Code release for Non-Autoregressive Neural Machine Translation by Jiatao Gu, James Bradbury, Caiming Xiong, Victor O.K. Li, and Richard Socher.
Requires PyTorch 0.3, torchtext 0.2.1, and SpaCy.
The pipeline for training a NAT model for a given language pair includes:
run_alignment_wmt_LANG.sh
(runsfast_align
for alignment supervision)run_LANG.sh
(trains an autoregressive model)run_LANG_decode.sh
(produces the distillation corpus for training the NAT)run_LANG_fast.sh
(trains the NAT model)run_LANG_fine.sh
(fine-tunes the NAT model)