OpenNMT-py

Why we included this project

For teams that would rather train their own neural machine translation model than pay for a hosted API, OpenNMT-py has long been the default PyTorch toolkit. The codebase spans the whole sequence-to-sequence pipeline, preprocessing through training to inference, and its modular layout makes it easy to try out new ideas for translation, language modeling, and summarization. Researchers and small NLP groups after a transparent, self-contained baseline for a custom translation system will find it straightforward to read and extend. One honest caveat: the project is no longer actively developed. The maintainers have moved all new work to its successor, Eole, so a fresh production project should factor in that handoff. As a mature, well-documented reference implementation to learn from or adapt, though, it still holds up.