ProxyBO: Accelerating Neural Architecture Search via Bayesian Optimization with Zero-Cost Proxies

Authors: Yu Shen, Yang Li, Jian Zheng, Wentao Zhang, Peng Yao, Jixiang Li, Sen Yang, Ji Liu, Bin Cui

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive empirical studies show that Proxy BO consistently outperforms competitive baselines on five tasks from three public benchmarks.
Researcher Affiliation Collaboration Yu Shen1,4, Yang Li5, Jian Zheng3, Wentao Zhang6,7, Peng Yao4, Jixiang Li4, Sen Yang4, Ji Liu4, Bin Cui1,2 1Key Lab of High Confidence Software Technologies, Peking University, China 2Institute of Computational Social Science, Peking University (Qingdao), China 3School of Computer Science and Engineering, Beihang University, China 4Kuaishou Technology, China 5Data Platform, TEG, Tencent Inc., China 6Mila Qu ebec AI Institute 7HEC, Montr eal, Canada {shenyu,bin.cui}@pku.edu.cn, zhengjian2322@buaa.edu.cn, {yaopeng,lijixiang,senyang}@kuaishou.com, jiliu@kwai.com, thomasyngli@tencent.com, wentao.zhang@mila.quebec
Pseudocode Yes Algorithm 1: Pseudo code for Sample in Proxy BO
Open Source Code No The paper does not include an explicit statement or link to the open-source code for the methodology described in this paper.
Open Datasets Yes To ensure reproducibility as in previous work (Abdelfattah et al. 2021), we conduct the experiments on four public NAS benchmarks: NAS-Bench-101 (Ying et al. 2019), NAS-Bench-201 (Dong and Yang 2020), NAS-Bench ASR (Mehrotra et al. 2021), and NAS-Bench-301 (Siems et al. 2020) with the real-world DARTS search space.
Dataset Splits No The paper mentions using NAS-Bench-101, NAS-Bench-201, NAS-Bench ASR, and NAS-Bench-301 benchmarks, but it does not explicitly provide the specific training/test/validation dataset splits (percentages or counts) used for these benchmarks within the text.
Hardware Specification Yes The experiments are conducted on a machine with 64 AMD EPYC 7702P CPU cores and two RTX 2080Ti GPUs.
Software Dependencies No The paper mentions software like Pytorch, Open Box, xgb, and lgb, but it does not provide specific version numbers for these software dependencies.
Experiment Setup Yes The population size for REA and A-REA is 20; in BRP and Warm-up BRP, 30 models are sampled to train the predictor during each iteration; in Warm-up BRP, we randomly sample 256 models to perform a warm start; if not mentioned, the zero-cost proxy-based methods apply all three proxies. The η and R are set to 3 and 27 in BOHB and MFES, respectively. The τ0 in Proxy BO is set to 0.05. In the last experiment using NAS-Bench-301, we use xgb as the performance predictor and lgb runtime as the runtime predictor.