Perform experiments with modelforge
Each experiment is performed and saved in the experiments directory. Each experiment has a descriptive name and contains config files inside the configs folder to reproduce the experiment. For each experiment, we provide a descriptive README file, and a bash script reproducing the experiment on a single node and GPU (this might differ from how we run the experiment in practice).
As a result, we provide test/validation set RMSE, training curve (MSE), training rate scheduling, and the number of epochs.
Note: we always train with the main branch of the modelforge repository. To make sure that we are able to reproduce the training code we will add the commit hash of the current main branch to each training.
The folder structure looks like the following:
experiments
- name_of_the_experiment
- config # define the training, deposit all toml files for dataset/potential that differ from the default values
- training.toml
- README # describe why this experiment is performed, and what is performed
- train.sh # reproduces the experiment on a single node/GPU