Analytics for Neural Networks (A4NN)

About

A4NN is a workflow designed to accelerate Neural Architecture Search (NAS) by integrating analytical engines that enhance search efficiency and reduce training time. A4NN leverages Wilkins to orchestrate the workflow by launching tasks and establishing communicators between them. The workflow consists of a multi-stage process involving a NAS algorithm, a parametric prediction engine, and a structural similarity engine--each implemented as distinct but interoperable tasks. In this implementation, we use NSGANet, a multi-objective search algorithm that balances accuracy with computational cost.

This repository includes example task definitions, configuration files, and scripts for running the A4NN workflow. A diagram of the workflow is provided below for reference.

Installation

Install Spack

1. Clone the Spack repository and checkout version v0.23.0:

   git clone https://github.com/spack/spack.git 
   cd spack/
   git checkout tags/v0.23.0

   Activate Spack:

   . path/to/spack/share/spack/setup-env.sh
   

2. Install [email protected] with Spack and add it to the list of visible compilers:

   spack install [email protected]
   spack compiler add <path-to-gcc>

   You can find the path to [email protected] using:

   spack location -i [email protected]

   It is recommended to make [email protected] the only visible compiler. To see a list of visible compilers:

   spack compilers

   To remove other compilers from this list:

   spack compiler remove <name-of-compiler>

3. Install Wilkins using Spack. Clone the Wilkins and LowFive repositories and add them to your local Spack repositories:

   cd /path/to/wilkins/
   git clone https://github.com/orcunyildiz/wilkins.git
   spack repo add /path/to/wilkins/
   
   cd /path/to/lowfive/
   git clone https://github.com/diatomic/LowFive.git
   spack repo add /path/to/lowfive/

   You can confirm that Spack can find Wilkins and LowFive:

   spack info wilkins
   spack info lowfive

   Then install Wilkins

   spack install wilkins

4. Use the create_env.sh script in the repository to create a spack environment to run A4NN using Wilkins. Feel free to modify the env.yaml to add additional package specs or indicate other package versions.

   # Clone repo and create spack environment
   git clone [email protected]:TauferLab/A4NN_workflow.git
   cd A4NN_workflow/
   source ./create_env.sh

   PyTorch and several additional libraries must be installed using pip. After activating the environment, install these dependencies as follows:

   spack env activate <name-of-the-environment>
   spack load py-pip
   
   # Install PyTorch for CPU
   python -m pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
   # Or for GPU support (CUDA 11.8)
   python -m pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
   
   # Install additional libraries
   python -m pip install pydot==2.0.0 igraph==0.11.6
   python -m pip install scikit-learn

Running example

We provide an example that demonstrates how to run the A4NN workflow on the CIFAR-10 dataset. This example generates 25 neural network architectures, each trained for 10 epochs, with A4NN performing early termination. Follow the steps below to run the example:

1. Activate the Spack environment and Python virtual environment for the project

   spack env activate <name-of-the-environment>
   spack load py-pip
   

2. Modify the configuration files as needed (e.g., add the corresponding paths). The configuration files are located in the configs/ directory. A README file is included in configs/, which describes each configuration field

3. Navigate to the nas_search/ directory and run the example script:

   cd /path/to/A4NN_worfklow/nas_search
   source ./run-wilkins-cifar10-example.sh
   

Publications

Kin Wai Ng, Orcun Yildiz, Tom Peterka, Florence Tama, Osamu Miyashita, Catherine Schuman, and Michela Taufer. 2025. Energy-Efficient Neural Network Training for Scientific Datasets with Advanced Similarity Analytics and Orchestration. In Proceedings of the International Conference on Computational Science (ICCS). Springer, Singapore, 1–15.

Copyright and License

Copyright (c) 2025, Global Computing Lab

A4NN is distributed under terms of the Apache License, Version 2.0

See LICENSE for more details.

Acknowledgments

This work was supported by the National Science Foundation (NSF) under grant numbers #2331152 and #2223704. This work was supported by Advanced Scientific Computing Research, Office of Science, U.S. Department of Energy, under Contract DE-AC02-06CH11357. Any opinions, findings, and conclusions, or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.

Code inspired and modified from pymoo, nsganet