GPyTorchWrapper

GPyTorchWrapper is a lightweight Python package designed to streamline the training of Gaussian Process (GP) models using GPyTorch. While it was developed for modeling potential energy surfaces (PES) in small molecular systems using custom permutationally invariant kernels (based on Bartók & Csányi [Bartok2015]), it is general enough for other regression tasks.

Features

• Modular YAML-based configuration

• Fully differentiable, permutationally invariant kernel functions

• Script for exporting models as TorchScript for deployment

Installation

Set up the environment locally:

conda env create -f environment.yml

conda activate gpytorchwrapper

pip install -e . --use-pep517

For an Intel-optimized HPC environment, use environment_hpc.yml.

HPC Support

A SLURM-based submission script is available in bash/sub_gp_training.sh. Use gp_training.sh to configure and launch training jobs with CLI options.

Docker Example

Run the example in Docker:

docker build -t gpytorchwrapper . && ./run-example-in-docker.sh

After training, 3d_plot.png will appear in the working directory.

Documentation

• Training
• Deployment
• Custom Models
• Permutationally Invariant Kernels
• API

Reference

[Bartok2015]

Bartók, A. P.; Csányi, G. Gaussian Approximation Potentials: A Brief Tutorial Introduction. Int. J. Quantum Chem. 2015, 115 (16), 1051–1057. https://doi.org/10.1002/qua.24927