YAHPO Gym

What is YAHPO Gym?

YAHPO Gym (Yet Another Hyperparameter Optimization Gym) is a collection of interesting problem sets for benchmark hyperparameter optimization / black-box optimization methods described in this paper. The underlying software with additional documentation and background can be found here. See the module Documentation for more info.

• Problem Variety: Optimization problems in YAHPO Gym stem from diverse Hyperparameter Optimization scenarios on tabular as well as image data.
• Multi-Fidelity: Allows for simulating low-fidelity approximations to the real target values to simulate multi-fidelity HPO.
• Multi-Objective: Benchmarks usually contain multiple objectives: performance metrics, runtime and memory consumption allowing for multi-objective and resource aware HPO.

YAHPO Gym distinguishes between scenarios and instances. A scenario is a collection of instances that share the same hyperparameter space. In practice, a scenario usually consists of a single algorithm optimized on a variety of datasets (= instances).

This repository contains three modules/packages:

• yahpo_gym (python): The core package allowing for inference on the surrogates.
• yahpo_train (python): Module for training surrogate models used in yahpo_gym.
• yahpo_gym_r(R): An R wrapper for yahpo_gym.

We also maintain a list of frequently asked questions.

NEWS:

• The paper accompanying YAHPO Gym was accepted at the 1st International Conference on Automated Machine Learning!
• YAHPO Gym (python) can now be installed via pip install yahpo-gym
• YAHPO Gym is now available in HPOBench
• YAHPO Gym is now available in syne-tune

Why should I use it?

YAHPO Gym (Yet Another Hyperparameter Optimization Gym) provides blazing fast and simple access to a variety of interesting benchmark problems for hyperparameter optimization. Since all our benchmarks are based on surrogate models that approximate the underlying HPO problems with very high fidelity, function evaluations are fast and memory friendly allowing for fast benchmarks across a large variety of problems.

Overview over benchmark instances

Scenario	Search Space	# Instances	Target Metrics	Fidelity	H	Source
rbv2_super	38D: Mixed	103	9: perf(6) + rt(2) + mem	fraction	✓	[1]
rbv2_svm	6D: Mixed	106	9: perf(6) + rt(2) + mem	fraction	✓	[1]
rbv2_rpart	5D: Mixed	117	9: perf(6) + rt(2) + mem	fraction		[1]
rbv2_aknn	6D: Mixed	118	9: perf(6) + rt(2) + mem	fraction		[1]
rbv2_glmnet	3D: Mixed	115	9: perf(6) + rt(2) + mem	fraction		[1]
rbv2_ranger	8D: Mixed	119	9: perf(6) + rt(2) + mem	fraction	✓	[1]
rbv2_xgboost	14D: Mixed	119	9: perf(6) + rt(2) + mem	fraction	✓	[1]
nb301	34D: Categorical	1	2: perf(1) + rt(1)	epoch	✓	[2], [3]
lcbench	7D: Numeric	34	6: perf(5) + rt(1)	epoch		[4], [5]
iaml_super	28D: Mixed	4	12: perf(4) + inp(3) + rt(2) + mem(3)	fraction	✓	[6]
iaml_rpart	4D: Numeric	4	12: perf(4) + inp(3) + rt(2) + mem(3)	fraction		[6]
iaml_glmnet	2D: Numeric	4	12: perf(4) + inp(3) + rt(2) + mem(3)	fraction		[6]
iaml_ranger	8D: Mixed	4	12: perf(4) + inp(3) + rt(2) + mem(3)	fraction	✓	[6]
iaml_xgboost	13D: Mixed	4	12: perf(4) + inp(3) + rt(2) + mem(3)	fraction	✓	[6]

The full, up-to-date overview can be obtained from the Documentation. The fidelity is given either as the dataset fraction fraction or the number of epochs epoch. Search spaces can be numeric, mixed and have dependencies (as indicated in the H column).

Original data sources are given by:

• [1] Binder M., Pfisterer F. & Bischl B. (2020). Collecting Empirical Data About Hyperparameters for Data Driven AutoML. 7th ICML Workshop on Automated Machine Learning.
• [2] Siems, J., Zimmer, L., Zela, A., Lukasik, J., Keuper, M., & Hutter, F. (2020). NAS-Bench-301 and the Case for Surrogate Benchmarks for Neural Architecture Search. arXiv preprint arXiv:2008.09777, 11.
• [3] Zimmer, L. (2020). nasbench301_full_data. figshare. Dataset. https://doi.org/10.6084/m9.figshare.13286105.v1, Apache License, Version 2.0.
• [4] Zimmer, L., Lindauer, M., & Hutter, F. (2021). Auto-Pytorch: Multi-Fidelity Metalearning for Efficient and Robust AutoDL. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(9), 3079-3090.
• [5] Zimmer, L. (2020). data_2k_lw.zip. figshare. Dataset. https://doi.org/10.6084/m9.figshare.11662422.v1, Apache License, Version 2.0.
• [6] None, simply cite Pfisterer, F., Schneider, L., Moosbauer, J., Binder, M., & Bischl, B. (2022). YAHPO Gym - An Efficient Multi-Objective Multi-Fidelity Benchmark for Hyperparameter Optimization. In International Conference on Automated Machine Learning.

Please make sure to always also cite the original data sources as YAHPO Gym would not have been possible without them!

What does this repository contain?

This repository contains two modules: yahpo_gym and yahpo_train. While we mainly focus on yahpo_gym, as it is provides an interface to the benchmark described in our paper, we also provide the full reproducible codebase used to generate the underlying surrogate neural networks in yahpo_train.

YAHPO Gym

YAHPO Gym is the module for inference and allows for evaluating a HPC configuration on a given benchmark instance.

Surrogate models (ONNX files), configspaces and metadata (encoding) can be obtained here (Github).

An example for evaluation and running HPO methods is given in the README of the YAHPO Gym module.

A quick introduction is given in the accompanying jupyter notebook.

YAHPO Train

YAHPO Train is the module for training new surrogate models.

YAHPO Train is still in a preliminary state but can already be used to reproduce and refit models introduced in our paper.

Docker

A docker image that allows accessing yahpo_gym is available from DockerHub at pfistfl/yahpo. This adds additional overhead but simplifies use and installation. The corresponding dockerfile to get you started can be found in docker/.

Roadmap

We want to add several features to yahpo_gym in future versions:

• Asynchronous Evaluation We would like to allow for faster-than-realtime asynchronous evaluation in future versions. This is currently available as an experimental feature via objective_function_timed, but requires additional (experimental) evaluation for release.

• Noisy Surrogate Models We would like to allow for surrogates that more closely reflect the underlying (noisy) nature of real HPO experiments. Currently, noisy evaluation are available using noisy = True during instantiation, but this feature is considered experimental and requires additional evaluation for release.

• Additional Scenarios We are always happy to include additional (interesting) scenarios. If you know of (or want to add) an additional scenario, get in touch!

  We welcome input, discussion or additions by the broader community. Get in touch via issues or emails if you have questions, comments or would like to collaborate!

Related Software

• rbv2 (R-Package) can be used to reproduce runs from all rbv2_* in a real setting.
• iaml (R-Package) can be used to reproduce runs from all iaml_* in a real setting.

Citation

If you use YAHPO Gym, please cite the following paper:

• Pfisterer, F., Schneider, L., Moosbauer, J., Binder, M., & Bischl, B. (2022). YAHPO Gym - An Efficient Multi-Objective Multi-Fidelity Benchmark for Hyperparameter Optimization. In International Conference on Automated Machine Learning.

Moreover, certain scenarios built upon previous work, e.g., the lcbench scenario uses data from:

• Zimmer, L., Lindauer, M., & Hutter, F. (2021). Auto-Pytorch: Multi-Fidelity Metalearning for Efficient and Robust AutoDL. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(9), 3079-3090.
• Zimmer, L. (2020). data_2k_lw.zip. figshare. Dataset. https://doi.org/10.6084/m9.figshare.11662422.v1, Apache License, Version 2.0.

Please make sure to always also cite the original data sources as YAHPO Gym would not have been possible without them!

Original data sources of a scenario that should also be cited are provided via the "citation" key within the config dictionary of a scenario, e.g.:

from yahpo_gym.configuration import cfg
lcbench = cfg("lcbench")
lcbench.config.get("citation")