redsho

Robust Evolutionary Direction Set Hyperparameter Optimizer

Redsho is a discrete optimizer, which means that it works with parameters that can only take on a certain set of values. This works well for evaluating neural networks and other large and complex models, since they take so long to train and test, and since the hyperparameters can have non-intuitive, strongly non-linear, and interactive effects.

Redsho (pictured in action above), an evolutionary search algorithm variant inspired by direction set methods like Powell's method, so much so that it was originally called Evolutionary Powell's method. Here is a detailed description of how it works.

Installation

It's on PyPI, so install with

pip install redsho

or as part of a uv environment

uv add redsho

If you want to experiment with tweaking the algorithm, clone the repository to your local machine and install it from there.

git clone https://codeberg.org/brohrer/redsho.git
python3 -m pip install -e redsho

Run the demo

In a python script

import redsho.demo

Parallelization

Redsho can seamlessly take advantage of multiple processor systems. Instead of

from redsho.optimizer import Redsho

try

from redsho.parallel_optimizer import ParallelRedsho

It automatically recruits all your processors (but one) to do its bidding.