Implementation of the dynamic training method applied to an equivariant graph neural network
Correspondance to: [email protected]
Clone this repository
git clone https://github.com/IZugec/DTEGNN.git
You can then build the environment with Anaconda:
cd DTEGNN
conda env create --file envs/env-cu121.yml --force
This command builds an environment with CUDA 12.1 version.
To train a DT-EGNN model you can run
python3.12 <path_to_cloned_dir>/scripts/train.py <path_to_input_file>/config.yml
An example of an input configuration file can be found in dtegnn/config/example.yml.
Current implementation reads AIMD simultions with the help of ASE and assumes Trajectory object. Velocities are assumed to be in the basis of lattice vectors multiplied by the integration time of the underlying simulation. An example of data required for a successful run can be downloaded through the following link
or directly by running
wget https://figshare.com/ndownloader/files/52738940
followed by renaming the downloaded file, and unzipping it.
mv 52738940 dataset.zip
unzip -d dataset.zip <path_to_desired_directory>
An example showing how to deploy the model trained via DT is in our examples directory.