This project provides a clean pytorch implementation of Learning mesh-based simulation with Graph Networks.
- Using pytorch-geometric data structure for graph representation and processing.
- Using hydra for hierarchical configuration and using Weigts&Biases to track and visualize experiments.
Prediction: right
Table below shows quantitative results of different rollout steps for both dataset cylinder_flow and flag_simple. Our results differ from the ones published in Paper. Results can vary under different hyperparameter settings, e.g. random seed and learning rate.
| rollout in scale(1e-3) |
1 | 10 | 50 | 100 | 200 | all |
|---|---|---|---|---|---|---|
| cylinder_flow | 3.99 | 8.90 | 18.10 | 25.36 | 34.97 | 63.56 |
| flag_simple | 0.98 | 12.15 | 133.51 | 157.07 | 157.42 | 165.09 |
More simply you can install this package via packaging tool pip.
# download this package
git clone ...
# change directory
cd mgn
# install mgn package
pip install -e .
You can train models with
python example/train.py model=cloth datamodule=flag_simple
or
python example/train.py model=cfd datamodule=cylinder_flow
For evaluating checkpoints see eval.ipynb.