作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2020, Vol. 46 ›› Issue (9): 95-100,109. doi: 10.19678/j.issn.1000-3428.0055152

• 人工智能与模式识别 • 上一篇    下一篇

基于深度可分离卷积的轻量级时间卷积网络设计

曹渝昆, 桂丽嫒   

  1. 上海电力大学 计算机科学与技术学院, 上海 200090
  • 收稿日期:2019-06-07 修回日期:2019-08-19 发布日期:2019-09-17
  • 作者简介:曹渝昆(1976-),女,副教授、博士,主研方向为自然语言处理、时间序列预测;桂丽嫒,硕士研究生。
  • 基金资助:
    国家自然科学基金(61702321)。

Design of Lightweight Temporal Convolutional Network Based on Depthwise Separable Convolution

CAO Yukun, GUI Liai   

  1. School of Computer Science and Technology, Shanghai University of Electric Power, Shanghai 200090, China
  • Received:2019-06-07 Revised:2019-08-19 Published:2019-09-17

摘要: 时间卷积网络(TCN)在处理时间序列预测问题时存在计算量大和参数冗余问题,导致其难以应用于存储空间和计算能力受限的手机、平板电脑、笔记本电脑等移动终端。为此,设计一种轻量级时间卷积网络(L-TCN)。采用深度可分离卷积代替TCN中的普通卷积,先通过通道卷积对普通卷积在空间维度上进行分离,以增加网络宽度并扩大特征提取范围,再利用逐点卷积降低普通卷积操作的计算复杂度。实验结果表明,与TCN网络相比,L-TCN在保证时间序列预测精度的同时,能减少网络模型的参数量和计算量,适用于存储空间和计算能力受限的移动终端。

关键词: 时间卷积网络, 深度可分离卷积, 空洞卷积, 因果卷积, 残差网络

Abstract: Application of existing Temporal Convolutional Network(TCN) to temporal sequence prediction results in a large amount of computation and redundant parameters,and therefore it is not applicable to mobile terminals with limited computing capabilities and storage space,including mobiles phones,tablets,and laptops.To address the problem,this paper proposes a Lightweight TCN(L-TCN).The network replaces the common convolution in TCN with depthwise separable convolution.The channel convolution is used to implement separation of common convolution on a spatial dimension,so as to broaden the network and extend the scope of feature extraction.Then the pointwise convolution is used to simplify the computation of common convolution operations.Experimental results show that compared with TCN,the proposed L-TCN can significantly reduce the number of parameters and the amount of calculation of network models while keeping the precision of the temporal sequence prediction,which demonstrates it is applicable to mobile terminals with limited computing capabilities and storage space.

Key words: Temporal Convolutional Network(TCN), depthwise separable convolution, dilated convolution, causal convolution, residual network

中图分类号: