基于代理模型进化的高效神经网络架构搜索
Efficient Neural Network Architecture Search Based on Proxy Model Evolution
王龙业 1肖舒 1曾晓莉 2王圳鹏3
作者信息
- 1. 西南石油大学电气信息学院,四川 成都 610500
- 2. 西藏大学信息科学技术学院,西藏 拉萨 850000
- 3. 成都市排水有限责任公司,四川 成都 610000
- 折叠
摘要
针对传统人工神经网络设计的时间复杂性与调试困难等问题,研究基于神经网络架构搜索的高效神经网络架构自动搜索方法,解决网络设计方案,提出了一种代理模型进化方法的高效神经网络搜索方法.上述方法将代理模型集成到遗传算法中,通过训练代理模型将预测卷积神经网络(Convolutional Neural Networks,CNN)性能问题转化为二值分类问题,加速网络评估过程.另外,在最新一代模型中,子模型通过继承父模型参数,在训练数据集上训练较少次数,加速最优网络的生成.所提方法在CIfar-10 数据集上分类精度为 95.6%,模型参数为3.4M;在Cifar-100 数据集上分类精度为 77.46%,模型参数为 4.3M.由于使用代理模型和参数共享,新方法在cirar-10 数据集上经过 2.5 天搜索获得性能优异的模型,避免了43.8%的候选CNN训练.
Abstract
Aiming at the problems of time complexity and debugging difficulties in traditional artificial neural net-work design,this paper studies an efficient neural network architecture automatic search method based on neural net-work architecture search,solves the network design scheme,and proposes an efficient neural network search method u-sing proxy model evolution method.In this method,the proxy model was integrated into the genetic algorithm,and the prediction of CNN performance was transformed into a binary classification problem,which can speed up the network evaluation process.A fitness evaluation mechanism was designed to measure the quality and diversity of candidate in-dividuals and retain those with good performance for further training.In addition,in the latest generation of parent model,the sub-model inherits the parameters of the parent model to train fewer times on the training data set and ac-celerate the generation of the optimal network.The classification accuracy of this method on Cifar-10 data set is 95.6%,and the model parameter is 3.4M.The classification accuracy of Cifar-100 data set is 77.46%,and the mod-el parameter is 4.3M.Due to the use of proxy model and parameter inheritance,the method takes only 2.5 days on the CIRAR-10 dataset and avoids 43.8%candidate individual training.
关键词
代理模型/遗传算法/适应度评估/参数继承Key words
Proxy model/Genetic algorithm/Fitness assessment/Parameter inheritance引用本文复制引用
基金项目
国家自然科学基金(61261021)
国家自然科学基金(61561045)
四川省科技计划(2019JDRC0012)
出版年
2024
NETL
NSTL
万方数据