Recipes

One of mlipx core functionality is providing you with pre-designed recipes. These define workflows for evaluating MLIP on specific tasks. You can get an overview of all available recipes using

(.venv) $ mlipx recipes --help

All recipes follow the same structure. It is recommended, to create a new directory for each recipe.

(.venv) $ mkdir molecular_dynamics
(.venv) $ cd molecular_dynamics
(.venv) $ mlipx recipes md --initialize

This will create the following structure:

molecular_dynamics/
├── .git/
├── .dvc/
├── models.py
└── main.py

After initialization, adapt the main.py file to point towards the requested data files. Define all models for testing in the models.py file.

Finally, build the recipe using

(.venv) $ python main.py
(.venv) $ dvc repro

Upload Results

Once the recipe is finished, you can persist the results and upload them to a remote storage. Therefore, you want to make a GIT commit and push it to your repository.

(.venv) $ git add .
(.venv) $ git commit -m "Finished molecular dynamics test"
(.venv) $ git push
(.venv) $ dvc push

Note

You need to define a GIT and DVC remote to push the results. More information on how to setup a DVC remote can be found at https://dvc.org/doc/user-guide/data-management/remote-storage.

In combination or as an alternative, you can upload the results to a parameter and metric tracking service, such as mlflow. Given a running mlflow server, you can use the following command to upload the results:

(.venv) $ zntrack mlflow-sync --help

Note

Depending on the installed packages, the mlflow command might not be available. This functionality is provided by the zntrack package, and other tracking services can be used as well. They will show up once the respective package is installed. See https://zntrack.readthedocs.io/ for more information.