NAS-Bench-Suite: NAS Evaluation is (Now) Surprisingly Easy

Authors: Yash Mehta, Colin White, Arber Zela, Arjun Krishnakumar, Guri Zabergja, Shakiba Moradian, Mahmoud Safari, Kaicheng Yu, Frank Hutter

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this work, we present an in-depth analysis of popular NAS algorithms and performance prediction methods across 25 different combinations of search spaces and datasets, finding that many conclusions drawn from a few NAS benchmarks do not generalize to other benchmarks.
Researcher Affiliation Collaboration 1 University of Freiburg, 2 Abacus.AI, 3 Bosch Center for AI
Pseudocode Yes Snippet 1: A minimal example on how one can run a NAS algorithm in NAS-Bench-Suite. Both the search space and the algorithm can be changed in one line of code.
Open Source Code Yes Our code is available at https://github.com/automl/naslib.
Open Datasets Yes This benchmark consists of 423 624 architectures trained on CIFAR-10. ... The search space consists of 8 242 architectures trained on the TIMIT dataset. ... evaluated across four datasets: Image Net50-1000, Cityscapes, KITTI, and HMDB51.
Dataset Splits Yes NAS-Bench-101 comes with precomputed validation and test accuracies at epochs 4, 12, 36, and 108 from training on CIFAR-10. ... Each architecture has precomputed train, validation, and test losses and accuracies for 200 epochs on CIFAR-10, CIFAR-100, and Image Net-16-120.
Hardware Specification No This is in contrast to the NAS-Bench-Suite, where NAS algorithms take at most 5 minutes on a CPU due to the use of queryable benchmarks. (Mentioned CPU and GPU but no specific models or configurations).
Software Dependencies No A search space is defined with a graph object using Py Torch and Network X (Hagberg et al., 2008)... We use the original implementation from the pybnn package. ... We use the Scikit-learn implementation (Pedregosa et al., 2011). ... We used the original code (Chen & Guestrin, 2016). (Software names are mentioned, but no specific version numbers are provided for PyTorch, NetworkX, pybnn, Scikit-learn, or XGBoost.)
Experiment Setup No For a list of the default hyperparameters and hyperparameter ranges, see https://github.com/automl/NASLib. (Specific hyperparameter values are explicitly deferred to an external link, not provided in the main text of the paper.)