HW-NAS-Bench
收藏资源简介:
HW-NAS-Bench是由莱斯大学电气与计算机工程系开发的首个公开硬件感知神经架构搜索(HW-NAS)研究数据集,旨在使非硬件专家也能参与HW-NAS研究,并提高研究的重复性和可访问性。该数据集涵盖了NAS-Bench-201和FBNet两个最先进的NAS搜索空间,为所有网络提供了在六种硬件设备上的测量/估计硬件性能数据,包括商用边缘设备、FPGA和ASIC。数据集的创建过程涉及精心收集和分析硬件性能数据,如能量成本和延迟,以提供对HW-NAS研究的深入见解。此外,HW-NAS-Bench还展示了如何通过简单查询预先测量的数据集,使非硬件专家能够执行HW-NAS,并验证了针对特定设备的HW-NAS可以实现最佳的精度-成本权衡。该数据集的应用领域包括深度神经网络的硬件加速器设计和优化,旨在解决如何在资源受限的日常设备中高效部署深度神经网络的问题。
HW-NAS-Bench is the first open-source hardware-aware neural architecture search (HW-NAS) research dataset developed by the Department of Electrical and Computer Engineering at Rice University. It aims to enable non-hardware experts to participate in HW-NAS research and improve the reproducibility and accessibility of related studies. This dataset covers two state-of-the-art NAS search spaces, NAS-Bench-201 and FBNet, and provides measured/estimated hardware performance data for all networks across six hardware devices, including commercial edge devices, FPGAs, and ASICs. The dataset creation process involves meticulous collection and analysis of hardware performance metrics such as energy cost and latency, to provide in-depth insights for HW-NAS research. Furthermore, HW-NAS-Bench demonstrates how non-hardware experts can conduct HW-NAS by simply querying the pre-measured dataset, and validates that device-specific HW-NAS can achieve optimal accuracy-cost trade-offs. The application scenarios of this dataset include hardware accelerator design and optimization for deep neural networks, aiming to address the problem of efficiently deploying deep neural networks in resource-constrained daily devices.

- 1HW-NAS-Bench:Hardware-Aware Neural Architecture Search Benchmark莱斯大学电气与计算机工程系 · 2021年



