通信学报 ›› 2022, Vol. 43 ›› Issue (4): 114-122.doi: 10.11959/j.issn.1000-436x.2022068

• 学术论文 • 上一篇    下一篇

卷积神经网络的损失最小训练后参数量化方法

张帆1, 黄赟2,3, 方子茁4,4, 郭威1   

  1. 1 国家数字交换系统工程技术研究中心,河南 郑州 450002
    2 信息工程大学,河南 郑州 450001
    3 紫金山实验室,江苏 南京 211111
    4 东南大学网络空间安全学院,江苏 南京 211189
  • 修回日期:2022-03-18 出版日期:2022-04-25 发布日期:2022-04-01
  • 作者简介:张帆(1981- ),男,河南郑州人,博士,国家数字交换系统工程技术研究中心副研究员,主要研究方向为主动防御、人工智能等
    黄赟(1993- ),男,江西新余人,信息工程大学硕士生,主要研究方向为神经网络模型量化压缩、网络内生安全等
    方子茁(1997- ),男,河南郑州人,东南大学硕士生,主要研究方向为网络内生安全、数据库安全、人工智能安全等
    郭威(1990- ),男,北京人,博士,国家数字交换系统工程技术研究中心副研究员,主要研究方向为主动防御、人工智能安全等
  • 基金资助:
    国家自然科学基金资助项目(61521003)

Lost-minimum post-training parameter quantization method for convolutional neural network

Fan ZHANG1, Yun HUANG2,3, Zizhuo FANG4,4, Wei GUO1   

  1. 1 National Digital Switching System Engineering &Technological R&D Center, Zhengzhou 450002, China
    2 Information Engineering University, Zhengzhou 450001, China
    3 Purple Mountain Laboratories, Nanjing 211111, China
    4 School of Cyber Science and Engineering, Southeast University, Nanjing 211189, China
  • Revised:2022-03-18 Online:2022-04-25 Published:2022-04-01
  • Supported by:
    The National Natural Science Foundation of China(61521003)

摘要:

针对数据敏感性场景下模型量化存在数据集不可用的问题,提出了一种不需要使用数据集的模型量化方法。首先,依据批归一化层参数及图像数据分布特性,通过误差最小化方法获得模拟输入数据;然后,通过研究数据舍入特性,提出基于损失最小化的因子动态舍入方法。通过对GhostNet等分类模型及M2Det等目标检测模型进行量化实验,验证了所提量化方法对图像分类及目标检测模型的有效性。实验结果表明,所提量化方法能够使模型大小减少75%左右,在基本保持原有模型准确率的同时有效地降低功耗损失、提高运算效率。

关键词: 卷积神经网络, 批归一化, 模拟输入数据, 动态舍入

Abstract:

To solve the problem that that no dataset is available for model quantization in data-sensitive scenarios, a model quantization method without using data sets was proposed.Firstly, according to the parameters of batch normalized layer and the distribution characteristics of image data, the simulated input data was obtained by error minimization method.Then, by studying the characteristics of data rounding, a factor dynamic rounding method based on loss minimization was proposed.Through quantitative experiments on classification models such as GhostNet and target detection models such as M2Det, the effectiveness of the proposed quantification method for image classification and target detection models was verified.The experimental results show that the proposed quantization method can reduce the model size by about 75%, effectively reduce the power loss and improve the computing efficiency while basically maintaining the accuracy of the original model.

Key words: convolutional neural network, batch normalization, simulation input data, dynamic rounding

中图分类号: 

No Suggested Reading articles found!