This is a basic implementation of attentional neural machine translation (Bahdanau et al., 2015, Luong et al., 2015) in Pytorch 0.4. It implements the model described in Luong et al., 2015, and supports label smoothing, beam-search decoding and random sampling. With 256-dimensional LSTM hidden size, it achieves 28.13 BLEU score on the IWSLT 2014 Germen-English dataset (Ranzato et al., 2015).

• nmt.py: contains the neural machine translation model and training/testing code.
• vocab.py: a script that extracts vocabulary from training data
• util.py: contains utility/helper functions

We provide a preprocessed version of the IWSLT 2014 German-English translation task used in (Ranzato et al., 2015) [script]. To download the dataset:

wget http://www.cs.cmu.edu/~pengchey/iwslt2014_ende.zip
unzip iwslt2014_ende.zip

Running the script will extract adata/ folder which contains the IWSLT 2014 dataset. The dataset has 150K German-English training sentences. The data/ folder contains a copy of the public release of the dataset. Files with suffix *.wmixerprep are pre-processed versions of the dataset from Ranzato et al., 2015, with long sentences chopped and rared words replaced by a special <unk> token. You could use the pre-processed training files for training/developing (or come up with your own pre-processing strategy), but for testing you have to use the original version of testing files, ie., test.de-en.(de|en).

The code is written in Python 3.6 using some supporting third-party libraries. We provided a conda environment to install Python 3.6 with required libraries. Simply run

conda env create -f environment.yml

Each runnable script (nmt.py, vocab.py) is annotated using dotopt. Please refer to the source file for complete usage.

First, we extract a vocabulary file from the training data using the command:

python vocab.py \
    --train-src=data/train.de-en.de.wmixerprep \
    --train-tgt=data/train.de-en.en.wmixerprep \
    data/vocab.json

This generates a vocabulary file data/vocab.json. The script also has options to control the cutoff frequency and the size of generated vocabulary, which you may play with.

To start training and evaluation, simply run data/train.sh. After training and decoding, we call the official evaluation script multi-bleu.perl to compute the corpus-level BLEU score of the decoding results against the gold-standard.