Parallel Propagator of Populations

Click here to watch our 3 min introduction video!

What Propulate can do for you

Propulate is an HPC-tailored software for solving optimization problems in parallel. It is openly accessible and easy to use. Compared to a widely used competitor, Propulate is consistently faster - at least an order of magnitude for a set of typical benchmarks - and in some cases even more accurate.

Inspired by biology, Propulate borrows mechanisms from biological evolution, such as selection, recombination, and mutation. Evolution begins with a population of solution candidates, each with randomly initialized genes. It is an iterative "survival of the fittest" process where the population at each iteration can be viewed as a generation. For each generation, the fitness of each candidate in the population is evaluated. The genes of the fittest candidates are incorporated in the next generation.

Like in nature, Propulate does not wait for all compute units to finish the evaluation of the current generation. Instead, the compute units communicate the currently available information and use that to breed the next candidate immediately. This avoids waiting idly for other units and thus a load imbalance. Each unit is responsible for evaluating a single candidate. The result is a fitness level corresponding with that candidate’s genes, allowing us to compare and rank all candidates. This information is sent to other compute units as soon as it becomes available. When a unit is finished evaluating a candidate and communicating the resulting fitness, it breeds the candidate for the next generation using the fitness values of all candidates it evaluated and received from other units so far.

Propulate can be used for hyperparameter optimization and neural architecture search at scale. It was already successfully applied in several accepted scientific publications. Applications include grid load forecasting, remote sensing, and structural molecular biology:

J. Debus, C. Debus, G. Dissertori, et al. PETNet–Coincident Particle Event Detection using Spiking Neural Networks. 2024 Neuro Inspired Computational Elements Conference (NICE), La Jolla, CA, USA, pp. 1-9 ( 2024). https://doi.org/10.1109/NICE61972.2024.10549584

D. Coquelin, K. Flügel, M. Weiel, et al. AB-Training: A Communication-Efficient Approach for Distributed Low-Rank Learning. arXiv preprint (2024). https://doi.org/10.48550/arXiv.2405.01067

D. Coquelin, K. Flügel, M. Weiel, et al. Harnessing Orthogonality to Train Low-Rank Neural Networks. arXiv preprint (2024). https://doi.org/10.48550/arXiv.2401.08505

A. Weyrauch, T. Steens, O. Taubert, et al. ReCycle: Fast and Efficient Long Time Series Forecasting with Residual Cyclic Transformers. 2024 IEEE Conference on Artificial Intelligence (CAI), Singapore, pp. 1187-1194 (2024). https://doi.org/10.1109/CAI59869.2024.00212

O. Taubert, F. von der Lehr, A. Bazarova, et al. RNA contact prediction by data efficient deep learning. Communications Biology 6(1), 913 (2023). https://doi.org/10.1038/s42003-023-05244-9

D. Coquelin, K. Flügel, M. Weiel, et al. Harnessing Orthogonality to Train Low-Rank Neural Networks. arXiv preprint (2023). https://doi.org/10.48550/arXiv.2401.08505

Y. Funk, M. Götz, and H. Anzt. Prediction of optimal solvers for sparse linear systems using deep learning. Proceedings of the 2022 SIAM Conference on Parallel Processing for Scientific Computing (pp. 14-24). Society for Industrial and Applied Mathematics (2022). https://doi.org/10.1137/1.9781611977141.2

D. Coquelin, R. Sedona, M. Riedel, and M. Götz. Evolutionary Optimization of Neural Architectures in Remote Sensing Classification Problems. IEEE International Geoscience and Remote Sensing Symposium IGARSS, Brussels, Belgium, pp. 1587-1590 (2021). https://doi.org/10.1109/IGARSS47720.2021.9554309

In more technical terms

Propulate is a massively parallel evolutionary hyperparameter optimizer based on the island model with asynchronous propagation of populations and asynchronous migration. In contrast to classical GAs, Propulate maintains a continuous population of already evaluated individuals with a softened notion of the typically strictly separated, discrete generations. Our contributions include:

• A novel parallel genetic algorithm based on a fully asynchronized island model with independently processing workers.
• Massive parallelism by asynchronous propagation of continuous populations and migration via efficient communication using the message passing interface.
• Optimized use efficiency of parallel hardware by minimizing idle times in distributed computing environments.

To be more efficient, the generations are less well separated than they usually are in evolutionary algorithms. New individuals are generated from a pool of currently active, already evaluated individuals that may be from any generation. Individuals may be removed from the breeding population based on different criteria.

You can find the corresponding publication here:

Taubert, O. et al. (2023). Massively Parallel Genetic Optimization Through Asynchronous Propagation of Populations. In: Bhatele, A., Hammond, J., Baboulin, M., Kruse, C. (eds) High Performance Computing. ISC High Performance 2023. Lecture Notes in Computer Science, vol 13948. Springer, Cham. doi.org/10.1007/978-3-031-32041-5_6

Documentation

Check out the full documentation at https://propulate.readthedocs.io/ :rocket:! Here you can find installation instructions, tutorials, theoretical background, and API references.

:point_right: If you have any questions or run into any challenges while using Propulate, don't hesitate to post an issue :bookmark:, reach out via GitHub discussions :octocat:, or contact us directly via e-mail :email: to propulate@lists.kit.edu.

Installation

• You can install the latest stable release from PyPI: pip install propulate
• If you need the latest updates, you can also install Propulate directly from the master branch. Pull and run pip install ..
• If you want to run the tutorials, you can install the required dependencies via: pip install ."[tutorials]"
• If you want to contribute to Propulate as a developer, you need to install the required dependencies with the package: pip install -e ."[dev]".

Propulate depends on mpi4py and requires an MPI implementation under the hood. Currently, it is only tested with OpenMPI.

Quickstart

Below, you can find a quick recipe for how to use Propulate in general. Check out the official ReadTheDocs documentation for more detailed tutorials and explanations.

Let's minimize the sphere function $f_\text{sphere}\left(x,y\right)=x^2 +y^2$ with Propulate as a quick example. The minimum is at $\left(x, y\right)=\left(0,0\right)$ at the orange star. First, we need to define the key ingredients that define our optimization problem:

• The search space of the parameters to be optimized as a Python dictionary. Propulate can handle three different parameter types:

  • A tuple of float for a continuous parameter, e.g., {"learning_rate": (0.0001, 0.01)}
  • A tuple of int for an ordinal parameter, e.g., {"conv_layers": (2, 10)}
  • A tuple of str for a categorical parameter, e.g., {"activation": ("relu", "sigmoid", "tanh")}

  Thus, an exemplary search space might look like this:

  search_space = {
      "learning_rate": (0.0001, 0.01),  # Search a continuous space between 0.0001 and 0.01.
      "num_layers": (2, 10),  # Search the integer space between 2 and 10 (inclusive).
      "activation": ("relu", "sigmoid",