RetroGAN

Code for paper: RetroGAN: A Cyclic Post-Specialization System for Improving Out-of-Knowledge and Rare Word Representations coming out in ACL FINDINGS 2021.

Getting Started

1. Create an environment with conda

   conda create -n retrogan python=3.8

2. Activate the environment

   conda activate retrogan

3. Install requirements:

   cd Code

   pip install -r requirements.txt

4. Run experiments:

   python retrogan_trainer.py --epochs 150 ../Data/ft_all_unseen.txt ../Data/ft_all_unseen_retrofitted.txt retrogan_nb_full_50epochs_alllosses models/retrogan_nb_full_50epochs_alllosses

Parameter Tuning

To run the parameter fine tuning, after installing the requirements, run:

python rg_ray_parameter_tuning.py

NOTE: It may be the case that you run into an error with the msgpack library (something about an unknown argument), if that is the case, uninstall, and reinstall the msgpack libary:

pip uninstall msgpack

pip install msgpack

Tests

You can also run the ablation, full/disjoint, and percentage tests that were reported in the paper with the respective scripts:

• ablationtests.sh
• full_and_oov_tests.sh
• oovtests.sh

Expected results

Here is a table with the expected results:

Ablation study results

One by one removal

Here is a higher resolution version of the plots given in ablation study. These results can be generated by running the script:

cd Scripts

sh ablationtests.sh

Toggling of losses

Here is a higher resolution version of the plots given in the toggle ablation study These results can be generated by running the script:

cd Scripts

sh ablationtests-2.sh

Citing

TBD