ARES with concurrency

This repository investigates faster CPU inference on the ARES neural network.

Installation

You need to install e3nn_ares before setting up ares_release.

Please follow their respective instructions:

• e3nn_ares/README.md
• ares_release/README.md

If you opt for a venv, you might find useful the requirements.txt and requirements-cpu.txt files at the root of the repository.

Usage

The same instructions for inferencing the network still apply, except that you would now pass two additional options:

• --torch_threads to set the number of PyTorch threads for intra-op parallelization
• --async_threads to set the number of concurrent threads to use in thread pools during forward pass

If the latter is not passed or set to a value less than 2, the original version of ARES is run.