太赫兹科学与电子信息学报2025，Vol.23Issue(10)：1051-1059,9.DOI:10.11805/TKYDA2024125
基于原位计算硬件架构的负载均衡剪枝
Load balancing pruning based on in situ computing hardware architecture
岑华峰 1贺迅 1刘友江 1杨春1
作者信息
- 1. 中国工程物理研究院 电子工程研究所,四川 绵阳 621999
- 折叠
摘要
Abstract
Deep neural networks are widely used in the field of computer vision,but their huge storage and computational costs hinder their further development in embedded environments.Pruning is an algorithm that can reduce the computational complexity of neural networks,but traditional pruning algorithms lack consideration for hardware architecture,resulting in mismatch and load imbalance issues,and low utilization of hardware resources.The new type of hardware architecture pruning often limits the pruning style to a certain range,and its accuracy performance is not as good as traditional algorithms.This article proposes a new load balancing pruning algorithm that can better utilize hardware resources,and compared to the previous pruning technology considering hardware architecture,it has better accuracy representation.Under the premise of less than 1%accuracy degradation,the algorithm achieves nearly 3.3×speedup on both YOLOv3-Tiny and VGG16 networks.关键词
剪枝/负载均衡/模型准确率/模型推理时间Key words
prune/load balance/model accuracy/model inference time分类
计算机与自动化引用本文复制引用
岑华峰,贺迅,刘友江,杨春..基于原位计算硬件架构的负载均衡剪枝[J].太赫兹科学与电子信息学报,2025,23(10):1051-1059,9.
