AmuNMT is an efficient Neural Machine Translation framework written in pure C++ with minimal dependencies. It has mainly been developed at the Adam Mickiewicz University in Poznań (AMU) and at the University of Edinburgh.
It is currently being deployed in multiple European projects and is the main translation and training engine behind the neural MT launch at the World Intellectual Property Organization.
Main features:
- Fast multi-gpu training and translation
- Compatible with Nematus and DL4MT
- Efficient pure C++ implementation
- Permissive open source license (MIT)
- more details...
If you use this, please cite:
Marcin Junczys-Dowmunt, Tomasz Dwojak, Hieu Hoang (2016). Is Neural Machine Translation Ready for Deployment? A Case Study on 30 Translation Directions (https://arxiv.org/abs/1610.01108)
@InProceedings{junczys2016neural,
title = {Is Neural Machine Translation Ready for Deployment? A Case Study
on 30 Translation Directions},
author = {Junczys-Dowmunt, Marcin and Dwojak, Tomasz and Hoang, Hieu},
booktitle = {Proceedings of the 9th International Workshop on Spoken Language
Translation (IWSLT)},
year = {2016},
address = {Seattle, WA},
url = {http://workshop2016.iwslt.org/downloads/IWSLT_2016_paper_4.pdf}
}
More information on https://amunmt.github.io
Ubuntu 16.04 LTS (tested and recommended). For Ubuntu 16.04 the standard packages should work. On newer versions of Ubuntu, e.g. 16.10, there may be problems due to incompatibilities of the default g++ compiler and CUDA.
- CMake 3.5.1
- GCC/G++ 5.4
- Boost 1.61
- CUDA 8.0
Ubuntu 14.04 LTS (tested). A newer CMake version than the default version is required and can be installed from source.
- CMake 3.5.1 (due to CUDA related bugs in earlier versions)
- GCC/G++ 4.9
- Boost 1.54
- CUDA 7.5
The CPU-only version will automatically be compiled if CUDA cannot be detected by CMake. Only the translator will be compiled, the training framework is strictily GPU-based.
Tested on different machines and distributions:
- CMake 3.5.1
- The CPU version should be a lot more forgiving concerning GCC/G++ or Boost versions.
Clone a fresh copy from github:
git clone https://github.com/amunmt/amunmt
The project is a standard CMake out-of-source build:
cd amunmt
mkdir build
cd build
cmake ..
make -j
If run for the first time, this will also download Marian -- the training framework for AmuNMT.
Marian is the training framework of AmuNMT. Assuming corpus.en and corpus.ro are
corresponding and preprocessed files of a English-Romanian parallel corpus, the
following command will create a Nematus-compatible neural machine translation model.
./amunmt/build/marian \
--train-sets corpus.en corpus.ro \
--vocabs vocab.en vocab.ro \
--model model.npz
See the documentation for a full list of command line options or the examples for a full example of how to train a WMT-grade model.
If a trained model is available, run:
./amunmt/build/amun -m model.npz -s vocab.en -t vocab.ro <<< "This is a test ."
See the documentation for a full list of command line options or the examples for a full example of how to use Edinburgh's WMT models for translation.
- Translating with Amun:
The files and scripts described in this section can be found in
amunmt/examples/translate. They demonstrate how to translate with Amun using Edinburgh's German-English WMT2016 single model and ensemble. - Training with Marian: The files
and scripts described in this section can be found in
amunmt/examples/training. They have been adapted from the Romanian-English sample from https://github.com/rsennrich/wmt16-scripts. We also add the back-translated data from http://data.statmt.org/rsennrich/wmt16_backtranslations/ as desribed in Edinburgh's WMT16 paper. The resulting system should be competitive or even slightly better than reported in that paper. - Winning system of the WMT 2016 APE shared task: This page provides data and model files for our shared task winning APE system described in Log-linear Combinations of Monolingual and Bilingual Neural Machine Translation Models for Automatic Post-Editing.
The development of AmuNMT received funding from the European Union's Horizon 2020 Research and Innovation Programme under grant agreements 688139 (SUMMA; 2016-2019) and 645487 (Modern MT; 2015-2017), the Amazon Academic Research Awards program, and the World Intellectual Property Organization.