CARP-S-example-optimizer

This is an example of an optimizer adhering to the carps interface and how this optimizer can be run in carps just via the commandline. The optimizer here is random search (which is just copy-pasted from carps).

Installation

# Create conda env
conda create -n carpsopt python=3.12 -c conda-forge

# Activate env
conda activate carpsopt

# Clone repo
git clone [email protected]:automl/CARP-S-example-optimizer.git
cd CARP-S-example-optimizer

# Install 
pip install -e .

Running a benchmark from carps

The command for running the optimizer on a benchmark, e.g., a BBOB function from carps looks as follows:

# we need to install the BBOB benchmark requirements
python -m carps.build.make benchmark_bbob

# Run the optimizer on the task
python -m carps.run 'hydra.searchpath=[pkg://carp_s_example_optimizer/configs]' +myoptimizer/randomsearch=config +task/BBOB=cfg_2_12_1 seed=1

The breakdown of the command:

• 'hydra.searchpath=[pkg://carp_s_example_optimizer/configs]': Let hydra know where to find the configs of the optimizer package. For this, carp_s_example_optimizer needs to be installed. It is in general: pkg://PACKAGE_NAME/PATH_INSIDE_PACKAGE_TO_CONFIGS_FOLDER .
• +myoptimizer/randomsearch=config: select an optimizer from your package. Follows the config folder structure in carp_s_example_optimizer. Maybe make sure that config folder names don't overlap with carps.
• +task/BBOB=cfg_2_12_1: Select a problem. Follows the config folder structure in carps. Beware, for other benchmarks you need to install dependencies (check the repo).
• seed=1: Set the seed to 1. 🙂

Of course, you can also specify the run dir etc. For more hydra overrides and parallelization, check their docs/tutorials.