Differences between different NMT frameworks

What are the differences between OpenNMT and Nematus? Does Nematus have more features than OpenNMT at this moment?

What are the differences between the OpenNMT and OpenNMT-Py? I’m mostly interested in the Python version. Is it possible to use it without forsaking too many important features?

Are there examples of OpenNMT or OpenNMT-Py being used in research? If so, can you point me to the repos and the corresponding papers?


OpenNMT is now widely used in research, most all of my groups work (nlp.seas.harvard.edu) and also work done by Systran.

I think Nematus of OpenNMT are relatively comparable frameworks, with OpenNMT having more support. See our report for comparison https://arxiv.org/pdf/1701.02810.pdf

OpenNMT-py is relatively new. Two major features missing are Multi-GPU and feature support. Also some advanced features. Unless you are changing the internals than OpenNMT and OpenNMT-py behave mostly the same. OpenNMT also has a server version to use from python.

1 Like