This is a PyTorch implementation of the paper "Unsupervised Neural Machine Translation".
NOTICE: This repository is still in development since we believe there are still some bugs.
Our program is tested on the following machines:
Ubuntu 16.04
macOS 10.13.1
Our program requires the following programs/packages to run successfully:
python 3.6
numpy v1.13
pytorch v0.2
-
Clone the repository
Notice: Please downoad the master branch. Thanks.
-
Download the dataset
Download the dataset from the following link:
https://drive.google.com/file/d/11bDQXpQOiUw5hEwtbmr1IMxDHk-uMFjF/view?usp=sharing
Then unzip the downloaded file and put all the files into the folder UNMT/data
The folder data should contain the following files
cross_en
cross_fr
data_en.subword.clean
data_fr.subword.clean
emb_en.dms
emb_fr.dms
test_en_100.proc
test_fr_100.proc
vocab_en.pkl
vocab_fr.pkl
- Download our trained model
Download our trained model from the following link:
https://drive.google.com/file/d/1KmcwT5QkEfSTxpdPXD4SNWp-Cj5FiIbk/view?usp=sharing
Then unzip the file and put Nonee_2000.pkl into the folder UNMT (the same folder as train.py, main.py…). That is, the folder UNMT should contain a file callled Nonee_2000.pkl.
- Run the training program
To run the training program, type the following command
python3 main.py --train
Note that our program only works on python 3. Specifically, we run our training and testing program on python 3.6 (for more information of packages version, please checkout the depencies section).
The training program will save the model file into the folder UNMT and the log file into save/
- Run the testing program
To run the testing program type the following command:
python3 main.py --test --model_path Nonee_2000.pkl
Then the program will translate each English sentence in test_en_100.proc into a French sentece and save the translation into the file l1l2.txt (in the folder UNMT). The ground-truth translation is in the file test_fr_100.proc, which will be used later to evaluate the results of translation
To evaluate the translation results, type the following command:
./mosesdecoder/scripts/generic/multi-bleu.perl data/test_fr_100.proc < l1l2.txt
Notice that the it is normal to see BLEU score around 0.00 to 0.01.