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., `{"learningrate": (0.0001, 0.01)} - A tuple ofintfor an ordinal parameter, e.g.,{"conv_layers": (2, 10)} - A tuple ofstrfor a categorical parameter, e.g.,{"activation": ("relu", "sigmoid", "tanh")}`

Thus, an exemplary search space might look like this: python 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", "tanh"), # Search the categorical space with the specified possibilities. }

The sphere function has two continuous parameters, $x$ and $y$, and we consider $x,y\in\left[-5.12,5.12\right]$. The search space in our example thus looks like this: python limits = { "x": (-5.12, 5.12), "y": (-5.12, 5.12) } - The loss function. This is the function we want to minimize in order to find the best parameters. It can be any Python function that - takes a set of parameters as a Python dictionary as an input. - returns a scalar loss value that determines how good the tested parameter set is.

In this example, the loss function whose minimum we want to find is the sphere function: ```python def sphere(params: Dict[str, float]) -> float: """ Sphere function: continuous, convex, separable, differentiable, unimodal

Input domain: -5.12 <= x, y <= 5.12
Global minimum 0 at (x, y) = (0, 0)

Parameters
----------
params: Dict[str, float]
    The function parameters.

Returns
-------
float
    The function value.
"""
return numpy.sum(numpy.array(list(params.values())) ** 2).item()

Next, we need to define the **evolutionary operator** or propagator that we want to use to breed new individuals during the optimization process. `Propulate` provides a reasonable default propagator via a utility function: python

Set up logger for Propulate optimization.

propulate.setloggerconfig()

Set up separate random number generator for Propulate optimization. DO NOT USE SOMEWHERE ELSE!

rng = random.Random( + mpi4py.MPI.COMM_WORLD.rank )

Set up evolutionary operator.

propagator = propulate.getdefaultpropagator( popsize=config.popsize, # The breeding population size limits=limits, # The search-space limits rng=rng, # Random number generator ) We also need to set up the asynchronous parallel evolutionary **optimizer**, that is a so-called ``Propulator`` instance: python

Set up Propulator performing actual optimization.

propulator = propulate.Propulator( lossfn=sphere, propagator=propagator, rng=rng, generations=config.generations, checkpointpath=config.checkpoint, ) Now we can run the actual optimization. Overall, ``generations * mpi4py.MPI.COMM_WORLD.size`` evaluations will be performed: python

Run optimization and print summary of results.

propulator.propulate() propulator.summarize() The output should look something like this: text

PROPULATE: Parallel Propagator of Populations

[2024-03-12 14:37:01,374][propulate.propulator][INFO] - No valid checkpoint file given. Initializing population randomly... [2024-03-12 14:37:01,374][propulate.propulator][INFO] - Island 0 has 4 workers. [2024-03-12 14:37:01,374][propulate.propulator][INFO] - Island 0 Worker 0: In generation 0... [2024-03-12 14:37:01,374][propulate.propulator][INFO] - Island 0 Worker 3: In generation 0... [2024-03-12 14:37:01,374][propulate.propulator][INFO] - Island 0 Worker 2: In generation 0... [2024-03-12 14:37:01,374][propulate.propulator][INFO] - Island 0 Worker 1: In generation 0... [2024-03-12 14:37:01,377][propulate.propulator][INFO] - Island 0 Worker 3: In generation 10... [2024-03-12 14:37:01,377][propulate.propulator][INFO] - Island 0 Worker 1: In generation 10... [2024-03-12 14:37:01,378][propulate.propulator][INFO] - Island 0 Worker 0: In generation 10... [2024-03-12 14:37:01,378][propulate.propulator][INFO] - Island 0 Worker 2: In generation 10...

... [2024-03-12 14:37:02,197][propulate.propulator][INFO] - Island 0 Worker 1: In generation 960... [2024-03-12 14:37:02,206][propulate.propulator][INFO] - Island 0 Worker 2: In generation 990... [2024-03-12 14:37:02,206][propulate.propulator][INFO] - Island 0 Worker 1: In generation 970... [2024-03-12 14:37:02,215][propulate.propulator][INFO] - Island 0 Worker 1: In generation 980... [2024-03-12 14:37:02,224][propulate.propulator][INFO] - Island 0 Worker 1: In generation 990... [2024-03-12 14:37:02,232][propulate.propulator][INFO] - OPTIMIZATION DONE. NEXT: Final checks for incoming messages... [2024-03-12 14:37:02,244][propulate.propulator][INFO] - ###########

SUMMARY

Number of currently active individuals is 4000. Expected overall number of evaluations is 4000. [2024-03-12 14:37:03,703][propulate.propulator][INFO] - Top 1 result(s) on island 0: (1): [{'a': '2.91E-3', 'b': '-3.05E-3'}, loss 1.78E-5, island 0, worker 0, generation 956] ```

Let's get your hands dirty

Do the following to run the example script:

• Make sure you have a working MPI installation on your machine.
• If you have not already done this, create a fresh virtual environment with Python: $ python3 -m venv best-venv-ever
• Activate it: $ source best-venv-ever/bin/activate
• Upgrade pip: $ pip install --upgrade pip
• Install Propulate: $ pip install propulate
• Run the example script propulator_example.py: $ mpirun --use-hwthread-cpus python propulator_example.py

Acknowledgments

This work is supported by the Helmholtz AI platform grant.