ild-domain-counterfactuals

Welcome! This is the official repository for the following paper:

Kulinski, S., Zhou, Z., Bai, R., Kocaoglu, M., & Inouye, D. I. Towards Characterizing Domain Counterfactuals For Invertible Latent Causal Models. ICLR 2024.

This repository will be updated soon! Feel free to email the authors for code.

General Guidance

All experiments are conducted using wandb sweep. TODO: Add some instruction on how to run the experiments.

Simulated Experiments

All corresponding code could be found in the simulated folder.

TODO: Add some instruction on how to run a single experiment.

To regenerate the figures in the paper:

Step 1

Run all sweeps in the directory simulated/configs.

Step 2

Follow the instruction in the notebook simulated/demo_results.ipynb to regenerate the figures.

Additional Experiments

In the rebuttal, we add some additional experiments using Normalizing Flows and VAEs as G. The corresponding code could be found in simulated/flow and simulated/vae respectively. Similarly, to regenerate the figures in the paper: (1) run the sweeps in the corresponding configs directory, and (2) follow the instruction in the notebook (share the same notebook with other simulated experiments).

TODO: Clean up the code for these experiments.

Image Experiments

Validation

Step 1

Go to directory images/validation, run

python run_prep_classifier.py

TODO: Fix path? Arguments?

Step 2

Run

python run_test.py

Notes