shengminp / vt-stower

Prerequisites

conda create -n stower python=3.7
conda activate stower

• transformers >= 4.4.2

  pip install transformers
  

• Install our fairseq repo

  cd VT-STOWER
  pip install --editable ./
  

• hydra = 1.0.3

  pip install hydra-core==1.0.3
  

• boto3

  pip install boto3
  

Our outputs

For simplity, we directly put our outputs under the folder outputs.

Data Preprocessing

Dataset is stored at preprocess folder. To preprocess dataset, run the command:

cd preprocess
bash prepare.sh

After preprocessing, three databins will be generated:

• yelp-databin: yelp dataset for sentiment transfer;
• gyafc-databin: gyafc dataset for formality transfer;
• en-hi-databin: Hindi-Hinglish dataset for code-switching transfer.

Training

Training scripts are stored in the train folder

cd train
bash train-{yelp,gyafc,en-hi}.sh

Training bash files include stage I, classifier, and stage II training process.

Style Transfer

To transfer style, please run the commands under generate folder:

cd generate
bash generate-{yelp,gyafc,en-hi}.sh

Outputs are named as *.out