Awesome Open Source
Awesome Open Source


English, 中文

License: MIT Build Status

NJUNMT-pytorch is an open-source toolkit for neural machine translation. This toolkit is highly research-oriented, which contains some common baseline model:

  • DL4MT-tutorial: A rnn-base nmt model widely used as baseline. To our knowledge, this is the only pytorch implementation which is exactly the same as original model.(nmtpytorch is another pytorch implementation but with minor structure difference.)

  • Attention is all you need: A strong nmt model introduced by Google, which only relies on attenion mechanism.

Table of Contents


  • python 3.5+
  • pytorch 0.4.0+
  • tqdm
  • tensorboardX
  • sacrebleu


0. Quick Start

We provide push-button scripts to setup training and inference of transformer model on NIST Chinese-English Corpus (only on NJUNLP's server). Just execute under root directory of this repo

bash ./scripts/

for training and

# 3 means decoding on NIST 2003. This value
# can also be 4,5,6, which represents NIST 2004, 2005, 2006 respectively. 
bash ./scripts/ 3 

1. Build Vocabulary

First we should generate vocabulary files for both source and target language. We provide a script in ./data/ to build them in json format.

See how to use this script by running:

python ./scripts/ --help

We highly recommend not to set the limitation of the number of words and control it by config files while training.

2. Write Configuration File

See examples in ./configs folder. We provide several examples:

  • dl4mt_nist_zh2en.yaml: to run a DL4MT model on NIST Chinese to Enligsh
  • transformer_nist_zh2en.yaml: to run a Transformer model on NIST Chinese to English
  • transformer_nist_zh2en_bpe.yaml: to run a Transformer model on NIST Chinese to English using BPE.
  • transformer_wmt14_en2de.yaml: to run a Transformer model on WMT14 English to German

To further learn how to configure a NMT training task, see this wiki page.

3. Training

We can setup a training task by running

python -m src.bin.train \
    --model_name <your-model-name> \
    --reload \
    --config_path <your-config-path> \
    --log_path <your-log-path> \
    --saveto <path-to-save-checkpoints> \
    --valid_path <path-to-save-validation-translation> \

See detail options by running python -m src.bin.train --help.

During training, checkpoints and best models will be saved under the directory specified by option ---saveto. Suppose that the model name is "MyModel", there would be several files under that directory:

  • MyModel.ckpt: A text file recording names of all the kept checkpoints

  • MyModel.ckpt.xxxx: Checkpoint stored in step xxxx

  • A text file recording names of all the kept best checkpoints

  • Best checkpoint stored in step xxxx.

  • Final best model, i.e., the model achieved best performance on validation set. Only model parameters are kept in it.

4. Translation

When training is over, our code will automatically save the best model. Usually you could just use the final best model, which is named as, to translate. This model achieves the best performance on the validation set.

We can translation any text by running:

python -m src.bin.translate \
    --model_name <your-model-name> \
    --source_path <path-to-source-text> \
    --model_path <path-to-model> \
    --config_path <path-to-configuration> \
    --batch_size <your-batch-size> \
    --beam_size <your-beam-size> \
    --alpha <your-length-penalty> \

See detail options by running python -m src.bin.translate --help.

Also our code support ensemble decoding. See more options by running python -m src.bin.ensemble_translate --help




If you have any question, please contact [email protected]

Get A Weekly Email With Trending Projects For These Topics
No Spam. Unsubscribe easily at any time.
Python (890,046
Pytorch (22,633
English (6,789
Chinese (6,477
Nmt (523
Neural Machine Translation (224