Welcome to CARP-S! This repository contains a benchmarking framework for optimizers. It allows flexibly combining optimizers and benchmarks via a simple interface, and logging experiment results and trajectories to a database. carps can launch experiment runs in parallel by using hydra, which offers launchers for slurm/submitit, Ray, RQ, and joblib.
The main topics of this README are:
For more details on CARP-S, please have a look at the documentation or our blog post.
To install CARP-S, you can simply use pip:
- Create virtual env with conda or uv
# Conda
conda create -n carps python=3.12
conda activate carps
# -OR -
# uv
pip install uv
export PIP="uv pip" # Env var needed for Makefile commands
uv venv --python=3.12 carpsenv
source carpsenv/bin/activate- Install carps.
pip install carpsIf you want to install from source, you can clone the repository and install CARP-S via:
git clone https://github.com/AutoML/CARP-S.git
cd CARP-S
export PIP="pip"
conda create -n carps python=3.12
conda activate carps
# Install for usage
$PIP install .git clone https://github.com/AutoML/CARP-S.git
cd CARP-S
pip install uv
export PIP="uv pip"
uv venv --python=3.12 carpsenv
source carpsenv/bin/activate
# Install for usage
$PIP install .
# Install as editable
$PIP install -e .If you want to install CARP-S for development, you can use the following command:
make install-devYou can also create a container with the env setup by running apptainer build container/env.sif container/env.def.
Then you can execute any carps commands as usual by add this prefix apptainer exec container/env.sif before the
command, e.g. apptainer exec container/env.sif python -m carps.run +task/... +optimizer/....
There is also an sbatch script to run experiments from the database using the apptainer on a slurm cluster
(sbatch scripts/container_run_from_db.sh). You might need to adapt the array size and the number of repetitions
according to the number of experiments you can run.
PS.: On some clusters you might need to load the module apptainer like so module load tools Apptainer.
Troubleshooting: If you have problems writing your cache directory, mount-bind it like so
apptainer shell --bind $XDG_CACHE_HOME container/env.sif. This binds the directory $XDG_CACHE_HOME in the
container to the directory $XDG_CACHE_HOME on the host.
If you have problems with /var/lib/hpobench, this bind might help:
<hpobench data dir>:/var/lib/hpobench/data. <hpobench data dir> can be found in
.hpobenchrc.
For python3.12, numpy should be numpy>=2.0.0. For python3.10, numpy must be numpy==1.26.4, you can simply
pip install numpy==1.26.4 after running the proposed install commands.
Additionally, you need to install the requirements for the benchmark and optimizer that you want to use.
⚠ You can specify the directory of the task data by export CARPS_TASK_DATA_DIR=.... Please use absolute dirnames.
The default location is <carps package location>/task_data. If you specify a custom dir, always export the env var.
For example, if you want to use the SMAC3 optimizer and the BBOB benchmark, you need to install the
requirements for both of them via:
# Install options for optimizers and benchmarks (these are Makefile commands, check the Makefile for more commands)
# The commands should be separated by a whitespace
python -m carps.build.make benchmark_bbob optimizer_smacThe benchmarks and optimizers can all be installed in one environment (tested with python3.12).
All possible install options for benchmarks are:
benchmark_bbob benchmark_hpobench benchmark_hpob benchmark_mfpbench benchmark_pymoo benchmark_yahpo
⚠ Some benchmarks require to download surrogate models and/or containers and thus might take disk space and time to download.
All possible install options for optimizers are:
optimizer_smac optimizer_dehb optimizer_nevergrad optimizer_optuna optimizer_ax optimizer_skopt optimizer_synetune
All of the above except optimizer_hebo work with python3.12.
You can also install all benchmarks in one go with benchmarks and all optimizers with optimizers.
Check the Makefile in carps for more details.
Once the requirements for both an optimizer and a benchmark, e.g. SMAC2.0 and BBOB, are installed, you can run
one of the following minimal examples to benchmark SMAC2.0 on BBOB directly with Hydra:
# Run SMAC BlackBoxFacade on certain BBOB task
python -m carps.run +optimizer/smac20=blackbox +task/BBOB=cfg_4_1_0 seed=1 task.optimization_resources.n_trials=25
# Run SMAC BlackBoxFacade on all available BBOB tasks for 10 seeds
python -m carps.run +optimizer/smac20=blackbox '+task/BBOB=glob(*)' 'seed=range(1,11)' -mFor the second command, the Hydra -m (or --multirun) option indicates that multiple runs will be
performed over a range of parameter values. In this case, it's indicating that the benchmarking
should be run for all available BBOB tasks (+task/BBOB=glob(*)) and for 10 different
seed values (seed=range(1,11)).
You can run a certain task and optimizer combination directly with Hydra via:
python -m carps.run +task=... +optimizer=... seed=... -mTo check whether any runs are missing, you can use the following command. It will create
a file runcommands_missing.sh containing the missing runs:
python -m carps.utils.check_missing <rundir>To collect all run data generated by the file logger into csv files, use the following command:
python -m carps.analysis.gather_data <rundir>The csv files are then located in <rundir>. logs.csv contain the trial info and values and
logs_cfg.csv contain the experiment configuration.
The experiments can be matched via the column experiment_id.
Per default, carps logs to files. This has its caveats: Checking experiment status is a bit more cumbersome (but
possible with python -m carps.utils.check_missing <rundir> to check for missing/failed experiments) and reading from
the filesystem takes a long time. For this reason, we can also control and log experiments to a MySQL database with
PyExperimenter.
Requirement: MySQL database is set up.
- Add a
credentials.yamlfile incarps/experimenterwith the following content:
CREDENTIALS:
Database:
user: someuser
password: amazing_password
Connection:
Standard:
server: mysql_server
port: 3306 (most likely)- Edit
carps/experimenter/py_experimenter.yamlby setting:
PY_EXPERIMENTER:
n_jobs: 1
Database:
use_ssh_tunnel: false
provider: mysql
database: your_database_name
...!!! Note: If you use an ssh tunnel, set use_ssh_tunnel to true in carps/experimenter/py_experimenter.yaml.
Set up carps/experimenter/credentials.yaml like this:
CREDENTIALS:
Database:
user: someuser
password: amazing_password
Connection:
Standard:
server: mysql_server
port: 3306 (most likely)
Ssh:
server: 127.0.0.1
address: some_host # hostname as specified in ~/.ssh/config
# ssh_private_key_password: null
# server: example.sshmysqlserver.com (address from ssh server)
# address: example.sslserver.com
# port: optional_ssh_port
# remote_address: optional_mysql_server_address
# remote_port: optional_mysql_server_port
# local_address: optional_local_address
# local_port: optional_local_port
# passphrase: optional_ssh_passphraseFirst, in order for PyExperimenter to be able to pull experiments from the database, we need to fill it. The general command looks like this:
python -m carps.experimenter.create_cluster_configs +task=... +optimizer=... -mAll subset runs were created with scripts/create_experiments_in_db.sh.
Now, execute experiments with:
python -m carps.run_from_db 'job_nr_dummy=range(1,1000)' -mThis will create 1000 multirun jobs, each pulling an experiment from PyExperimenter and executing it.
!!! Note: On most slurm clusters the max array size is 1000.
!!! Note: On our mysql server location, at most 300 connections at the same time are possible. You can limit your number
of parallel jobs with hydra.launcher.array_parallelism=250.
!!! carps/configs/runfromdb.yaml configures the run and its resources. Currently defaults for our slurm cluster are
configured. If you run on a different cluster, adapt hydra.launcher.
Experiments with error status (or any other status) can be reset via:
python -m carps.experimenter.database.reset_experimentsFor instructions on how to add a new optimizer or benchmark, please refer to the contributing guidelines for benchmarks and optimizers.
In the case when you are developing your optimizer or benchmark in a standalone package, you can use carps without directly working in the carps repo. For a custom benchmark we have an example repo. It shows how to use your own benchmark with carps optimizers. For a custom optimizer check this example repo. Information is also available here.
For each task_type (blackbox, multi-fidelity, multi-objective and multi-fidelity-multi-objective) and set (dev and test), we run selected optimizers and provide the data.
Here we provide the link to the meta data
that contains the detailed optimization setting for each run
and the running results that
records the running results of each optimization-benchmark combination.