A PPROX C ALIPER : A P ROGRAMMABLE F RAMEWORK FOR A PPLICATION - AWARE N EURAL N ETWORK O PTIMIZATION
A PPROX C ALIPER : A P ROGRAMMABLE F RAMEWORK FOR A PPLICATION - AWARE N EURAL N ETWORK O PTIMIZATION
复制标题
DOI：
--
复制
发表时间：
2023
期刊：
影响因子：
--
通讯作者：
Yifan Zhao;Hashim Sharif;Peter Pao-Huang;Vatsin Ninad Shah;A. N. Sivakumar;M. V. Gasparino;Abdulrahman Mahmoud;Nathan Zhao;S. Adve;Girish V. Chowdhary;Sasa Misailovic;Vikram S. Adve
中科院分区：
文献类型：
--
作者：
Yifan Zhao;Hashim Sharif;Peter Pao-Huang;Vatsin Ninad Shah;A. N. Sivakumar;M. V. Gasparino;Abdulrahman Mahmoud;Nathan Zhao;S. Adve;Girish V. Chowdhary;Sasa Misailovic;Vikram S. Adve
为了在资源受限的边缘系统上部署计算密集型神经网络，开发人员使用模型优化技术来减少模型大小和计算成本。现有的优化工具与应用无关-它们仅根据神经网络的准确性来优化模型参数-因此可能错过优化机会。我们提出了ApproxCaliper，这是第一个用于应用感知神经网络优化的可编程框架。通过结合应用程序特定的目标，与应用程序无关的技术相比，ApproxCaliper有助于更积极地优化神经网络。我们在两个真实世界的机器人系统中使用的五种不同的神经网络上进行实验：商业农业机器人和自动电动车的模拟。与学习率回退（LRR）相比，这是一种用于应用程序不可知设置的最先进的结构化修剪工具，ApproxCaliper实现了5.3倍的加速比和2.9倍的GPU资源利用率，以及36倍和6.1倍的额外模型大小减少。
To deploy compute-intensive neural networks on resource-constrained edge systems, developers use model optimization techniques that reduce model size and computational cost. Existing optimization tools are application-agnostic – they optimize model parameters solely in view of the neural network accuracy – and can thus miss optimization opportunities. We propose ApproxCaliper , the first programmable framework for application-aware neural network optimization . By incorporating application-specific goals, ApproxCaliper facilitates more aggressive optimization of the neural networks compared to application-agnostic techniques. We perform experiments on five different neural networks used in two real-world robotics systems: a commercial agriculture robot and a simulation of an autonomous electric cart. Compared to Learning Rate Rewinding (LRR), a state-of-the-art structured pruning tool used in an application agnostic setting, ApproxCaliper achieves 5.3 × higher speedup and 2.9 × lower GPU resource utilization, and 36 × and 6.1 × additional model size reduction for the two evaluated benchmarks, respectively.