A Neural Machine Translation framework for training large-scale networks on multiple nodes with multiple GPUs.

• Python 2.7
• deepy >= 0.2
• theano
• numpy

1. Clone neuralmt

git clone https://github.com/zomux/neuralmt
export PYTHONPATH="$PYTHONPATH:/path/to/neuralmt"

2. Create a directory for WMT data

export WMT_ROOT="/path/to/your_wmt_folder"
mkdir $WMT_ROOT/text
mkdir $WMT_ROOT/models

3. Tokenize de-en training corpus, and rename them to following filenames

• $WMT_ROOT/text/wmt15.de-en.de
• $WMT_ROOT/text/wmt15.de-en.en

4. Build training data

cd /path/to/neuralmt
python examples/gru_search/preprocess.py

5. Train on 3 GPUs

python -m deepy.multigpu.launch examples/gru_search/train.py gpu0 gpu1 gpu2

6. Wait for several days

7. Test your model

python examples/gru_search/test.py

(The test script only translate one sample sentence, you can modify it to translate a text file)

Training on multiple machine is still in development.

Although the current framework for parallelism shall be extended to multiple machine easily, it require some works.

• WMT15 German-English task (using the model in the example)
• BLEU: 21.29
• Duration: 2.5 days with 3 Titan X GPUs

Raphael Shu, 2016