tensorflow-tutorials

AboutTensorFlowTensorFlow™isanopensourcesoftwarelibraryfornumericalcomputationusingdataflowgraphs.Nodesinthegraphrepresentmathematicaloperations,whilethegraphedgesrepresentthemultidimensionaldataarrays(tensors)communicatedbetweenthem.TheflexiblearchitectureallowsyoutodeploycomputationtooneormoreCPUsorGPUsinadesktop,server,ormobiledevicewithasingleAPI.TensorFlowwasoriginallydevelopedbyresearchersandengineersworkingontheGoogleBrainTeamwithinGoogle'sMachineIntelligenceresearchorganizationforthepurposesofconductingmachinelearninganddeepneuralnetworksresearch,butthesystemisgeneralenoughtobeapplicableinawidevarietyofotherdomainsaswell.WhatisaDataFlowGraph?Dataflowgraphsdescribemathematicalcomputationwithadirectedgraphofnodes&edges.Nodestypicallyimplementmathematicaloperations,butcanalsorepresentendpointstofeedindata,pushoutresults,orread/writepersistentvariables.Edgesdescribetheinput/outputrelationshipsbetweennodes.Thesedataedgescarrydynamically-sizedmultidimensionaldataarrays,ortensors.TheflowoftensorsthroughthegraphiswhereTensorFlowgetsitsname.Nodesareassignedtocomputationaldevicesandexecuteasynchronouslyandinparallelonceallthetensorsontheirincomingedgesbecomesavailable.TensorFlowFeaturesDeepFlexibilityTensorFlowisn'tarigidneuralnetworkslibrary.Ifyoucanexpressyourcomputationasadataflowgraph,youcanuseTensorFlow.Youconstructthegraph,andyouwritetheinnerloopthatdrivescomputation.Weprovidehelpfultoolstoassemblesubgraphscommoninneuralnetworks,butuserscanwritetheirownhigher-levellibrariesontopofTensorFlow.DefininghandynewcompositionsofoperatorsisaseasyaswritingaPythonfunctionandcostsyounothinginperformance.Andifyoudon'tseethelow-leveldataoperatoryouneed,writeabitofC++toaddanewone.TruePortabilityTensorFlowrunsonCPUsorGPUs,andondesktop,server,ormobilecomputingplatforms.Wanttoplayaroundwithamachinelearningideaonyourlaptopwithoutneedofanyspecialhardware?TensorFlowhasyoucovered.Readytoscale-upandtrainthatmodelfasteronGPUswithnocodechanges?TensorFlowhasyoucovered.Wanttodeploythattrainedmodelonmobileaspartofyourproduct?TensorFlowhasyoucovered.Changedyourmindandwanttorunthemodelasaserviceinthecloud?ContainerizewithDockerandTensorFlowjustworks.ConnectResearchandProductionGonearethedayswhenmovingamachinelearningideafromresearchtoproductrequireamajorrewrite.AtGoogle,researchscientistsexperimentwithnewalgorithmsinTensorFlow,andproductteamsuseTensorFlowtotrainandservemodelslivetorealcustomers.UsingTensorFlowallowsindustrialresearcherstopushideastoproductsfaster,andallowsacademicresearcherstosharecodemoredirectlyandwithgreaterscientificreproducibility.Auto-DifferentiationGradientbasedmachinelearningalgorithmswillbenefitfromTensorFlow'sautomaticdifferentiationcapabilities.AsaTensorFlowuser,youdefinethecomputationalarchitectureofyourpredictivemodel,combinethatwithyourobjectivefunction,andjustadddata--TensorFlowhandlescomputingthederivativesforyou.Computingthederivativeofsomevaluesw.r.t.othervaluesinthemodeljustextendsyourgraph,soyoucanalwaysseeexactlywhat'sgoingon.LanguageOptionsTensorFlowcomeswithaneasytousePythoninterfaceandano-nonsenseC++interfacetobuildandexecuteyourcomputationalgraphs.Writestand-aloneTensorFlowPythonorC++programs,ortrythingsoutinaninteractiveTensorFlowiPythonnotebookwhereyoucankeepnotes,code,andvisualizationslogicallygrouped.Thisisjustthestartthough--we’rehopingtoenticeyoutocontributeSWIGinterfacestoyourfavoritelanguage--beitGo,Java,Lua,JavaScript,orR.MaximizePerformanceWanttouseeveryounceofmuscleinthatworkstationwith32CPUcoresand4GPUcards?Withfirst-classsupportforthreads,queues,andasynchronouscomputation,TensorFlowallowsyoutomakethemostofyouravailablehardware.FreelyassigncomputeelementsofyourTensorFlowgraphtodifferentdevices,andletTensorFlowhandlethecopies.WhoCanUseTensorFlow?TensorFlowisforeveryone.It'sforstudents,researchers,hobbyists,hackers,engineers,developers,inventorsandinnovatorsandisbeingopensourcedundertheApache2.0opensourcelicense.TensorFlowisnotcomplete;itisintendedtobebuiltuponandextended.Wehavemadeaninitialreleaseofthesourcecode,andcontinuetoworkactivelytomakeitbetter.Wehopetobuildanactiveopensourcecommunitythatdrivesthefutureofthislibrary,bothbyprovidingfeedbackandbyactivelycontributingtothesourcecode.WhyDidGoogleOpenSourceThis?IfTensorFlowissogreat,whyopensourceitratherthankeepitproprietary?Theanswerissimplerthanyoumightthink:Webelievethatmachinelearningisakeyingredienttotheinnovativeproductsandtechnologiesofthefuture.Researchinthisareaisglobalandgrowingfast,butlacksstandardtools.Bysharingwhatwebelievetobeoneofthebestmachinelearningtoolboxesintheworld,wehopetocreateanopenstandardforexchangingresearchideasandputtingmachinelearninginproducts.GoogleengineersreallydouseTensorFlowinuser-facingproductsandservices,andourresearchgroupintendstoshareTensorFlowimplementationsalongsidemanyofourresearchpublications.OverviewMNISTForMLBeginnersIfyou'renewtomachinelearning,werecommendstartinghere.You'lllearnaboutaclassicproblem,handwrittendigitclassification(MNIST),andgetagentleintroductiontomulticlassclassific