CopyRight 2012-2014 DS文库版权所有
ShuffleNet An Extremely Efficient Convolutional Neural Network for Mobile Devices
(0 次评价)2175 人阅读0 次下载

 ShuffleNet:AnExtremelyEfficientConvolutionalNeuralNetworkforMobileDevices XiangyuZhang ∗ XinyuZhou ∗ MengxiaoLinJianSunMegviiInc(Face++) { zhangxiangyu,zxy,linmengxiao,sunjian } @megvii.com Abstract Weintroduceanextremelycomputation-efficientCNNarchitecturenamedShuffleNet,whichisdesignedspeciallyformobiledeviceswithverylimitedcomputingpower(e.g.,10-150MFLOPs).Thenewarchitectureutilizestwonewoperations,pointwisegroupconvolutionandchannelshuf-fle,togreatlyreducecomputationcostwhilemaintainingaccuracy.ExperimentsonImageNetclassificationandMSCOCOobjectdetectiondemonstratethesuperiorperfor-manceofShuffleNetoverotherstructures,e.g.lowertop-1error(absolute7.8%)thanrecentMobileNet[ 12 ]onIma-geNetclassificationtask,underthecomputationbudgetof40MFLOPs.OnanARM-basedmobiledevice,ShuffleNetachieves ∼ 13 × actualspeedupoverAlexNetwhilemain-tainingcomparableaccuracy. 1.Introduction Buildingdeeperandlargerconvolutionalneuralnet-works(CNNs)isaprimarytrendforsolvingmajorvisualrecognitiontasks[ 21 , 9 , 33 , 5 , 28 , 24 ].Themostaccu-rateCNNsusuallyhavehundredsoflayersandthousandsofchannels[ 9 , 34 , 32 , 40 ],thusrequiringcomputationatbillionsofFLOPs.Thisreportexaminestheoppositeex-treme:pursuingthebestaccuracyinverylimitedcompu-tationalbudgetsattensorhundredsofMFLOPs,focusingoncommonmobileplatformssuchasdrones,robots,andsmartphones.Notethatmanyexistingworks[ 16 , 22 , 43 , 42 , 38 , 27 ]focusonpruning,compressing,orlow-bitrepresent-inga“basic”networkarchitecture.Hereweaimtoexploreahighlyefficientbasicarchitecturespeciallydesignedforourdesiredcomputingranges.Wenoticethatstate-of-the-artbasicarchitecturessuchas Xception [ 3 ]and ResNeXt [ 40 ]becomelessefficientinex-tremelysmallnetworksbecauseofthecostlydense 1 × 1 convolutions.Weproposeusing pointwisegroupconvolu- *Equallycontribution. tions toreducecomputationcomplexityof 1 × 1 convolu-tions.Toovercomethesideeffectsbroughtbygroupcon-volutions,wecomeupwithanovel channelshuffle opera-tiontohelptheinformationflowingacrossfeaturechannels.Basedonthetwotechniques,webuildahighlyefficientar-chitecturecalled ShuffleNet .Comparedwithpopularstruc-tureslike[ 30 , 9 , 40 ],foragivencomputationcomplexitybudget,ourShuffleNetallowsmorefeaturemapchannels,whichhelpstoencodemoreinformationandisespeciallycriticaltotheperformanceofverysmallnetworks.WeevaluateourmodelsonthechallengingImageNetclassification[ 4 , 29 ]andMSCOCOobjectdetection[ 23 ]tasks.Aseriesofcontrolledexperimentsshowstheeffec-tivenessofourdesignprinciplesandthebetterperformanceoverotherstructures.Comparedwiththestate-of-the-artarchitecture MobileNet [ 12 ],ShuffleNetachievessuperiorperformancebyasignificantmargin,e.g.absolute7.8%lowerImageNettop-1erroratlevelof40MFLOPs.Wealsoexaminethespeeduponrealhardware,i.e.anoff-the-shelfARM-basedcomputingcore.TheShuffleNetmodelachieves ∼ 13 × actual speedup(theoreticalspeedupis18 × )overAlexNet[ 21 ]whilemaintainingcomparableaccuracy. 2.RelatedWork EfficientModelDesigns Thelastfewyearshaveseenthesuccessofdeepneuralnetworksincomputervisiontasks[ 21 , 36 , 28 ],inwhichmodeldesignsplayanim-portantrole.Theincreasingneedsofrunninghighqual-itydeepneuralnetworksonembeddeddevicesencour-agethestudyonefficientmodeldesigns[ 8 ].Forex-ample, GoogLeNet [ 33 ]increasesthedepthofnetworkswithmuchlowercomplexitycomparedtosimplystack-ingconvolutionlayers. SqueezeNet [ 14 ]reducesparame-tersandcomputationsignificantlywhilemaintainingaccu-racy. ResNet [ 9 , 10 ]utilizestheefficientbottleneckstruc-turetoachieveimpressiveperformance. SENet [ 13 ]in-troducesanarchitecturalunitthatboostsperformanceatslightcomputationcost.Concurrentwithus,averyre- 1 arXiv:1707.01083v2 [cs.CV] 7 Dec 2017

打分:

0 星

用户评论:

小飞飞
于 2020-07-15 上传

版权及免责声明|RISC-V单片机中文网 |网站地图

GMT+8, 2022-10-5 06:41 , Processed in 44.115301 second(s), 30 queries .

返回顶部