MGNMT

This repo is the implementation of Mirror-Generative Neural Machine Translation.

0. Preparation

MGNMT is implemented based on NJUNMT. As for data preprocessing and others please refer to README-NJUNMT

This repo has been tested with python 3.7 and pytorch 1.2.0 on 2080ti and V100.

We should prepare following datasets

1. parallel bilingual training dataset with bpe codes, vocabs
2. dev and test datasets
3. (optional) non-paralell bilingual training dataset, i.e., monolingual dataset for source and target language

Please see configurations in configs/ed/ for example.

1. Training

Parallel Data Only

sh train_mgnmt_iwslt_de2en_ed.sh

This script uses configuration in configs/ed/iwslt_de2en_mgnmt.yaml

Using Non-Parallel Data

sh train_mgnmt_np_iwslt_de2en_ed.sh

This script uses configuration in configs/ed/iwslt_de2en_mgnmt_np.yaml

• for domain adaptation, using NP_train_mode as "finetune" may be better
• for other scenarios, using "hybrid" to enable hybrid training of both P and NP data
• it would be better to start using NP data after the model has been well trained on parallel data, which should probably take a bit larger NP_warmup_step.

Multi-GPU Training

Specify the environment variable CUDA_VISIBLE_DEVICES. For example:

export CUDA_VISIBLE_DEVICES=0,1,2,3

Currently this repo only support multi gpus on a single node.

2. Generation

sh translate_mgnmt.sh

Note that we have both src2tgt and tgt2src translation direction, so we should specify --src_lang to either "src" or "tgt" to specify the expected translation direction.

• Use beta to control the weight of decoding target LM fusion.

• Use --reranking to enable reconstructive reranking.

• Use gamma to control the weight of source LM in reconstructive reranking.

• --batch_size determines batch size based on the number of tokens.

3. Others

Important files

src/tasks/mgnmt.py

src/models/mgnmt.py

src/decoding/iterative_decoding.py

src/modules/variational_inferrer.py

and configs in ./configs/

Tips

1. How well the inference model learns affects the performance of MGNMT. So having a proper KL loss is important. However, KL annealing is VERY tricky. Try to adjust x0 (KL annealing step) to find an optimal KL loss that is neither large nor small too much. A KL loss of ~3 to 6 works well in practice.

2. For the best training efficiency, it is highly recommended that do not use reranking and LM fusion in back-translation or bleu_validation. Plus you may not know the optimal weights to interpolate LMs. You can find the best weights in inference after you get the training done.

TODO

• continue refactoring codebase
• pretrained models
• hyperparameters and settings for other experiments in the paper

Citation

@inproceedings{zheng2020mirror,
    title={Mirror-Generative Neural Machine Translation},
    author={Zaixiang Zheng and Hao Zhou and Shujian Huang and Lei Li and Xin-Yu Dai and Jiajun Chen},
    booktitle={International Conference on Learning Representations},
    year={2020},
    url={https://openreview.net/forum?id=HkxQRTNYPH}
}