This repository includes an implementation of a convolutional seq2seq model by Chainer. This is derived from Chainer's official seq2seq example.

See Convolutional Sequence to Sequence Learning, Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, Yann N. Dauphin, arxiv, 2017. blog post, Torch code.

PYTHONIOENCODING=utf-8 python -u seq2seq.py -g=0 -i DATA_DIR -o SAVE_DIR -b 48 -e 100

DATA_DIR directory needs to include training pair data giga-fren.release2.fixed.en and giga-fren.release2.fixed.fr, in addition to validation pair data dev/newstest2013.en and dev/newstest2013.fr. Other pair dataset which has line-by-line sentence alignment can be used.

JSON log report file and a model snapshot will be saved in SAVE_DIR directory (if it does not exist, it will be automatically made).