Learning Mesh-Based Simulation with Graph Networks

Setup

conda env create -f utils/envs/gnn.yml
conda activate gnn

Download the dataset

bash download_dataset.sh cylinder_flow data/

It will create a folder cylinder_flow/ inside the folder data/ and will take some time to download the simulations. You will obtain four files:

• meta.json
• test.tfrecord
• train.tfrecord
• valid.tfrecord

Rename the folder cylinder_flow/ to raw/.

Parameters

Open main.py to update the paths.

The dataset is composed of 1000 simulations, each of them having 600 time steps. Pre-processing and training on all the dataset is not possible on standards GPU in a short time. To select only a part of the dataset, have a look at the data.* parameters. These are already set up, but you can change to reduce the number of data and have a faster training.

Train the model

Train the model by running:

python main.py fit

You can get help on the command line arguments by running:

python main.py fit --help

It will create a new folder in the logs/ folder containing the checkpoints of the model and a configuration file containing the parameters used for the training, that you can use later if you want.

Evaluate the model

To evaluate the model training, run:

tensorboard --logdir=logs/

You can stop the training whenever you are satisfied with the learning. The model is saved in logs\version_*\checkpoints\.

Test the model

To test the model, run:

python main.py test --ckpt_path $ckpt_path

where $ckpt_path is the path to the checkpoint file located in the logs/version_$version/checkpoints/ folder.

Contact

Ugo Pelissier
<ugo.pelissier@minesparis.psl.eu>