Designing Machine Learning Systems_ An Iterative Process for Production-Ready Applications.pdf

Designing Machine Learning Systems_ An Iterative Process for Production-Ready Applications.pdf

《Designing Machine Learning Systems_ An Iterative Process for Production-Ready Applications.pdf》由会员分享，可在线阅读，更多相关《Designing Machine Learning Systems_ An Iterative Process for Production-Ready Applications.pdf（389页珍藏版）》请在冰点文库上搜索。

ChipHuyenDesigningMachineLearningSystemsAnIterativeProcessforProduction-ReadyApplicationsMACHINELEARNING“Thisis,simply,theverybestbookyoucanreadabouthowtobuild,deploy,andscalemachinelearningmodelsatacompanyformaximumimpact.”JoshWillsSoftwareEngineeratWeaveGridandformerDirectorofDataEngineering,Slack“Inabloomingbutchaoticecosystem,thisprincipledviewonend-to-endMLisbothyourmapandyourcompass:

amust-readforpractitionersinsideandoutsideofBigTech.”JacopoTagliabueDirectorofAI,CoveoDesigningMachineLearningSystemsUS$59.99CAN$74.99ISBN:

978-1-098-10796-3Twitter:

Machinelearningsystemsarebothcomplexandunique.Complexbecausetheyconsistofmanydifferentcomponentsandinvolvemanydifferentstakeholders.Uniquebecausetheyredatadependent,withdatavaryingwildlyfromoneusecasetothenext.Inthisbook,youlllearnaholisticapproachtodesigningMLsystemsthatarereliable,scalable,maintainable,andadaptivetochangingenvironmentsandbusinessrequirements.AuthorChipHuyen,co-founderofClaypotAI,considerseachdesigndecisionsuchashowtoprocessandcreatetrainingdata,whichfeaturestouse,howoftentoretrainmodels,andwhattomonitorinthecontextofhowitcanhelpyoursystemasawholeachieveitsobjectives.Theiterativeframeworkinthisbookusesactualcasestudiesbackedbyamplereferences.Thisbookwillhelpyoutacklescenariossuchas:

EngineeringdataandchoosingtherightmetricstosolveabusinessproblemAutomatingtheprocessforcontinuallydeveloping,evaluating,deploying,andupdatingmodelsDevelopingamonitoringsystemtoquicklydetectandaddressissuesyourmodelsmightencounterinproductionArchitectinganMLplatformthatservesacrossusecasesDevelopingresponsibleMLsystemsChipHuyenisco-founderofClaypotAI,aplatformforreal-timemachinelearning.ThroughherworkatNVIDIA,Netflix,andSnorkelAI,sheshelpedsomeoftheworldslargestorganizationsdevelopanddeployMLsystems.ChipbasedthisbookonherlecturesforCS329S:

MachineLearningSystemsDesign,acoursesheteachesatStanfordUniversity.PraiseforDesigningMachineLearningSystemsThereissomuchinformationoneneedstoknowtobeaneffectivemachinelearningengineer.Itshardtocutthroughthechafftogetthemostrelevantinformation,butChiphasdonethatadmirablywiththisbook.IfyouareseriousaboutMLinproduction,andcareabouthowtodesignandimplementMLsystemsendtoend,thisbookisessential.LaurenceMoroney,AIandMLLead,GoogleOneofthebestresourcesthatfocusesonthefirstprinciplesbehinddesigningMLsystemsforproduction.Amust-readtonavigatetheephemerallandscapeoftoolingandplatformoptions.GokuMohandas,FounderofMadeWithMLChipsmanualisthebookwedeserveandtheoneweneedrightnow.Inabloomingbutchaoticecosystem,thisprincipledviewonend-to-endMLisbothyourmapandyourcompass:

amust-readforpractitionersinsideandoutsideofBigTechespeciallythoseworkingat“reasonablescale.”Thisbookwillalsoappealtodataleaderslookingforbestpracticesonhowtodeploy,manage,andmonitorsystemsinthewild.JacopoTagliabue,DirectorofAI,Coveo;Adj.ProfessorofMLSys,NYUThisis,simply,theverybestbookyoucanreadabouthowtobuild,deploy,andscalemachinelearningmodelsatacompanyformaximumimpact.Chipisamasterfulteacher,andthebreadthanddepthofherknowledgeisunparalleled.JoshWills,SoftwareEngineeratWeaveGridandformerDirectorofDataEngineering,SlackThisisthebookIwishIhadreadwhenIstartedasanMLengineer.ShreyaShankar,MLOpsPhDStudentDesigningMachineLearningSystemsisawelcomeadditiontothefieldofappliedmachinelearning.Thebookprovidesadetailedguideforpeoplebuildingend-to-endmachinelearningsystems.ChipHuyenwritesfromherextensive,hands-onexperiencebuildingreal-worldmachinelearningapplications.BrianSpiering,DataScienceInstructoratMetisChipistrulyaworld-classexpertonmachinelearningsystems,aswellasabrilliantwriter.Bothareevidentinthisbook,whichisafantasticresourceforanyonelookingtolearnaboutthistopic.AndreyKurenkov,PhDCandidateattheStanfordAILabChipHuyenhasproducedanimportantadditiontothecanonofmachinelearningliteratureonethatisdeeplyliterateinMLfundamentals,buthasamuchmoreconcreteandpracticalapproachthanmost.Thefocusonbusinessrequirementsaloneisuncommonandvaluable.ThisbookwillresonatewithengineersgettingstartedwithMLandwithothersinanypartoftheorganizationtryingtounderstandhowMLworks.ToddUnderwood,SeniorEngineeringDirectorforMLSRE,Google,andCoauthorofReliableMachineLearningChipHuyenDesigningMachineLearningSystemsAnIterativeProcessforProduction-ReadyApplicationsBostonFarnhamSebastopolTokyoBeijingBostonFarnhamSebastopolTokyoBeijing978-1-098-10796-3LSIDesigningMachineLearningSystemsbyChipHuyenCopyright2022HuyenThiKhanhNguyen.Allrightsreserved.PrintedintheUnitedStatesofAmerica.PublishedbyOReillyMedia,Inc.,1005GravensteinHighwayNorth,Sebastopol,CA95472.OReillybooksmaybepurchasedforeducational,business,orsalespromotionaluse.Onlineeditionsarealsoavailableformosttitles（http:

/）.Formoreinformation,contactourcorporate/institutionalsalesdepartment:

800-998-9938or.AcquisitionsEditor:

NicoleButterfieldDevelopmentEditor:

JillLeonardProductionEditor:

GregoryHymanCopyeditor:

nSight,Inc.Proofreader:

PiperEditorialConsulting,LLCIndexer:

nSight,Inc.InteriorDesigner:

DavidFutatoCoverDesigner:

KarenMontgomeryIllustrator:

KateDulleaMay2022:

FirstEditionRevisionHistoryfortheFirstEdition2022-05-17:

FirstReleaseSeehttp:

/forreleasedetails.TheOReillylogoisaregisteredtrademarkofOReillyMedia,Inc.DesigningMachineLearningSystems,thecoverimage,andrelatedtradedressaretrademarksofOReillyMedia,Inc.Theviewsexpressedinthisworkarethoseoftheauthor,anddonotrepresentthepublishersviews.Whilethepublisherandtheauthorhaveusedgoodfaitheffortstoensurethattheinformationandinstructionscontainedinthisworkareaccurate,thepublisherandtheauthordisclaimallresponsibilityforerrorsoromissions,includingwithoutlimitationresponsibilityfordamagesresultingfromtheuseoforrelianceonthiswork.Useoftheinformationandinstructionscontainedinthisworkisatyourownrisk.Ifanycodesamplesorothertechnologythisworkcontainsordescribesissubjecttoopensourcelicensesortheintellectualpropertyrightsofothers,itisyourresponsibilitytoensurethatyourusethereofcomplieswithsuchlicensesand/orrights.TableofContentsPreface.ix1.OverviewofMachineLearningSystems.1WhentoUseMachineLearning3MachineLearningUseCases9UnderstandingMachineLearningSystems12MachineLearninginResearchVersusinProduction12MachineLearningSystemsVersusTraditionalSoftware22Summary232.IntroductiontoMachineLearningSystemsDesign.25BusinessandMLObjectives26RequirementsforMLSystems29Reliability29Scalability30Maintainability31Adaptability31IterativeProcess32FramingMLProblems35TypesofMLTasks36ObjectiveFunctions40MindVersusData43Summary463.DataEngineeringFundamentals.49DataSources50DataFormats53JSON54iiiRow-MajorVersusColumn-MajorFormat54TextVersusBinaryFormat57DataModels58RelationalModel59NoSQL63StructuredVersusUnstructuredData66DataStorageEnginesandProcessing67TransactionalandAnalyticalProcessing67ETL:

Extract,Transform,andLoad70ModesofDataflow72DataPassingThroughDatabases72DataPassingThroughServices73DataPassingThroughReal-TimeTransport74BatchProcessingVersusStreamProcessing78Summary794.TrainingData.81Sampling82NonprobabilitySampling83SimpleRandomSampling84StratifiedSampling84WeightedSampling85ReservoirSampling86ImportanceSampling87Labeling88HandLabels88NaturalLabels91HandlingtheLackofLabels94ClassImbalance102ChallengesofClassImbalance103HandlingClassImbalance105DataAugmentation113SimpleLabel-PreservingTransformations114Perturbation114DataSynthesis116Summary1185.FeatureEngineering.119LearnedFeaturesVersusEngineeredFeatures120CommonFeatureEngineeringOperations123HandlingMissingValues123Scaling126iv|TableofContentsDiscretization128EncodingCategoricalFeatures129FeatureCrossing132DiscreteandContinuousPositionalEmbeddings133DataLeakage135CommonCausesforDataLeakage137DetectingDataLeakage140EngineeringGoodFeatures141FeatureImportance142FeatureGeneralization144Summary1466.ModelDevelopmentandOfflineEvaluation.149ModelDevelopmentandTraining150EvaluatingMLModels150Ensembles156ExperimentTrackingandVersioning162DistributedTraining168AutoML172ModelOfflineEvaluation178Baselines179EvaluationMethods181Summary1887.ModelDeploymentandPredictionService.191MachineLearningDeploymentMyths194Myth1:

YouOnlyDeployOneorTwoMLModelsataTime194Myth2:

IfWeDontDoAnything,ModelPerformanceRemainstheSame195Myth3:

YouWontNeedtoUpdateYourModelsasMuch196Myth4:

MostMLEngineersDontNeedtoWorryAboutScale196BatchPredictionVersusOnlinePrediction197FromBatchPredictiontoOnlinePrediction201UnifyingBatchPipelineandStreamingPipeline203ModelCompression206Low-RankFactorization206KnowledgeDistillation208Pruning208Quantization209MLontheCloudandontheEdge212CompilingandOptimizingModelsforEdgeDevices214MLinBrowsers222Summary223TableofContents|v8.DataDistributionShiftsandMonitoring.225CausesofMLSystemFailures226SoftwareSystemFailures227ML-SpecificFailures229DataDistributionShifts237TypesofDataDistributionShifts237GeneralDataDistributionShifts241DetectingDataDistributionShifts242AddressingDataDistributionShifts248MonitoringandObservability