Installation

causalig needs Python ≥ 3.10. Its only hard dependencies are PyTorch ≥ 2.0 and NumPy.

From a clone (recommended for now)

git clone https://github.com/sifrimlab/CausalIG.git
cd CausalIG
pip install -e .

Or install straight from GitHub without cloning:

pip install "causalig @ git+https://github.com/sifrimlab/CausalIG.git"

Tip

On Linux, pip install torch pulls the CUDA build by default, which is several GB. On a CPU-only machine, install the CPU wheel before installing causalig:

pip install torch --index-url https://download.pytorch.org/whl/cpu

With conda

The repository ships an environment.yml that creates a causalig environment:

conda env create -f environment.yml
conda activate causalig
pip install -e ".[shap,experiments]"   # adds the extras needed for the paper's experiments

Optional extras

Extra	Installs	Needed for
shap	shap	The Shapley-value baselines CausalSHAP and ConditionalSHAP.
experiments	captum, sempler, matplotlib, scipy	Running experiments/ and scripts/, which reproduce the paper’s synthetic benchmark (see Reproducing the paper).
dev	pytest, pytest-cov, ruff, black, shap	Running the test suite and linters (see Development).

pip install -e ".[shap]"
pip install -e ".[experiments]"
pip install -e ".[dev]"

import causalig never imports shap: the SHAP baselines are loaded lazily the first time you access causalig.attr.CausalSHAP or causalig.attr.ConditionalSHAP.

Check the install

python -c "import causalig; print(causalig.__version__)"