/linear-causal-disentanglement-via-interventions

Primary LanguagePython

Linear Causal Disentanglement via Interventions

Setup

To set up a virtual environment with the required packages, run:

bash setup.sh

Note that this project depends on two packages which can be difficult to install, pygraphviz and gurobipy. The dependence on pygraphviz is only necessary for plotting the learned latent DAG over the real data, so almost everything can be run without this dependency. The dependence on gurobipy is required for efficient computation to see how well two permutations match, up to the partial order defined by a DAG. This dependence can be removed by changing DEFAULT_PERMUTATION_METHOD in src/run_experiment.py from "ilp" to "naive".

Synthetic Data Result Reproduction

To reproduce the synthetic data results, run:

python3 -m experiments.experiment1.noisy_recovery --seed 8164 --nnodes 5 --nnodes_obs 10

Real Data Result Reproduction

To reproduce the real data results, first download the data:

python3 -m experiments.real_data.step1_download_and_pickle

For the semi-synthetic results, run

bash experiments/real_data/semisynthetic.sh

For real-data results, run

bash experiments/real_data/real.sh