Greedy Pruning Algorithm for DETR Architecture Networks Based on Global Optimization  

基于全局优化的DETR架构网络贪心剪枝算法

作　　者：HUANG Qiubo XU Jingsai ZHANG Yakui WANG Mei CHEN Dehua 黄秋波;徐敬赛;张亚魁;王梅;陈德华(东华大学计算机科学与技术学院,上海201620)

机构地区：[1]School of Computer Science and Technology,Donghua University,Shanghai 201620,China

出　　处：《Journal of Donghua University(English Edition)》2025年第1期96-105,共10页东华大学学报(英文版)

基　　金：Shanghai Municipal Commission of Economy and Information Technology,China(No.202301054)。

摘　　要：End-to-end object detection Transformer(DETR)successfully established the paradigm of the Transformer architecture in the field of object detection.Its end-to-end detection process and the idea of set prediction have become one of the hottest network architectures in recent years.There has been an abundance of work improving upon DETR.However,DETR and its variants require a substantial amount of memory resources and computational costs,and the vast number of parameters in these networks is unfavorable for model deployment.To address this issue,a greedy pruning(GP)algorithm is proposed,applied to a variant denoising-DETR(DN-DETR),which can eliminate redundant parameters in the Transformer architecture of DN-DETR.Considering the different roles of the multi-head attention(MHA)module and the feed-forward network(FFN)module in the Transformer architecture,a modular greedy pruning(MGP)algorithm is proposed.This algorithm separates the two modules and applies their respective optimal strategies and parameters.The effectiveness of the proposed algorithm is validated on the COCO 2017 dataset.The model obtained through the MGP algorithm reduces the parameters by 49%and the number of floating point operations(FLOPs)by 44%compared to the Transformer architecture of DN-DETR.At the same time,the mean average precision(mAP)of the model increases from 44.1%to 45.3%.目标检测Transformer(detection Transformer,DETR)成功地定义了Transformer架构在目标检测领域的范式,其端对端检测的流程和集合预测的思想成为近些年最热门的网络架构之一。对DETR的改进工作层出不穷。然而,DETR及其变种需较多的内存资源和计算消耗,这些网络巨大的参数量对于模型的部署十分不利。为解决该问题提出一种贪心剪枝算法,运用于一个变种DN-DETR上,可剔除DN-DETR中Transformer架构的冗余参数。考虑到Transformer架构中多头注意力(multi-head attention,MHA)模块和前馈网络(feed-forward network,FFN)模块的不同作用,进一步提出一种分模块贪心剪枝算法,将两个模块分开考虑,应用各自最优的策略与参数。在COCO 2017数据集上验证了所提方法的有效性。通过分模块贪心剪枝算法得到的模型,对比DN-DETR的Transformer架构,参数量减少了49%,浮点运算次数减少了44%,同时模型的平均精度由44.1%提高到45.3%。

关 键 词：model pruning object detection Transformer(DETR) Transformer architecture object detection 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TP301.6[自动化与计算机技术—计算机科学与技术]