




版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领
文档简介
AutomateMachineLearningwithH2ODriverlessAIonDellInfrastructure
DellValidatedDesignforAI
July2022H19252
WhitePaper
Abstract
Thistechnicalwhitepaperdiscussesthebenefitsofautomatedmachinelearningandthechallengesofnon-automatedmodeldevelopmentthatitovercomes.ThepaperpresentsanoverviewoftheH2ODriverlessAIproductfromH2O.ai,alongwithasolutionarchitectureforH2ODriverlessAIbuiltontheDellValidatedDesignforAI.Italsoprovidesseveralvalidatedusecasesusingthesolution.
DellTechnologiesSolutions
Copyright
Contents
PAGE
2
AutomateMachineLearningwithH2ODriverlessAIonDellInfrastructure
DellValidatedDesignforAIWhitePaper
AutomateMachineLearningwithH2ODriverlessAIonDellInfrastructure
DellValidatedDesignforAI
WhitePaper
PAGE
3
Theinformationinthispublicationisprovidedasis.DellInc.makesnorepresentationsorwarrantiesofanykindwithrespecttotheinformationinthispublication,andspecificallydisclaimsimpliedwarrantiesofmerchantabilityorfitnessforaparticularpurpose.
Use,copying,anddistributionofanysoftwaredescribedinthispublicationrequiresanapplicablesoftwarelicense.Copyright©2022DellInc.oritssubsidiaries.PublishedintheUSA07/22.WhitePaperH19252.
DellInc.believestheinformationinthisdocumentisaccurateasofitspublicationdate.Theinformationissubjecttochangewithoutnotice.
Contents
Introduction 5
Executivesummary 5
Documentpurpose 6
Audience 6
ThechallengesofAIadoption 6
Machinelearningchallenges 6
Talent 6
Time 6
Trust 7
OverviewofAutoMLandH2ODriverlessAI 7
AutoMLworkflowwithH2ODriverlessAI 7
Keyfeatures 10
SolutionarchitectureforAutoML 11
Kubernetes-baseddeploymentusingEnterpriseSteam 11
Dockerimage 12
Security 12
GPUsupport 12
Storageandnetworkconfiguration 13
Licensing 13
InvokingH2ODriverlessAIfromcnvrg.ioMLOpsPlatform 13
AutoMLonanoptimizedDellinfrastructure 15
SizingofAutoMLinfrastructure 16
ValidatedusecasesforAutoML 17
SentimentanalysiswithNLP 17
Imageclassification 20
DellTechnologiesservicesandsupport 21
Deploymentandsupport 21
TheDellTechnologiesCustomerSolutionsCenter 22
Conclusion 22
Wevalueyourfeedback 23
References 24
Contents
Introduction
PAGE
4
AutomateMachineLearningwithH2ODriverlessAIonDellInfrastructure
DellValidatedDesignforAIWhitePaper
AutomateMachineLearningwithH2ODriverlessAIonDellInfrastructure
DellValidatedDesignforAI
WhitePaper
PAGE
5
DellTechnologiesdocumentation 24
H2O.aidocumentation 24
NVIDIAdocumentation 24
AppendixA–Modelservingincnvrg.io 25
Introduction
Executivesummary
Artificialintelligence(AI)andmachinelearninghaverevolutionizedhoworganizationsareusingtheirdata.Automatedmachinelearning(AutoML)facilitatesandimprovestheend-to-enddatascienceprocess.Thisprocessincludeseverythingfrompreprocessingandcleaningthedata,selectingandengineeringappropriatefeatures,tuningandoptimizingthemodel,analyzingresults,explaininganddocumentingthemodel,andofcourse,deployingitintoproduction.
AutoMLacceleratesyourAIinitiativesbyprovidingmethodsandprocessestomakemachinelearningaccessibletobothexpertsandnonexpertsalike.OrganizationslookingtoapplymachinelearningquicklyandaccuratelywithoutemployinglargenumbersofdatascientistscanbenefitfromAutoMLcapabilities.Fororganizationsthathavedatascientists,AutoMLequipsandempowersthemtocreatemorerobustmodelswithaccuracy,speed,andtransparencytodeliverbetterperformanceandoutcomes.Inallcases,AutoMLhelpsorganizationsquicklydiscoverbusinessvaluehiddeninsidetheirdataandeasilyusethatdatatoaddresscomplexproblems.
H2ODriverlessAIisacomprehensiveautomatedmachinelearningproductthatusesAItodoAI,optimizingdatascienceworkflowstoincreaseboththequantityandqualityofdatascienceprojectsdeliveredtobusinessstakeholders.Itempowersdatascientiststoworkonprojectsfasterandmoreefficientlybyusingautomationtoaccomplishkeymachinelearningtasksinminutesorhours,notmonths.
H2ODriverlessAIprovidescapabilitiessuchas:
Exploratorydataanalysis(AutoViz)
Automaticfeatureengineering
Modelbuildingandvalidation
Automaticmodeldocumentation(AutoDoc)
Modelselectionanddeployment
Machinelearninginterpretability(MLI)
AutoMLdoesnotreplacemachinelearningoperations(MLOps).AutoMLfocusesonautomatingandacceleratingthemodeldevelopmentportionoftheMLpipeline,whileMLOpsprovidesanoveralllifecyclemanagementframeworkfordatapreparation,modeldevelopment,andcoding.AutoMLcomplementsMLOpsandcanrunsuccessfullyandefficientlywithvariousMLOpsframeworkssuchascnvrg.io.MLOpsprovidesanoveralllifecyclemanagementframeworkfordatapreparation,modeldevelopment,andcoding.
WithH2ODriverlessAIbring-your-ownrecipes,andtimeseriesandautomaticpipelinegenerationformodelscoring,H2ODriverlessAIprovidescompanieswithanextensibleandcustomizabledatascienceplatformthataddressestheneedsofvarioususecasesforeveryenterpriseineveryindustry.
ThechallengesofAIadoption
OverviewofAutoMLandH2ODriverlessAI
PAGE
6
AutomateMachineLearningwithH2ODriverlessAIonDellInfrastructure
DellValidatedDesignforAIWhitePaper
AutomateMachineLearningwithH2ODriverlessAIonDellInfrastructure
DellValidatedDesignforAI
WhitePaper
PAGE
7
Documentpurpose
Audience
ThiswhitepaperdiscussesAutoML,includingitsbenefitsandthechallengesofmoretraditionalmodeldevelopmentprocessesthatitovercomes.ThewhitepaperprovidesanoverviewoftheH2ODriverlessAIproduct,presentsasolutionarchitectureforH2ODriverlessAIbuiltontheDellValidatedDesignforAIwithVMware,anddescribesseveralvalidatedusecasesusingthesolution.Bydeployingthissolution,datascientistsandITprofessionalscanmovemachinelearningmodelsoutofthelabandintoproductionfasterandmoreeasily,thusbringingabetterreturnoninvestment(ROI)foranorganization’smachinelearninginvestments.
Thiswhitepaperisintendedfordatascientists,solutionarchitects,systemadministrators,andothersdevelopingandsupportingAIandmachinelearningapplications.
ThechallengesofAIadoption
Machinelearningchallenges
AsorganizationsstreamlinedecisionmakingandimprovecustomerexperienceswithAI,theyarerunningintothreecorechallenges:talent,time,andtrust.First,thereisnotenoughdatasciencetalenttobuildmodelsforeveryusecasebyhand.Evenwiththerightpeople,hand-codingtakestoomuchtimeandispronetoerrors.Then,thebusinessmustexplainandvalidateeachmodelsothatuserscantrustthedecisionsthatthemodelsupports.Thekeytobreakingthroughthetalent,time,andtrustbarriersistheautomationofadvancedmachinelearningtechniqueswithH2ODriverlessAI.
Talent
Datascientistsareinshortsupplyforallbutthelargesttechnologycompanies.WithH2ODriverlessAI,bothexpertandnovicedatascientistscanautomaticallybuildhighlyandtransparentaccuratemodelsquickly.H2ODriverlessAIisanaward-winningAutoMLproductthatembedsdatasciencebestpracticesfromtheworld’sleadingexpertsinengineeringanddatascience,includingtheworld’stopKaggleGrandmasters.Itusesauniquegeneticalgorithmthatdeterminesthebestcombinationoffeatures,models,andtuningparametersforeachusecase.Integratedbestpracticesandguardrailsensurethatmodelsdonotoverfitthedataandhelpwithothercommonissueswithwhichnovicedatascientistsmightneedassistance.H2ODriverlessAIenablescompaniestoundertakemoreusecaseswiththetalentthattheyalreadyhaveorcaneasilyfind.
Time
Reducingthetimetodevelopaccurate,production-readymodelsiscriticaltodeliveringAIatscale.H2ODriverlessAIautomatestime-consumingdatasciencetaskssuchasadvancedfeatureengineering,modelselection,hyperparametertuning,modelstacking,andcreationofaneasy-to-deploy,low-latencyscoringpipeline.Withhigh-performancecomputingusingbothCPUsandGPUs,H2ODriverlessAIcomparesthousandsofcombinationsanditerationstofindthebestmodelinminutesorhours.EvenexperienceddatascientistscanuseH2ODriverlessAItoexploremoretechniques,featurecombinations,andtuningparameters.H2ODriverlessAIalsostreamlinesmodeldeploymentthatincludeseverythingneededtorunthemodelinproduction,takingtheprocesstimefromexperimentationtoproductionfrommonthstodays.
Trust
FororganizationstoadoptAIatscale,datateams,businessleaders,andregulatorsmustbeabletoexplain,interpret,andtrustAIresults.H2ODriverlessAIdeliversindustry-leadingcapabilitiesforunderstanding,debugging,andsharingmodelresults,includinganextensivemachinelearninginterpretability(MLI)toolkit,fairnessdashboards,automatedmodeldocumentation,andreasoncodesforeachpredictionforservicerepresentativesandcustomers.WithH2ODriverlessAI,datateamshaveeverythingtheyneedtobuildtrustwithbusinessstakeholdersandregulators.
OverviewofAutoMLandH2ODriverlessAI
H2ODriverlessAIdeliversenterprise-ready,scalable,andsecureAutoMLthatcanrunonanycloudplatformorinon-premisesenvironments,usingthearchitecturethatthisdocumentdescribes.Withanon-premisesenvironment,youdonotneedtomoveyourdatatothecloud;youcanperformAutoMLsecurelywhereveryourdataresides.
H2ODriverlessAIenablesdatascientiststoworkonprojectsfasterandmoreefficientlybyusingautomationtoperformkeymachinelearningtasksinminutesorhours,notmonths.
H2ODriverlessAIincreasestheproductivityofdatapractitionersbyautomatingdataprocessing,featureengineering,modelbuilding,andhyperparametertuning.Itisastand-aloneplatformthatcanbeappliedforusecasessuchasNaturalLanguageProcessing(NLP),timeseriesforecasting,andimageclassifications.EnterprisescanchoosetodeployanMLOPsplatformtoenablecross-functionalcollaborationandtomanagetheend-to-endlifecycleoftheirAIapplications.Inthosecases,userscanintegrateH2ODriverlessAIwiththeirMLOpsplatformsuchascnvrg.io(see
InvokingH2ODriverlessAI
fromcnvrg.ioMLOpsPlatform
).
AutoMLworkflowwith
ThefollowingfigureshowsthestepsinatypicalAutoMLworkflowandhowH2ODriverlessAIenablesthesesteps:
H2ODriverlessAI
Figure1. AutoMLworkflowinH2ODriverlessAI
OverviewofAutoMLandH2ODriverlessAI
OverviewofAutoMLandH2ODriverlessAI
PAGE
10
AutomateMachineLearningwithH2ODriverlessAIonDellInfrastructure
DellValidatedDesignforAIWhitePaper
AutomateMachineLearningwithH2ODriverlessAIonDellInfrastructure
DellValidatedDesignforAI
WhitePaper
PAGE
9
Dataingestion—Theworkflowbeginswiththedata.Dataingestionconsistsofimportingandobtainingdatatoperformanalysisandtraining.
H2ODriverlessAIcaningestdatafromdatasetsinvariousformatsandfilesystemsincludingHadoopHDFS,AmazonS3compatiblestorage,AzureBlobStorage,GoogleBigQuery,GoogleCloudStorage,ApacheHive,JDBC,kdb+,MinIO,Snowflake,DataRecipe,DataRecipeFile,andNFS.ForlargerdatasetsthatarealreadyavailableinPowerScalestorage,H2ODriverlessAIprovidesdataconnectorsforaccessingandingestingdata.
Datapreparation—Whenthedataisdefined,thenextstepisdatapreparation.Thedatasetcanbedividedintotraining,test,andvalidationdatasets.Datascientistscaninteractivelymodelthedataforexploration,analysis,andvisualizationusingdataplotsandstatistics.AutoMLtoolsautomaticallyperformfeatureengineeringbyextractingfeatures(domain-specificattributes)fromrawdataanddatatransformationstosuiteMLalgorithms.
H2ODriverlessAIdeterminesthebestpipelineforadataset,includingautomaticdatatransformationandfeatureengineering.Datascientistscancontrolthenumberoforiginalfeaturesusedinmodelbuildingbyselectingorexcludingcolumnsinthedataset.H2ODriverlessAIusesauniquegeneticalgorithmtoautomaticallyfindnew,high-valuefeaturesandfeaturecombinationsforaspecificdatasetthatarevirtuallyimpossibletofindmanually.Theinterfaceincludesaneasy-to-readvariableimportancechartthatshowsthesignificanceoforiginalandnewlyengineeredfeatures.
Automaticvisualizations(AutoViz)inH2ODriverlessAIproviderobustexploratorydataanalysiscapabilitiesbyautomaticallyselectingdataplotsbasedonthemostrelevantdatastatisticsthatarebasedonthedatashape.Inspecificcases,AutoVizcansuggeststatisticaltransformationforsomedata.Experienceduserscanalsocustomizevisualizationstomeettheirneeds.AutoVizhelpsusersdiscovertrendsandissuessuchaslargenumbersofmissingvaluesorsignificantoutliersthatcanimpactmodelingresults.
Modelbuilding—Whenthedataisprepared,thenextstepismodelbuilding.AutomaticmodelbuildingincludesdatatransformationsandhyperparametertuningforthevariousmodelsavailableintheAutoMLproduct.Itautomaticallytrainsseveralin-builtmodelsandselectsthebestmodelorafinalensembleofmodelsbasedonuser-definedparameterssuchasmodelaccuracy.
AutomaticmodeldevelopmentinH2ODriverlessAIisaccomplishedbyrunningexperiments.H2ODriverlessAItrainsmultiplemodelsandincorporatesmodelhyperparametertuning,scoring,andensembling.Datascientistscanconfigureparameterssuchastheaccuracy,time,lossfunction,andinterpretabilityforaspecificexperiment.Thispreviewisautomaticallyupdatedwhenanyoftheexperiment’ssettingschange(includingtheknobs).Userscanalsorunmultiplediverseexperimentsthatprovideanoverviewofthedataset.Thisfeatureprovidesdatascientistswithrelevantinformationfordeterminingcomplexity,accuracy,size,andtimetradeoffswhenputtingmodelsintoproduction.H2ODriverlessAIusesageneticalgorithmthatincorporatesa‘survivalofthefittest’concepttodeterminethebestmodelforspecificdatasetandconfiguredoptionsautomatically.
Productization—Whentheexperimentiscompleted,youcanmakenewpredictionsandpushthemodelforproduction,eitherinthecloud,on-premises,orattheedge.
H2ODriverlessAIoffersconvenientoptionsfordeployingmachinelearningmodels,dependingonwheretheAIapplicationisrun:
Downloadthemodelandbuildyourowncontainer.
Downloadascoringpipeline.
Whentheexperiment(modelbuildingstep)iscomplete,H2ODriverlessAIcanbuildascoringpipelinethatcanbedeployedtoproduction.Ascoringpipelineisapackagedexperimentwhichincludesartifactsnecessaryformodeldeployment,includingmodelbinary,runtime,readme,example,scripts,andsoon.Youcandownloadtwodifferenttypesofscoringpipelines:
PythonScoringPipeline
MOJOScoringPipeline,whichisavailablewithbothJavaandC++backends
Thedecisionaboutwhichtypeofpipelinetousecomesfromvariousfactorsincludingthetypeofmodelbeingbuiltintheexperiment,usecase,latencyrequirements,andsoon.Ingeneral,MOJOScoringPipelinesarefasterbutmightrequireadditionalsetup,whilePythonScoringPipelinesarebuiltintoa
.whlfile,whicheasilyinstallableinPython.H2ODriverlessAIalsoallowsyoutovisualizethescoringpipelineasadirectionalgraph,asshowninthefollowingfigure:
Figure2. VisualizationofH2ODriverlessAIscoringpipeline
Deploythemodeldirectlyinacloudservice.
ConfigurethemodeltorunonalocalRESTserverwithacoupleofclicks.
Keyfeatures
TheH2ODriverlessAIplatformenablesthefollowingelementsofAutoML:
SupportforNVIDIAGPUs—AImodelsareexplodingincomplexity,andautomateddatatransformationanddeeplearningrequiremassivecomputepowerandscalability.H2ODriverlessAIsupportsthelatestNVIDIAGPUstoacceleratefeatureengineeringandtrainingofneuralnetworks.NVIDIA’sMulti-InstanceGPU(MIG)featurecanbeusedtopartitiontheGPUs,increaseoverallGPUutilization,andsupportseveraltypesofusecasesanddeploymentswithguaranteedqualityofservice.
Integratedcatalogofrecipesandmodels—H2ODriverlessAIoffersarichcatalogofAImodels,transformers,andscorersforautomaticfeatureengineeringandmodelbuilding.
Machinelearninganddeeplearning—H2ODriverlessAIincludesleadingopen-sourcetransformers,embeddings,andframeworksformachinelearninganddeeplearningtechniquestohandlevariousdatascienceusecases.WithH2ODriverlessAI,youcanautomaticallybuildmodelsforIndependentandIdenticallyDistributed(IID)data,images,text,andmore.Forexample,H2ODriverlessAIincludesTensorFlowCNNsforimagemodelingandNLPlibrariesfromPyTorch,includingBERTandotherstate-of-the-arttechniques.
MachineLearningInterpretability(MLI)—H2ODriverlessAIprovidesrobustexplainabilityandfairnessanalysisformachinelearningmodelsandhelpsexploreanddemystifymodelingresults.Itincludesstraightforwarddisparateimpactanalysistotestformodelbiasandprovidesreasoncodesforeveryprediction.Maximumtransparencyandminimaldisparateimpactarecrucialdifferentiatorsifyoumustjustifyyourmodelstobusinessstakeholdersandregulators.
Automaticmodeldocumentation(AutoDoc)—Datascientistsmustdocumentthedata,algorithms,andprocessesusedtocreatemachinelearningmodelsforbusinessusersandregulators.H2ODriverlessAIautomaticmodeldocumentationrelievesyoufromthetime-consumingtaskofrecordingandsummarizingyourworkflowwhilebuildingmachinelearningmodels.Thedocumentationincludesdetailsaboutthedataused,thevalidationschemaselected,modelandfeaturetuning,MLI,andthefinalmodelcreated.AutoDocsavesdatascientiststimeandremovestediousworksothattheycanspendmoretimepracticingdatascienceanddrivemorevalueforthebusiness.
Bring-Your-OwnRecipes—ExperienceddatascientistscaneasilyextendH2ODriverlessAIwithcustomizationsthatrunwithintheH2ODriverlessAIplatform,includingdatapreparation,models,transformers,andscorers.Thesecustomizations,calledrecipes,arePythoncodesnippetsthatcanbeuploadedintoH2ODriverlessAIatruntime,likeplugins.H2ODriverlessAIcanconsumerecipeswithmultipleconvenientoptions:uploadingfromalocalmachine,consumingfrompublishedcodeinasourcecontrolhub(Bitbucket)andlinkingtoareciperawcode.YoucanchecktheGitHubrepositoryfortheavailableandoptimizedH2O.airecipes.Duringtrainingofasupervisedmachinelearningmodelingpipeline,H2ODriverlessAIcanusetheserecipesasbuildingblockswithorinsteadofallintegratedcodepieces.Theyareusedintheautomaticmachinelearningoptimizationprocess,eventuallycreatingthewinningmodel.Datascienceteamscandevelopcustomizationsspecifictotheiruse-cases,industry,orbusiness.
SolutionarchitectureforAutoML
SolutionarchitectureforAutoML
AutomateMachineLearningwithH2ODriverlessAIonDellInfrastructure
DellValidatedDesignforAI
WhitePaper
PAGE
11
PAGE
12
AutomateMachineLearningwithH2ODriverlessAIonDellInfrastructure
DellValidatedDesignforAIWhitePaper
SolutionarchitectureforAutoML
H2ODriverlessAIprovidesanenterprise-readyAutoMLproductfordatascientistsandmachinelearningengineerstodevelopandpublishAIapplications.ItcanbedeployedeitherinKubernetesaspodsorasastand-alonecontainer.
Kubernetes-baseddeploymentusingEnterpriseSteam
EnterpriseSteamfromH2O.aiisaserviceforsecurelymanaginganddeployinginfrastructureforH2ODriverlessAIonKubernetes.EnterpriseSteamofferssecurity,accesscontrol,resourcecontrol,andresourcemonitoringoutoftheboxsothatorganizationscanfocusonthecoreoftheirdatasciencepractice.Itenablessecure,streamlinedadoptionofH2ODriverlessAIandotherH2O.aiproductsthatcomplieswithcompanypolicies.
Fordatascientists,EnterpriseSteamprovidesPython,R,andwebclientsformanagingclustersandinstances.ItallowsdatascientiststopracticedatascienceintheirownH2ODriverlessAIinstance.Foradministrators,EnterpriseSteamcontrolswhichproductversionsandcomputeresourcesareavailable.
EnterprisesteamisasinglepodthatisdeployedusingHelm.WhenEnterpriseSteamisdeployed,youcanlaunchanewH2ODriverlessAIinstanceandmanageexistinginstances.
Youcanuseeachinstanceformodelbuildingforaspecificproject.Inthefollowingfigure,weshowthreeinstancesofH2ODriverlessAIdeployedforautomatedmodelbuildingforthreedifferentusecases:NLP,timeseriesforecasting,andimageclassification.
Figure3. SolutionarchitectureforKubernetes-basedDriverlessAIdeployment
Datasetsaremadeavailabletotheinstanceeitherbydownloadingthemintothecontainerorthroughseveralofthedataconnectors,asexplainedinthefollowingsections.Datavisualization,featureengineering,andmodeldevelopmentareperformed
onthisinstance.H2ODriverlessAIsupportsNVIDIAGPUaccelerationandsomeusecasessuchasimageclassificationcanbenefitfromGPUresources.Fortheseusecases,GPUsareconfiguredandmadeavailabletothecontainer.
Afterthemodelistrained,youcandownloadthePythonorMOJOScoringPipelineandbuildaDockercontainer.YoucandeploythisDockercontaineroutsideoftheKubernetesenvironmentoraspodexposedasaKubernetesservice.
H2ODriverlessAIcanalsobedeployedasastand-alonecontainereitheronbaremetalorvirtualmachines.Thisdeploymentoptionisoutsidethescopeofthisvalidateddesign.Seethe
H2ODriverlessAIdocumentation
formoreinformation.
Dockerimage
Security
GPUsupport
H2ODriverlessAIDockerimagesareavailablethroughEnterpriseSteam.TheDockerimagescomewithalltherequiredlibrariesandsoftwareinstalled,includinglibrariesfortheGPU.
EnterpriseSteamprovidesaccesscontrol.Userscanbecreatedwithdifferentroles,andresourcescanbeallocatedtoeachuser.H2ODriverlessAIsupportsclientcertificate,LDAP,andotherauthenticationoptions.TheseoptionscanbeconfiguredbyspecifyingtheenvironmentvariableswhenstartingtheH2ODriverlessAIDockerimageorbyspecifyingtheappropriateoptionsintheconfigurationfile.Seethe
H2ODriverlessAI
documentation
formoreinformation.
H2ODriverlessAIcanrunonmachineswithonlyCPUsormachineswithCPUsandGPUs.H2ODriverlessAIsupportsNVIDIAA100andA30GPUs.OnlyoneGPUissupportedperinstance.ImageandNLPusecasesinH2ODriverlessAIbenefitsignificantlyfromGPUusage.ModelbuildingalgorithmssuchasXGBoost(GBM/DART/RF/GLM),LightGBM(GBM/DART/RF),PyTorch(BERTmodels),andTensorFlow(CNN/BiGRU/ImageNet)modelsuseGPU.
NVIDIA’sMulti-InstanceGPU(MIG)featurecanbeusedtopartitiontheGPUs,increaseoverallGPUutilization,andsupportseveraltypesofusecasesanddeploymentswithguaranteedqualityofservice.FormoreinformationaboutGPUpartitioningrecommendations,seetothe
NVIDIAMulti-InstanceGPUandNVIDIATechnicalBrief.
ImageandNLPusecasesinH2ODriverlessAIbenefitsignificantlyfromGPUusage.ModelbuildingalgorithmssuchasXGBoost(GBM/DART/RF/GLM),LightGBM(GBM/DART/RF),PyTorch(BERTmodels),andTensorFlow(CNN/BiGRU/ImageNet)modelsuseGPU.
InvokingH2ODriverlessAIfromcnvrg.ioMLOpsPlatform
InvokingH2ODriverlessAIfromcnvrg.ioMLOpsPlatform
AutomateMachineLearningwithH2ODriverlessAIonDellInfrastructure
DellValidatedDesignforAI
WhitePaper
PAGE
13
PAGE
14
AutomateMachineLearningwithH2ODriverlessAIonDellInfrastructure
DellValidatedDesignforAIWhitePaper
Storageandnetworkconfiguration
Licensing
H2ODriverlessAIrequiresnospecialnetworkconsiderations.TheKubernetes-baseddeploymentusesingresscontrolandloadbalancerstogovernaccesstothedeployment.
H2ODriverlessAIusespersistentvolumestosavetherequireddataandtoconnecttoexternaldatasourcessuchasNFS.
H2ODriverlessAIislicensedperuser.EachusercandeployaninstanceofH2ODriverlessAI.H2ODriverlessAImanagestheGPUsinthedeployment.Itensuresthatdifferentexperimentsbydifferentuserscanrunsafelysimultaneouslyanddonotinterferewitheachother.NospeciallicensingisrequiredforGPUsupport.
EnterpriseSteamislicensedseparately.UsersrequireonelicenseperEnterpriseSteamdeployment.
InvokingH2ODriverlessAIfromcnvrg.ioMLOpsPlatform
Asshownin
Figure1,
AutoMLenablesautomaticmodelbuilding.However,itdoesnotofferthecompletelifecycleforamachinelearningapplication.Also,AutoMLautomatedmodelbuildingdoesnotsupportallscenariosandusecases.Forexample,AutoMLsupportstrainingonlyforsuperviseddataandunsupervisedlearning.Itdoesnotsupportreinforcementlearning.
ForbuildingmodelsforsuchcomplexusecasesandtomaintainacompletelifecycleofAImodels,enterprisesrelyonanMLOPsplatform.MLOpsisadefinedprocessandlifecycleformachinelearningdata,models,andcoding.TheMLOpslifecyclebeginswithdataextractionandpreparationasthedatasetismassagedintoastructurethatcaneffectivelyfeedthemodel.MLOpsplatformsprovideconstantmonitoringtoensurethattheprocessisrunningsmoothly.MLOpsenablesdatascientiststobuildcomplexpipelinesthatallowforcontinuouslearning.Automaticretrainingcanbeimplementedtohelpadjustthedeployedprocessandimprovetheaccuracywitheachiteration.
EnterprisesthathavemultipleongoingAIprojectstosupportprogresstowardstheirbusinessintelligencegoalscanusebothMLOpsandAutoMLplatformstotheirrespectivestrengths.DellTechnologieshasworkedcloselywithcnvrg.iotodeliverMLOpsforAIandmachinelearningadoptersthroughajointlyengineeredandtestedsolutiontohelporganizationscapitalizeonthebenefitsofMLOpsformachinelearningandAIworkloads.TheOptimizeMachineLearningThroughMLOpswithDellTechnologiesandcnvrg.io
WhitePaper
and
DesignGuide
provideguidanceforarchitecting,deploying,andoperatingMLOps
温馨提示
- 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
- 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
- 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
- 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
- 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
- 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
- 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。
最新文档
- 班主任班级安全防范与应急处理协议
- 旧村改造项目拆迁补偿与二手房购买合同
- 财务软件研发保密协议及劳动合同模板
- 电玩城免责协议书范本
- 部门副总经理员工晋升与发展规划劳动合同范本
- 残疾人劳动合同签订与劳动争议预防与处理
- 桩基成孔智能监测仪
- 学校缴费流程规范说明
- 快递店店员培训
- 2025消防知识培训
- 休克诊疗指南规范2025
- 2024年辽宁省普通高等学校招生录取普通类本科批(物理学科类)投档最低分
- 保安培训考试内容解析及试题及答案
- 电梯维护保养服务投标文件(技术方案)
- 2025年中国PCR仪市场全面调研及行业投资潜力预测报告
- 2025年医院信息科数据安全管理计划
- 智能工厂整体解决方案
- 2025年四川大学自主招生个人陈述的自我定位
- 2025年福建省建工集团及下属集团招聘235人高频重点提升(共500题)附带答案详解
- 紧急情况的处理措施、预案以及抵抗风险的措施
- 2025中智集团招聘重要岗位高频重点提升(共500题)附带答案详解
评论
0/150
提交评论