Installation

To install franken, start by setting up your environment with the correct version of PyTorch. This is especially necessary if you wish to use GPUs. Then install franken by running

pip install franken

The basic installation comes bare-bones without any GNN backbone installed. You can install franken with a specific backbone directly, by running one of the following commands

pip install franken[cuda,mace]
pip install franken[cuda,sevenn]
pip install franken[cuda,pet]

In more detail:

  • the cuda qualifier installs dependencies which are only relevant on GPU-enabled environments and can be omitted.

  • the three supported backbones are MACE, UPET, and SevenNet. They are explained in more detail below.

Warning

Each backbone seems to have mutually incompatible requirements, particularly with regards to e3nn - but also pytorch versions might be a problem. To minimize incompatibilities, we suggest that the users who wishes to use multiple backbones create independent python environments for each. In particular, the mace-torch package requires an old version of e3nn (0.4.4) which may conflict with other backbones. If you encounter errors with model loading, simply upgrade e3nn by running pip install -U e3nn.

Supported pre-trained models

MACE

We support several models which use the MACE architecture:

  • The MACE-MP0 models trained on the materials project data by Batatia et al. Additional informations on the pre-training of MACE-MP0 are available on its HuggingFace model card.

  • The MACE-OFF (paper and github) models which are pretrained on organic molecules.

  • The Egret (github) family of models (Egret-1, Egret-1e, Egret-1t), also tuned for organic molecules.

To use any MACE model as a backbone for franken just pip-install mace-torch in franken’s environment

pip install mace-torch

or directly install franken with mace support (pip install franken[cuda,mace]).

In addition to MACE-MP0 trained on the materials project dataset, Franken also supports the MACE-OFF models for organic chemistry.

PET

Franken supports UPET models through the Metatomic/Metatrain ecosystem. To use PET models as a backbone for franken, install the required dependencies with pip install franken[cuda,pet] or follow the instructions on metatomic and metatrain documentation.

SevenNet

Franken also supports the SevenNet model by Park et al. as implemented in the sevennet library. We have only tested the SevenNet-0 model trained on the materials project dataset, but support for other models should be possible (open an issue if you encounter any problem).