ParamRepulsor

This is the code repository for the NeurIPS 2024 paper "Navigating the Effect of Parametrization for Dimensionality Reduction". Our paper can be found here.

How to install

This repository can be installed locally via pip by the following command:

git clone https://github.com/hyhuang00/ParamRepulsor.git
cd ParamRepulsor
pip install .

Note: this will not install torch, as this is highly platform-dependent. This project provides optionals:

pip install .[cpu]    # cpu-only pytorch
pip install .[cu118]  # cuda 118
pip install .[cu121]  # cuda 121
pip install .[cu124]  # cuda 124
pip install .[mps]    # arm64/aarch64 (Apple M-Series chips)

This project also supports uv (pip install uv):

echo "3.11" > .python-version  # supported: [3.9, 3.12)
uv sync (--extra cpu)  # as appropriate for your system
uv run pytest
TORCH_DEVICE=cpu uv run pytest  # disable accelerator

How to use our algorithm

ParamPaCMAP/ParamRepulsor is fully scikit-learn compatible, meaning that it can be used as any other scikit-learn based algorithm. After the installation, you can use our algorithm by:

import parampacmap

# Initialize the reducer. Notice that by default, the stronger paramrepulsor
# algorithm will be used.
reducer = parampacmap.ParamPaCMAP()
X_low = reducer.fit_transform(X)  # Substitute your data here.

Citation

If you have referred to our research in your publication, or you used the ParamRepulsor/ParamPaCMAP algorithm in this repository, please cite our paper using the following bibtex:

@inproceedings{huang2024navigating,
  title={Navigating the Effect of Parametrization for Dimensionality Reduction},
  author={Huang, Haiyang and Wang, Yingfan and Rudin, Cynthia},
  booktitle={The Thirty-eighth Annual Conference on Neural Information Processing Systems},
  year={2024},
}

Project Contributor

A full list of project contributors can be found here.