Source code for the Transformer model used in the paper "Neural Machine Translation of Text from Non-Native Speakers" https://arxiv.org/abs/1808.06267. Require Python3.6 and Pytorch 1.0.

Pretty much just the transformer model, but the validator validates on different dev sets at a time and keeps best checkpoint for each dev set.

How to train a new model

Write a config function in configurations.py. Then run python3 -m nmt --proto config_name.

The best checkpoints are saved in nmt/saved_models/model_name. To decode with a checkpoint: python3 -m nmt --proto config_name --mode translate --model-file nmt/saved_models/model_name/checkpoint_name.path --input-file path_to_input_file.

References

A lot of code / scripts are borrowed from:

• http://nlp.seas.harvard.edu/2018/04/03/attention.html
• https://github.com/pytorch/fairseq
• https://github.com/EdinburghNLP/nematus
• https://github.com/moses-smt/mosesdecoder

...