Data-based parallel acceleration for physics-informed neural networks (PINNs) via Horovod. Its code comes with the preprint entitled:

$h$-analysis and data-parallel physics-informed neural networks, P. Escapil-Inchauspé and G. A. Ruz.

We apply data-based Horovod acceleration to pioneer PINNs code by Raissi, Perdikaris and Karniadakis. Horovod acceleration is inspired by Xihui Meng Distributed-training-Horovod code.

This repository is intended to help users understand how to apply Horovod based data-parallel acceleration to PINNs or physics-informed machine learning schemes. It is also meant to enable users to fully replicate and reproduce the results in our manuscript, including the figures generation.

Backend: tensorflow.compat.v1.

This repository contains the following folders:

• PINNs submodule;
• Distributed-training-Horovod submodule;
• Laplace: An improved version of pinn_hvd_data.py;
• Schrodinger: Data-parallel Schrodinger.py forward problem;
• NavierStokes: Data-parallel NavierStokes.py inverse problem.

To clone this repository along with its submodules, run:

git clone https://github.com/pescap/horovodPINNs.git --recurse-submodules 

To run the experiments, a Docker image with horovod==0.26.1 can be downloaded and run throughout command:

nvidia-docker run -p 8888:8888 pescap/dist-training-horovod-master