Manuscript received January 4, 2024; revised January 20, 2024; accepted February 13, 2024; published September 20, 2024
Abstract—With the development of deep neural networks, network compilation plays as an important role for achieving faster execution time. Black-box optimization aims to find an optimal solution by searching the design space. However, it suffers from countless costly hardware measurements, which greatly increase the compilation time. This paper aims to reduce the compilation time by reducing hardware measurements. Our solution includes adaptive early-stop, a self-tuning module that controls tuning workflow according to real-time measurements, a K-means cluster sampling module, a history database that records and organizes measurement results for later usages, a decoupled online tuning service. We extend the work leveraging multiple users of online services, including shared history data and module hyperparameter suggestions. Experiments show our proposed approach achieves 52.39% reduction in hardware measurements for auto-tuning with AutoTVM.
Keywords—TVM, neural network compilation, auto-tuning, auto-optimization.
Cite: Yu-Hsiang Chen, Tay-Jyi Lin, Juin-Ming Lu, Tien-Fu Chen, "Optimizing Neural Network Compilation via Adaptive Workflow with AutoTVM," International Journal of Machine Learning vol. 14, no. 3, pp. 92-96, 2024.
Copyright © 2024 by the authors. This is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited (CC BY 4.0).