slight_smile: but OpenNMT-TF will do the job. ===> onmt-tf DONE


(jean.senellart) #2

Install Torch 7 with CUDA 9.1