📜 We term our method Aether, inspired by the postulated medium that permeates all throughout space and allows for the propagation of light. 💨 🌊 🪨 🔥
Official source code for
Latent Field Discovery in Interacting Dynamical Systems with Neural Fields Miltiadis Kofinas, Erik J Bekkers, Naveen Shankar Nagaraja, Efstratios Gavves NeurIPS 2023 https://arxiv.org/abs/2310.20679
TL;DR: We discover global fields in interacting systems, inferring them from the dynamics alone, using neural fields.
For a reference implementation of Aether, see here.
Create a new conda environment and install dependencies:
conda create -n aether python=3.9 conda activate aether conda install pytorch==2.0.1 torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia conda install pyg pytorch-scatter -c pyg pip install plotly tensorboard matplotlib pandas Then, download the repo and install it:
git clone https://https://github.com/mkofinas/aether.git cd aether pip install -e . To run a specific experiment, please follow the README file within its corresponding experiment directory. It provides full instructions and details for downloading/generating the data and reproducing the results reported in the paper.
- Electrostatic field:
experiments/electrostatic - Lorentz force field:
experiments/lorentz - Traffic scenes (inD):
experiments/ind - Gravitational field:
experiments/gravitational
The scripts directory contains scripts for running experiments for the electrostatic field setting, the traffic scenes, and the gravitational field. From the root directory, you can run the following:
./scripts/electrostatic_field_aether.sh./scripts/ind_aether.sh./scripts/gravitational_field_3d_aether.sh
Each script will train a model, evaluate it, save results, and (in the case of Aether) visualize the discovered field(s). You can also find scripts for other baseline models in the scripts directory.
Our codebase is based on the code from the papers:
- LoCS, https://github.com/mkofinas/locs
- dNRI, https://github.com/cgraber/cvpr_dNRI
- NRI, https://github.com/ethanfetaya/NRI
- ClofNet, https://github.com/mouthful/ClofNet
- EGNN, https://github.com/vgsatorras/egnn
If you find our work or this code to be useful in your own research, please consider citing the following paper:
@inproceedings{kofinas2023latent, title={{L}atent {F}ield {D}iscovery in {I}nteracting {D}ynamical {S}ystems with {N}eural {F}ields}, author={Kofinas, Miltiadis and Bekkers, Erik J, and Nagaraja, Naveen Shankar and Gavves, Efstratios}, booktitle = {Advances in Neural Information Processing Systems 36 (NeurIPS)}, year={2023}, }