You can use your own pre-trained word-embeddings or use
(i) GloVe embeddings (download here)
(ii) Vietnamese wiki embeddings (download here),
which are the default embeddings in the implementation
use python vocab.py --train-src=[vietnamese texts] --train-tgt =[english texts] [output_file] to generate vocab dictionaries as json
use sh run.sh train_local_cuda to train using pre-configured settings on the given toy dataset, or
use python run.py train --train-src =[vietnamese training data] --train-tgt =[english training data] --dev-src =[vietnamese dev data] --dev-tgt =[english dev data] --vocab =[json vocab file] (Optional: include --cuda to train on GPU)
use python run.py decode [model_path] [Vietnamese text file] [English text file] [output_file] to perform prediction and validation
use python run.py translate [model_path] [ input vietnamese text file] to translate
This implementation is based on the starter code given by Stanford cs224n's assignment 4