基于改进自适应遗传算法的结构化剪枝方法
Improved Adaptive Genetic-Algorithm based Structured Pruning Method
袁沁 1孙闽红 1滕旭阳 1朱万乾1
作者信息
- 1. 杭州电子科技大学通信工程学院,浙江 杭州 310018
- 折叠
摘要
网络剪枝是当前深度神经网络模型压缩的主要方法之一.针对现有的基于遗传算法的网络剪枝方法效率较低的问题,提出了一种基于改进自适应遗传算法的结构化剪枝方法.首先设计了新的适应度函数,平衡了模型损失值和参数量对最终结果的影响;其次,利用自适应的交叉概率和变异概率来代替固定的超参数,提高了剪枝效率和模型准确率;最后,通过实验验证了方法的可行性,得到了精度更高、参数量更少的网络模型.
Abstract
Network pruning is one of the main methods for compressing deep neural network models.To address the issue of low efficiency in existing genetic algorithm-based network pruning methods,we propose a structured pruning method based on an improved adaptive genetic algorithm.Firstly,we design a new fitness function that balances the impact of model loss and parameter quantity of the final result.Secondly,adaptive crossover and mutation probabilities are used instead of fixed hyperparameters to improve pruning efficiency and model accuracy.Finally,the feasibility of the method is verified through experiments,obtaining network models with higher accuracy and fewer parameters.
关键词
深度学习/模型压缩/网络剪枝/遗传算法Key words
deep learning/model compression/network pruning/genetic algorithm引用本文复制引用
出版年
2024
NETL
NSTL
万方数据