| 刘晨,陆杰,李炼.深度学习程序内存预测方法[J].高技术通讯(中文),2024,34(10):1036~1045 |
| 深度学习程序内存预测方法 |
| Memory estimation for deep learning program |
| |
| DOI：10. 3772 / j. issn. 1002-0470. 2024. 10. 003 |
| 中文关键词: 深度学习; 静态分析; 内存预测 |
| 英文关键词: deep learning, static analysis, memory estimation |
| 基金项目: |
| 作者 | 单位 | | 刘晨 | (中国科学院计算技术研究所处理器芯片全国重点实验室 北京 100190)
(中国科学院大学 北京 100049) | | 陆杰 | | | 李炼 | |
|
| 摘要点击次数: 270 |
| 全文下载次数: 262 |
| 中文摘要: |
| 深度学习程序在广泛领域取得了巨大成功,然而其内部错误可能导致严重的资源浪费,甚至引发灾难性故障。 本文分析了导致程序在实际运行中出现任务执行失败的典型缺陷及其关键影响因素,提出基于静态分析与自注意力机制网络的深度学习程序内存预测方法,在程序内存估计任务上达到平均 8. 38% 的相对预测误差,可以有效预防内存溢出问题、协助合理优化硬件资源配置。 |
| 英文摘要: |
| Deep learning programs have achieved tremendous success in a wide range of fields. However, internal errors
in these programs can lead to significant resource wastage and even result in catastrophic failures. This paper analy-
zes typical defects that cause task execution failures during the practical operation of programs, as well as their key
influencing factors. A deep learning program memory prediction method is proposed based on static analysis and a
self-attention mechanism network. This method achieves a mean relative error of 8. 38% in memory estimation
tasks, effectively preventing memory overflow issues and assisting in the rational optimization of hardware resource
allocation. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
