Reformer

PyTorch original implementation of Neural Machine Translation with Joint Representation. It is modified from fairseq-0.6.0.

Requirements

• PyTorch version == 1.1.0
• Python version == 3.6.7

Get started

Installation

You need to install it first:

git clone https://github.com/lyy1994/reformer.git
cd reformer
pip install -r requirements.txt
python setup.py build develop

Structure

Before running the training and decoding scripts, we implicitly make assumptions on the file directory structure:

|- reformer (code)
|- data
   |- data-bin
      |- BINARIZED_DATA_FOLDER
   |- RAW_DATA_FOLDER
      |- train (training set raw text)
      |- valid (validation set raw text)
      |- test (test set raw text)
|- checkpoints
   |- torch-1.1.0
      |- EXPERIMENT_FOLDER
|- toolkit
   |- multi-bleu.perl

Usage

To train a model, run:

cd reformer/scripts
sh train.sh

To decode from the trained model, run:

sh decode.sh

If you would like to customize the configuration, please modify train.sh for training and decode.sh for decoding.

The table below summarizes the scripts for reproducing our experiments:

| Dataset | Script | |---|---| | IWSLT14 German-English | iwslt-train.sh | | NIST12 Chinese-English | nist-train.sh |

Citation

@inproceedings{li2020aaai,
  title = {Neural Machine Translation with Joint Representation},
  author = {Yanyang Li and Qiang Wang and Tong Xiao and Tongran Liu and Jingbo Zhu},
  booktitle = {Proceedings of the Thirty-Fourth AAAI Conference on Artificial Intelligence},
  year = {2020},
}