1. 您所在位置：
  2. 网站首页
  3. 文档分类
  4. 计算机
  5. 数据结构与算法

基于Winograd算法卷积神网络加速方法研究.pdfVIP

  • 2
  • 0
  • 约12.76万字
  • 约 26页
  • 2025-10-22 发布于北京
  • 举报

基于Winograd算法卷积神网络加速方法研究.pdf

    FastAlgorithmsforConvolutionalNeuralNetworks

    AndrewLavinScottGray

    alavin@NervanaSystems

    sgray@

    causetheyachieveteraccuracywithfewerweightsthan

    shallownetworkswithlargerfilters[12,8].

    DeepconvolutionalneuralnetworkstakeGPU-daysofThereforethereisastrongneedforfastconvnetalgo-

    computationtotrainonlargedatasets.Pedestriandetec-rithmsforsmallbatchsizesandsmallfilters.However

    tionforselfdrivingcarsrequiresverylowlatency.Imageconventionalconvnetlibrariesrequirelargebatchsizesand

    recognitionformobilephonesisconstrainedbylimitedpro-largefiltersforfastoperation.

    cessingresources.Thesuccessofconvolutionalneuralnet-

    Thispaperintroducesanewclassoffastalgorithmsfor

    worksinthesesituationsislimitedbyhowfastwecancom-

    convolutionalneuralnetworksbasedontheminimalfilter-

    putethem.ConventionalFFTbasedconvolutionisfastfor

    ingalgorithmsdiscovere

    您可能关注的文档

    最近下载

    文档评论（0）

    1亿VIP精品文档

    更多 >

    相关文档

    更多 >