README.md

Attention: below commands should be executed under fairseq folder.

Enviroments

We recommend to create a new conda enviroment (named eisl):

conda create -n eisl python==3.7  pytorch==1.5.0 torchvision==0.6.0 cudatoolkit=10.2 -c pytorch

Then activate the conda enviroment:

conda activate eisl

Install the required package by running the script:

bash install_pkgs.sh

Download the Pretrained Models

For Multi30k dataset, we provide all the trained models (BART) mentioned in the Section 4.1. You can run the below command to download and extract the models ($noise should be one of shuffle, repetition, blank, multiple)

cd ckpts 
bash download_models.sh $noise

Download the Processed Datasets

You can download the processed data (noisy Multi30k data) by run the script

bash download_datasets.sh

Please refer to fairseq/examples/translation for more details of the data preprocessing.

Code

Work In Progress

Generated Results

The generated files from test set are in log/hypo/hypo. The original source is test.de and the target is test.en. For different noise (e.g., shuffle), *hypo.txt is the generated files of different loss and different scale of noise, and *bleu is the BLEU score of each target sentence.