Skip to content

Latest commit

 

History

History
230 lines (159 loc) · 7.17 KB

README.md

File metadata and controls

230 lines (159 loc) · 7.17 KB

Boosted Control Functions:
Distribution generalization and invariance in confounded models

The goal of this repository is to reproduce the numerical experiments from [1].

If you use this software in your work, please cite it using the following metadata.

@misc{gnecco2024boostedcontrolfunctionsdistribution,
      title={Boosted Control Functions: Distribution generalization and invariance in confounded models}, 
      author={Nicola Gnecco and Jonas Peters and Sebastian Engelke and Niklas Pfister},
      year={2024},
      eprint={2310.05805},
      archivePrefix={arXiv},
      primaryClass={stat.ML},
      url={https://arxiv.org/abs/2310.05805}, 
}

Table of Contents

Directory Structure

.
├── R-code     
│   ├── R               # R functions
│   └── main            # R scripts
|
├── data
│   ├── processed       # Processed data
│   └── raw             
│       └── prism_temp  # Temperature data 
|
├── python
│   ├── main            # Python scripts
│   └── src             
│       ├── bcf         # Classes and functions for BCF
│       ├── scenarios   # Functions to generate data
│       ├── simulations # Functions to run simulations
│       └── utils       # Helper functions
|
└── results
    ├── figures         # Saved figures
    └── output_data     # Saved results

Dependencies

  • Python: version (3.9.1)
  • R: version (4.0.2)

Installing Python Requirements

  1. Ensure Python and pip are installed: Before installing the requirements, make sure you have Python installed on your system. You can check by running in the terminal:
python --version

Also, ensure you have pip (Python package manager) installed by running in the terminal:

pip --version

If you don't have Python or pip installed, please download Python and install it. Pip is included in Python versions 3.4+.

  1. Navigate to python directory: Use the terminal to navigate to the directory containing your requirements.txt file.
cd python
  1. Install the requirements: Once inside the directory, run the following command to install the required packages:
pip install -r requirements.txt

pip install -e .

The first line will install all the Python packages listed in requirements.txt. The second line will install the python code located in ./python/src/ as a package.

⚠️ Consider creating a new Python environment before installing the Python packages.

Installing R Requirements

  1. Ensure R is installed: Make sure you have R installed on your system by running in the terminal:
R --version

If you don't have R installed please download R and install it.

  1. Install the requirements: Move to the R-code directory and launch the R script dependencies.R by running in the terminal:
cd R-code
Rscript --vanilla main/dependencies.R

Instructions

⚠️ Experiment 1 and 2 were originally run on a cluster with a PBS/TORQUE job schedular. The instructions below only apply to this type of scheduler.

Experiment 1: Predicting unseen interventions

⚠️ This experiment was run in parallel using 80 cores and took around 2 hours.

1a) How to run on a PBS/TORQUE cluster:

  • Open ./python/main/run_job.sh and modify lines 29–31 depending on your cluster.
  • From the root of the project, in the terminal type the following:
cd python

qsub main/run_job.sh "python main/experiment_1.py"

1b) How to run locally

  • Open ./python/main/experiment_1.py and on line 13 set the number of cores according to your machine, e.g., N_WORKERS = 4.

  • From the root of the project, in the terminal type the following:

cd python

python main/experiment_1.py

💡 To run only a few iterations of the experiment, change the number of repetitions on line 29 of ./python/main/experiment_1.py, e.g., "n_reps": range(2).

2) Plot results

  • From the root of the project, in the terminal type the following:
cd R-code

Rscript --vanilla main/plot-interv_sim.R

Experiment 2: Estimating M0

⚠️ This experiment was run in parallel using 10 cores and took around 20 minutes.

1a) How to run on a PBS/TORQUE cluster:

  • Open ./python/main/run_job.sh and modify lines 29–31 depending on your cluster.
  • From the root of the project, in the terminal type the following:
cd python

qsub main/run_job.sh "python main/experiment_2.py"

1b) How to run locally

  • Open ./python/main/experiment_2.py and on line 13 set the number of cores according to your machine, e.g., N_WORKERS = 4.

  • From the root of the project, in the terminal type the following:

cd python

python main/experiment_2.py

💡 To run only a few iterations of the experiment, change the number of repetitions on line 31 of ./python/main/experiment_2.py, e.g., "n_reps": range(2).

2) Plot results

  • From the root of the project, in the terminal type the following:
cd R-code

Rscript --vanilla main/plot-nullspace_sim.R

California housing dataset

To run the experiment and produce the figures for the California housing dataset, follow these steps.

cd python

python main/housing_data-import.py

python main/asc2csv.py ../data/raw/prism_temp/ "../data/processed/temp-data-california.csv" "Mean_Temp"

cd ../R-code

Rscript --vanilla main/prepare-temp_housing_data.R

cd ../python

python main/housing_data_analysis_regions.py

cd ../R-code

Rscript --vanilla main/plot-temp_housing_data.R

Contributors

License

This project is licensed under the MIT License - see the LICENSE file for details.

References

[1]: Nicola Gnecco, Jonas Peters, Sebastian Engelke, and Niklas Pfister. 2024. "Boosted Control Functions: Distribution generalization and invariance in confounded models." arXiv Preprint [https://arxiv.org/abs/2310.05805].