Computer Engineering and Applications ›› 2025, Vol. 61 ›› Issue (6): 210-219.DOI: 10.3778/j.issn.1002-8331.2311-0002

• Pattern Recognition and Artificial Intelligence • Previous Articles     Next Articles

Fusion of Deep Reinforcement Learning in Joint Compression Method for Convolutional Neural Network

MA Zuxin, CUI Yunhe, QIN Yongbin, SHEN Guowei, GUO Chun, CHEN Yi, QIAN Qing   

  1. 1.Engineering Research Center of Ministry of Education for Text Computing and Cognitive Intelligence, School of Computer Science and Technology, Guizhou University, Guiyang 550025, China
    2.State Key Laboratory of Public Big Data, Guizhou University, Guiyang 550025, China
    3.Provincial Key Laboratory of Software Engineering and Information Security, Guizhou University, Guiyang 550025, China
    4.School of Information, Guizhou University of Finance and Economics, Guiyang 550025, China
  • Online:2025-03-15 Published:2025-03-14

融合深度强化学习的卷积神经网络联合压缩方法

马祖鑫,崔允贺,秦永彬,申国伟,郭春,陈意,钱清   

  1. 1.贵州大学 计算机科学与技术学院 文本计算与认知智能教育部工程研究中心,贵阳 550025
    2.贵州大学 公共大数据国家重点实验室,贵阳 550025
    3.贵州大学 贵州省软件工程与信息安全特色重点实验室,贵阳 550025
    4.贵州财经大学 信息学院,贵阳 550025

Abstract: With the rise of concepts such as edge computing and edge intelligence, the lightweight deployment of convolutional neural network has gradually become a research hotspot. The traditional convolutional neural network compression technique usually performs pruning and quantization strategies in stages and independently, but this method does not consider the interaction between pruning and quantification processes, so that it cannot achieve the optimal pruning and quantification results, which affects the performance of the compressed model. In order to solve the above problems, this paper proposes CoTrim, a joint compression method for neural networks based on deep reinforcement learning. CoTrim performs channel pruning and weight quantization at the same time, and uses the deep reinforcement learning algorithm to search for the global optimal pruning and quantization strategy to balance the impact of pruning and quantization on network performance. Experiments are conducted on the CIFAR-10 dataset on VGG and ResNet, and the experimental results show that for common single branch convolution and residual convolution structures, CoTrim is able to compress the model size of VGG16 to the original 1.41% with a precision loss of only 2.49?percentage points. Experiments are conducted on compact networks MobileNet and DenseNet on a complex dataset Imagenet-1K. The experimental results show that for deep separable convolutional structures and densely connected structures, CoTrim can still ensure accuracy loss within an acceptable range and achieve from 1/5 to 1/8 model size compression.

Key words: convolutional neural network, deep reinforcement learning, model compression, channel pruning, weight quantization, edge intelligence

摘要: 随着边缘计算、边缘智能等概念的兴起,卷积神经网络的轻量化部署逐渐成为研究热点。传统的卷积神经网络压缩技术通常分阶段地、独立地执行剪枝与量化策略,但这种方式没有考虑剪枝与量化过程的相互影响,使其无法达到最优的剪枝与量化结果,影响压缩后的模型性能。针对以上问题,提出一种基于深度强化学习的神经网络联合压缩方法——CoTrim。CoTrim同时执行通道剪枝与权值量化,利用深度强化学习算法搜索出全局最优的剪枝与量化策略,以平衡剪枝与量化对网络性能的影响。在CIFAR-10数据集上对VGG和ResNet进行实验,实验表明,对于常见的单分支卷积和残差卷积结构,CoTrim能够在精度损失仅为2.49个百分点的情况下,将VGG16的模型大小压缩至原来的1.41%。在复杂数据集Imagenet-1K上对紧凑网络MobileNet和密集连接网络DenseNet进行实验,实验表明,对于深度可分离卷积结构以及密集连接结构,CoTrim依旧能保证精度损失在可接受范围内将模型压缩为原始大小的1/5~1/8。

关键词: 卷积神经网络, 深度强化学习, 模型压缩, 通道剪枝, 权值量化, 边缘智能