📘Documentation | 🛠️Installation | 🚀Model Zoo | 🤗Huggingface | 👀Visualization | 🆕News
NOTE: We use the concepts of VQ-VAE and memory to improve Earthfarseer and add more experiments. Currently, it is submitted to TPAMI 2024 as Earthfarseer-V2.
Efficiently modeling spatio-temporal physical processes presents a challenge for deep learning. Recent models often lack simplicity and practicality. To address this, we propose EarthFarseer, a concise framework combining parallel local convolutions and global Fourier-based transformer architectures. This approach dynamically captures local-global spatial interactions and dependencies. EarthFarseer also incorporates multi-scale fully convolutional and Fourier architectures for efficient temporal evolution. It demonstrates strong adaptability, fast convergence, and improved local fidelity in long-term predictions across various datasets, achieving state-of-the-art performance.
- Install Python 3.8. For convenience, execute the following command.
pip install -r requirements.txt- Prepare Data. You can obtain experimental datasets from the following links.
| Dataset | Task | Geometry | Link |
|---|---|---|---|
| Navier-Stokes equation | Predict future fluid vorticity | Regular Grid | [Google Cloud] |
| Shallow-water equations | Predict future fluid height | Regular Grid | [Google Cloud] |
| Moving MNIST | Predict future image | Regular Grid | [Google Cloud] |
- Use the following instructions to quickly run the code.
python train_main.py --data_path Dataset/NavierStokes_V1e-5_N1200_T20.mat --num_epochs 100 --batch_size 5If you are interested in our repository or our paper, please cite the following paper:
@inproceedings{wu2024earthfarsser,
title={Earthfarsser: Versatile Spatio-Temporal Dynamical Systems Modeling in One Model},
author={Wu, Hao and Liang, Yuxuan and Xiong, Wei and Zhou, Zhengyang and Huang, Wei and Wang, Shilong and Wang, Kun},
booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
volume={38},
number={14},
pages={15906--15914},
year={2024}
}
If you have any questions or want to use the code, please contact [email protected].