an-seunghwan / cdg-vae

Official pytorch implementation codes for ECAI-2023 accepted paper "Causally Disentangled Generative Variational AutoEncoder"

Causally Disentangled Generative Variational AutoEncoder

This repository is the official implementation of Causally Disentangled Generative Variational AutoEncoder (26th European Conference on Artificial Intelligence ECAI 2023 accepted paper) with pytorch.

NOTE: This repository supports WandB MLOps platform!

Appendix

See appendix.pdf for the appendix file for the main manuscript.

Training & Evaluation

1. How to Training & Evaluation

1. pendulum datset

training

• training CDG-VAE

python main.py --model "CDGVAE"

• training CDG-VAE in order to evaluate distributional robustness of downstream task

python DR/main.py --model "CDGVAE"

(Note) The file names with _semi means training CDG-VAE under semi-supervised learning setting.

evaluation

• counterfactual image generation: inference.py
• sample efficiency of downstream task: sample_efficiency.py
• distributional robustness of downstream task: DR/robustness.py

2. tabular dataset

training

• CDG-VAE

python tabular/main.py --model "CDGVAE"

• CDG-TVAE

python tabular/main_tvae.py 

evaluation (SHD and synthetic data quality)

• CDG-VAE: tabular/inference.py
• CDG-TVAE: tabular/inference_tvae.py

Results

directory and codes

.
+-- assets 
+-- modules 
|       +-- datasets.py
|       +-- model.py
|       +-- pendulum_real.py
|       +-- pendulum.py
|       +-- simulation.py
|       +-- train.py
|       +-- viz.py

+-- main.py
+-- main_semi.py
+-- main_classifier.py
+-- inference.py
+-- metric.py
+-- sample_efficiency.py
+-- LICENSE
+-- README.md
+-- appendix.pdf

+-- DR (folder which contains source codes for distributional robustness experiments)
|   +-- assets 
|   +-- main.py
|   +-- main_semi.py
|   +-- robustness.py
|   +-- toy_DR.py

+-- tabular (folder which contains source codes for tabular dataset experiments)
|   +-- assets 
|   +-- modules
|       +-- adult_datasets.py
|       +-- covtype_datasets.py
|       +-- loan_datasets.py
|       +-- errors.py
|       +-- numerical.py
|       +-- transformer_base.py
|       +-- transformer_null.py
|       +-- data_transformer.py
|       +-- model.py
|       +-- train.py
|       +-- evaluation.py
|       +-- viz.py
|       +-- simulation.py
|   +-- dag_adult.py
|   +-- dag_covtype.py
|   +-- dag_loan.py
|   +-- main.py
|   +-- inference.py
|   +-- main_tvae.py
|   +-- inference_tvae.py

Citation

@inproceedings{An2023CDGVAE,
  author       = {SeungHwan An and
                  Kyungwoo Song and
                  Jong{-}June Jeon},
  title        = {Causally Disentangled Generative Variational AutoEncoder},
  booktitle    = {{ECAI} 2023 - 26th European Conference on Artificial Intelligence,
                  September 30 - October 4, 2023, Krak{\'{o}}w, Poland - Including
                  12th Conference on Prestigious Applications of Intelligent Systems
                  {(PAIS} 2023)},
  series       = {Frontiers in Artificial Intelligence and Applications},
  volume       = {372},
  pages        = {93--100},
  publisher    = {{IOS} Press},
  year         = {2023},
  url          = {https://doi.org/10.3233/FAIA230258},
  doi          = {10.3233/FAIA230258},
}