FedHPO-Bench: A Benchmark Suite for Federated Hyperparameter Optimization

Authors: Zhen Wang, Weirui Kuang, Ce Zhang, Bolin Ding, Yaliang Li

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct extensive experiments based on FEDHPO-BENCH to provide the community with more insights into Fed HPO.
Researcher Affiliation Collaboration 1Alibaba Group 2ETH Zürich. Correspondence to: Yaliang Li <yaliang.li@alibaba-inc.com>.
Pseudocode Yes Figure 6. A general algorithmic view for Fed HPO methods: They are allowed to concurrently explore different client-side configurations in the same round of FL, but the clients are heterogeneous, i.e., corresponding to different functions f (c) i ( ). Operators in brackets are optional.
Open Source Code Yes We open-sourced FEDHPO-BENCH at https://github.com/alibaba/Federated Scope/tree/ master/benchmark/Fed HPOBench.
Open Datasets Yes All these datasets are publicly available and can be downloaded and preprocessed by our prepared scripts.
Dataset Splits Yes FEMNIST... And we use the default train/valid/test splits for each client, where the ratio is 60% : 20% : 20%.
Hardware Specification Yes In creating them, we spent about two months of computation time on six machines, each with four Nvidia V100 GPUs.
Software Dependencies Yes Table 5. Overview of the optimizers from widely adopted libraries. Name Model Packages version RS (Bergstra & Bengio, 2012) HPBandster 0.7.4 BOGP (Hutter et al., 2011; Lindauer et al., 2022) GP SMAC3 1.3.3 BOKDE (Falkner et al., 2018) KDE HPBandster 0.7.4 TPEHB (Akiba et al., 2019) TPE Optuna 2.10.0
Experiment Setup Yes As a full-fidelity function evaluation consumes 500 rounds on these datasets, we specify the total budget to 2,500 (i.e., 5 times the budget of a full-fidelity evaluation) in terms of #round. Precisely, each BBO method consists of 50 trials, each of which runs for 50 rounds. For MF optimizers, we set η of Successive Halving Algorithm (SHA) (Jamieson & Talwalkar, 2016) to 3, the min and max budget to 9 and 81 rounds, respectively.