Neural Causal Abstractions

This repository contains the code for the paper "Neural Causal Abstractions" by Kevin Xia and Elias Bareinboim.

Please cite our work if you found this code useful.

Setup

Run the following code to install python requirements.

python -m pip install -r requirements.txt

To run the ColoredMNIST experiments, place the MNIST data files in dat/mnist.

Running the code

All experiment procedures can be run using the main.py file with the desired arguments entered. For any of the commands below, feel free to modify the hyperparameters from main.py.

BMI Experiment

To run the identification part of the BMI experiment, navigate to the base directory of the repository and run

python -m src.main <NAME> identify bmi gan --u-size 16 --batch-norm --gan-mode wgan --custom-query -r 2 --gpu 0

where <NAME> is replaced with the name of the folder in which the results will be saved.

For estimation, run

python -m src.main <NAME> sampling bmi gan --u-size 16 --batch-norm --gan-mode wgan --custom-query --gpu 0

To turn off normalization of the data, use the --no-normalize tag, and to apply abstractions, use the --use-tau tag.

To compile the visualization of the ID results, name the folders of the results naive, naive_nonormal and tau for each setting respectively, and place all three folders into a single folder. Then run the command

python -m src.experiment.experiment1_id_results <DIR>

where <DIR> refers to the outer folder with all three results. Similarly, for estimation, name the three folders identically and place them in a different directory. Then run

python -m src.experiment.experiment1_est_results <DIR>

ColoredMNIST Experiment

To run the GAN-NCM without representation learning, run the following command

python -m src.main <NAME> sampling mnist gan --h-layers 3 --h-size 2 --scale-h-size --scale-u-size --batch-norm --gan-mode wgan --gan-arch biggan --disc-type biggan --img-size 32 --gpu 0

The command can be run with sampling_noncausal in place of sampling to run a standard conditional causal-agnostic GAN.

To run the GAN-RNCM, run the following command

python -m src.main <NAME> sampling mnist gan --h-layers 3 --h-size 2 --scale-h-size --scale-u-size --batch-norm --gan-mode wgan --gan-arch biggan --repr auto_enc_conditional --rep-size 64 --rep-image-only --rep-h-layers 3 --rep-h-size 128 --img-size 32 --gpu 0

To compile the visualization of the results, name the folders of the results naive, noncausal, and representational for each setting respectively, and place all three folders into a single folder. Then run the command

python -m src.experiment.experiment2_results <DIR>

where <DIR> refers to the outer folder with all three results.