Point Set Diffusion

This is the reference implementation of our ICLR 2025 paper Unlocking Point Processes through Point Set Diffusion.

Citation

If you build upon this work, please cite our paper as follows:

@inproceedings{luedke2025psdiff,
    title={Unlocking Point Processes through Point Set Diffusion},
    author={David L{\"u}dke and Enric Rabasseda Ravent{\'o}s and Marcel Kollovieh and Stephan G{\"u}nnemann},
    booktitle={The Thirteenth International Conference on Learning Representations},
    year={2025},
    url={https://openreview.net/pdf?id=4anfpHj0wf}
}

Getting started

Basic requirements

The code has been tested on a cluster of Linux nodes using SLURM.
We cannot guarantee the functioning of the code if the following requirements are not met:

Installation

To properly install and run our code we recommend using a virtual environment (e.g., created via pyenv-virtualenv or conda).

The entire installation process consists of 3 steps. You can skip step 0 at you own "risk".

(Optional) Step 0: create a virtual environment

In the following we show how to create the environment via pyenv and pyenv-virtualenv. The steps are the following:

• install pyenv (if you don't have it yet) by following the original guidelines;
• install the correct Python version:

  pyenv install 3.10.4

• create a virtual environment with the correct version of Python:

  pyenv virtualenv 3.10.4 ps-diff

Step 1: clone the repository, change into it and (optional) activate the environment

This step allows you to download the code in your machine, move into the correct directory and (optional) activate the correct environment. The steps are the following:

• clone the repository:

  git clone https://github.com/davecasp/ps-diff.git

• change into the repository:

  cd point-set-diff

• (optional) activate the environment

  pyenv activate ps-diff

Step 2: install the code as a local package

All the required packages are defined in the pyproject.toml file and can be easily installed via pip as following:

pip install -r requirements.txt

Run code

Configuring experiments and running the code is done via hydra. If you are unfamiliar with how hydra works please check out the documentation.

Train model on paper configs

To run our Model for a datatype:

./train.py -m --config-name config_name

where config_name should be spp_train or stpp_train or tpp_train. All seeds and datasets are scheduled as a gridsearch via the multirun flag.

Condional modelling

Example will be added in due time