Skip to content

Benchmarks used for `HPOBench: A Collection of Reproducible Multi Fidelity Benchmark Problems for HPO`

Katharina Eggensperger edited this page Sep 19, 2022 · 4 revisions

For HPOBench: A Collection of Reproducible Multi-Fidelity Benchmark Problems for HPO we used the following

existing community benchmarks

family container version name in the paper benchmarks Reference
nas.nasbench_201 0.0.5 NB201 [Cifar10ValidNasBench201BenchmarkOriginal, Cifar100NasBench201BenchmarkOriginal, ImageNetNasBench201BenchmarkOriginal] paper
nas.nasbench_101 0.0.4 NB101 [NASCifar10ABenchmark, NASCifar10BBenchmark, NASCifar10CBenchmark] paper
nas.tabular_benchmarks 0.0.5 NBHPO [SliceLocalizationBenchmarkOriginal, ProteinStructureBenchmarkOriginal, NavalPropulsionBenchmarkOriginal, ParkinsonsTelemonitoringBenchmarkOriginal] paper
nas.nasbench_1shot1 0.0.4 NB1SHOT1 [NASBench1shot1SearchSpace1Benchmark, NASBench1shot1SearchSpace2Benchmark, NASBench1shot1SearchSpace3Benchmark] paper
ml.pybnn 0.0.4 BNN [BNNOnProteinStructure, BNNOnYearPrediction] paper
rl.cartpole 0.0.4 Cartpole [CartpoleReduced] paper
surrogates.paramnet_benchmark 0.0.4 Net [ParamNetReducedAdultOnTimeBenchmark, ParamNetReducedHiggsOnTimeBenchmark, ParamNetReducedLetterOnTimeBenchmark, ParamNetReducedMnistOnTimeBenchmark, ParamNetReducedOptdigitsOnTimeBenchmark, ParamNetReducedPokerOnTimeBenchmark] paper

and new benchmarks

family container version name in the paper benchmarks Reference
ml.tabular_benchmark - LogReg TabularBenchmark for model = 'lr' -
ml.tabular_benchmark - SVM TabularBenchmark for model = 'svm' -
ml.tabular_benchmark - XGBoost TabularBenchmark for model = 'xgb' -
ml.tabular_benchmark - RF TabularBenchmark for model = 'rf' -
ml.tabular_benchmark - MLP TabularBenchmark for model = 'nn' -

We host all code to recreate the experiments in this repo: https://github.com/automl/HPOBenchExperimentUtils

The best-known values for all benchmarks we used to plot regret can be found here: https://github.com/automl/HPOBenchExperimentUtils/blob/master/HPOBenchExperimentUtils/utils/plotting_utils.py