This repository contains the official implementation of LapSum, a novel approach for differentiable ranking, sorting, and top-k selection, as presented in our ICML 2025 paper:
"LapSum - One Method to Differentiate Them All: Ranking, Sorting and Top-k Selection"
arXiv preprint | ICML 2025 proceedings
.
├── experiments_of_paper/ # Reproduction scripts for paper experiments
│ └── README.md # Instructions
├── lapsum/ # Core library implementation
│ └── README.md # Installation and usage instructions
├── tutorial_notebooks/ # Pure PyTorch implementations with usage examples
└── README.md # are you here
The core implementation is available as the lapsum Python package. For detailed installation instructions, see ./lapsum/README.md
Tested Environment:
- OS: Ubuntu 24.04 LTS
- Python: 3.10+
- PyTorch: 2.3+
- CUDA: 12.3 (when using GPU acceleration)
For practical examples and educational purposes, see our Jupyter ./tutorial_notebooks/
🚀 Pro Tip: If you don't want to compile the CPP/CUDA implementation, you can use the pure PyTorch version of the library - it's almost as fast and memory-efficient as the CUDA implementation!
For users less experienced with building libraries from source, we especially recommend using the this implementation - it works out-of-the-box with standard PyTorch installations and doesn't require any additional compilation steps!
All experimental results from the paper can be reproduced using scripts in ./experiments_of_paper/
This project is open-source under the MIT License. For full details, see the LICENSE file.
If you use LapSum in your research, please cite our work:
@inproceedings{lapsum2025,
title={LapSum - One Method to Differentiate Them All: Ranking, Sorting and Top-k Selection},
author={\L{}ukasz Struski and Micha\l{} B. Bednarczyk and Igor T. Podolak and Jacek Tabor},
booktitle={The International Conference on Machine Learning (ICML) 2025},
year={2025}
}