Frugal Flows

This repository is the official implementation of Marginal Causal Flows for Inference and Validation, accepted at NeurIPS 2024.

Set-up the environment

Micromamba

Environment requirements to run the paper experiments can be found in the environment.yaml file. This file can be used to set up an environment with any environment manager e.g., venv, Conda, Mamba, Micromamba. With Micromamba, you can create and activate the environment as follows:

micromamba create -f environment.yaml

micromamba activate <name-environment>

This will automatically install the Frugal Flows package together with its dependencies and all the Python packages required to run the experiments in the paper.

Some experiments require additional R packages. To install these, please refer to the section "Rerunning Experiments".

Rerunning Experiments

Rerunning all the experiments presented in the paper requires the installation of rpy2 and other relevant R libraries. To install these, please run:

python install_rpy2_libraries.py

Pip install Frugal Flows

Alternatively, you can solely install the Frugal Flow package using pip:

git clone <URL-repository>

cd frugal-flows

pip install -e ./
 

The dependencies of frugal-flows can be found in the pyproject.toml file.

General Structure

• The main bulk of the frugal flow implementation can be found in frugal_flows.
• The script containing functions to generate the simulated data for the inference experiments can be found here.
• The main class which allows the user to implement Frugal Flows at ease can be found in benchmarking.py

Reproduce paper experiments

• To reproduce Table 1: Continous_Frugal_Flows.ipynb
• To reproduce Figure 3: Lalonde_Data_Pipeline.ipynb
• To reproduce Figure 4: e401k_Data_Pipeline.ipynb
• To reproduce Table 3 in the Appendix: Logistic_Sampling.ipynb

To reproduce comparisons to Causal Flows

To recover Causal Flows ATE values reported in Table 1:

• Clone our forked causal-flow repository
• Build your environment from the environment.yaml file
• Run run.sh to reproduce experiments
• Run ate_FF_loop.ipynb to produce ATE values

Acknowledgement

This repository is developed mainly based on the FlowJAX repository. Many thanks to its contributors!