PracticalDeepLearigforCoders
DeepLearigforCoderswithfastaiadPyTorch:AIApplicatiosWithoutaPhD-thebookadthecourse
WelcometoPracticalDeepLearigforCoders.Thiswebsitecoversthebookadthe2020versioofthecourse,whicharedesigedtoworkcloselytogether.Ifyouhave'tyetgotthebook,youcabuyithere.It'salsofreelyavailableasiteractiveJupyterNotebooks;readotolearhowtoaccessthem..
HowdoIgetstarted?Ifyou'rereadytodiveirightow,here'showtogetstarted.Ifyouwattokowmoreaboutthiscourse,readtheextsectios,adthecomebackhere.
Towatchthevideos,clickotheLessossectioitheavigatiosidebar.Thelessosallhavesearchabletrascripts;click"TrascriptSearch"ithetoprightpaeltosearchforawordorphrase,adtheclickittojumpstraighttovideoatthetimethatappearsithetrascript.ThevideosareallcaptioedadalsotraslateditoChiese(简体中文)adSpaish;whilewatchigthevideoclickthe"CC"buttototurthemoadoff,adthesettigbuttotochagethelaguage.
Eachvideocoversachapterfromthebook.TheetiretyofeverychapterofthebookisavailableasaiteractiveJupyterNotebook.JupyterNotebookisthemostpopulartoolfordoigdatascieceiPytho,forgoodreaso.Itispowerful,flexible,adeasytouse.Wethikyouwillloveit!Sicethemostimportatthigforlearigdeeplearigiswritigcodeadexperimetig,it'simportatthatyouhaveagreatplatformforexperimetigwithcode.
Togetstarted,werecommedusigaJupyterServerfromoeoftherecommededolieplatforms(clicktheliksforistructiosohowtousetheseforthecourse):
Colab:ApopularfreeservicefromGoogle.Googlehavechagedtheotebookplatformquitealot,sokeyboardshortcutsaredifferet,adoteverythigworks(e.g.muchofchapter2does'tworkbecauseColabdoes'tsupportcreatigwebappGUIs)Gradiet:UlikeColab,thisisa"real"JupyterNotebooksoeverythigithecourseworks.Italsoprovidesspacetosaveyourotebooksadmodels.However,sometimesthefreeserversgetover-loaded,adwhethathappesit'simpossibletocoect.IfyouareiteresteditheexperieceofruigafullLiuxserver,youcacosiderDataCruch.io(veryewservicesowedo'tkowhowgooditis,osetuprequired,extremelygoodvalueadextremelyfastGPUs),orGoogleCloud(extremelypopularservice,veryreliable,butthefastestGPUsarefarmoreexpesive).Westroglysuggestusigoeoftherecommededolieplatformsforruigtheotebooks,adtootuseyourowcomputer,ulessyou'reveryexperiecedwithLiuxsystemadmistratioadhadligGPUdrivers,CUDA,adsoforth.
Ifyoueedhelp,there'sawoderfuloliecommuityreadytohelpyouatforums.fast.ai.Beforeaskigaquestiootheforums,searchcarefullytoseeifyourquestiohasbeeasweredbefore.(Theforumsystemwo'tletyoupostutilyou'vespetafewmiutesothesitereadigexistigtopics.)OebitthatmaystudetsfidtrickyisgettigsigedupfortheBigAPIfortheimagedowloadtaskilesso2;here'sahelpfulforumpostexplaiighowtogettheBigAPIkeyyou'lleedfordowloadigimages.
Isthiscourseforme?Thakyouforlettigusjoiyouoyourdeeplearigjourey,howeverfaralogthatyoumaybe!Previousfast.aicourseshavebeestudiedbyhudredsofthousadsofstudets,fromallwalksoflife,fromallpartsoftheworld.Maystudetshavetoldusabouthowthey'vebecomemultiplegoldmedalwiersofiteratioalmachielearigcompetitios,receivedoffersfromtopcompaies,adhavigresearchpaperspublished.Foristace,IsaacDimitrovskytoldusthathehad"beeplayigaroudwithMLforacoupleofyearswithoutreallygrokkigit...[the]wetthroughthefast.aipart1courselatelastyear,aditclickedforme".HewetotoachievefirstplaceitheprestigiousiteratioalRA2-DREAMChallegecompetitio!Hedevelopedamultistagedeeplearigmethodforscorigradiographichadadfootjoitdamageirheumatoidarthritis,takigadvatageofthefastailibrary.
Itdoes'tmatterifyoudo'tcomefromatechicaloramathematicalbackgroud(thoughit'sokayifyoudotoo!);wewrotethiscoursetomakedeeplearigaccessibletoasmaypeopleaspossible.Theolyprerequisiteisthatyoukowhowtocode(ayearofexperieceiseough),preferablyiPytho,adthatyouhaveatleastfollowedahighschoolmathcourse.Thefirstthreechaptershavebeeexplicitlywritteiawaythatwillallowexecutives,productmaagers,etc.touderstadthemostimportatthigsthey'lleedtokowaboutdeeplearig--ifthat'syou,justskipoverthecodeithosesectios.
Deeplearigisacomputertechiquetoextractadtrasformdata–-withusecasesragigfromhumaspeechrecogitiotoaimalimageryclassificatio–-byusigmultiplelayersofeuraletworks.Alotofpeopleassumethatyoueedallkidsofhard-to-fidstufftogetgreatresultswithdeeplearig,butasyou'llseeithiscourse,thosepeoplearewrog.Here'safewthigsyouabsolutelydo'teedtodoworld-classdeeplearig:
Myth(do'teed)TruthLotsofmathJusthighschoolmathissufficietLotsofdataWe'veseerecord-breakigresultswith<50itemsofdataLotsofexpesivecomputersYoucagetwhatyoueedforstateoftheartworkforfreeDeeplearighaspower,flexibility,adsimplicity.That'swhywebelieveitshouldbeappliedacrossmaydisciplies.Theseicludethesocialadphysicalscieces,thearts,medicie,fiace,scietificresearch,admaymore.Here'salistofsomeofthethousadsoftasksidifferetareasatwhichdeeplearig,ormethodsheavilyusigdeeplearig,isowthebestitheworld:
Naturallaguageprocessig(NLP)Aswerigquestios;speechrecogitio;summarizigdocumets;classifyigdocumets;fidigames,dates,etc.idocumets;searchigforarticlesmetioigacoceptComputervisioSatelliteaddroeimageryiterpretatio(e.g.,fordisasterresiliece);facerecogitio;imagecaptioig;readigtrafficsigs;locatigpedestriasadvehiclesiautoomousvehiclesMedicieFidigaomaliesiradiologyimages,icludigCT,MRI,adX-rayimages;coutigfeaturesipathologyslides;measurigfeaturesiultrasouds;diagosigdiabeticretiopathyBiologyFoldigproteis;classifyigproteis;maygeomicstasks,suchastumor-ormalsequecigadclassifyigcliicallyactioablegeeticmutatios;cellclassificatio;aalyzigprotei/proteiiteractiosImagegeeratioColorizigimages;icreasigimageresolutio;removigoisefromimages;covertigimagestoartithestyleoffamousartistsRecommedatiosystemsWebsearch;productrecommedatios;homepagelayoutPlayiggamesChess,Go,mostAtarivideogames,admayreal-timestrategygamesRoboticsHadligobjectsthatarechallegigtolocate(e.g.,trasparet,shiy,lackigtexture)orhardtopickupOtherapplicatiosFiacialadlogisticalforecastig,texttospeech,admuchmore...WhoweareWeareSylvaiGuggeradJeremyHoward,yourguidesothisjourey.We'retheco-authorsoffastai,thesoftwarethatyou'llbeusigthroughoutthiscourse.
Jeremyhasbeeusigadteachigmachielearigforaroud30years.Hestartedusigeuraletworks25yearsago.Durigthistime,hehasledmaycompaiesadprojectsthathavemachielearigattheircore,icludigfoudigthefirstcompaytofocusodeeplearigadmedicie,Elitic,adtakigotheroleofPresidetadChiefScietistoftheworld'slargestmachielearigcommuity,Kaggle.Heistheco-fouder,alogwithDr.RachelThomas,offast.ai,theorgaizatiothatbuiltthecoursethiscourseisbasedo.
Sylvaihaswritte10mathtextbooks,coverigtheetireadvacedFrechmathscurriculum!HeisowaresearcheratHuggigFace,adwaspreviouslyaresearcheratfast.ai.
Wecarealotaboutteachig.Ithiscourse,westartbyshowighowtouseacomplete,workig,veryusable,state-of-the-artdeeplearigetworktosolvereal-worldproblems,usigsimple,expressivetools.Adthewegraduallydigdeeperaddeeperitouderstadighowthosetoolsaremade,adhowthetoolsthatmakethosetoolsaremade,adsoo…Wealwaysteachigthroughexamples.Weesurethatthereisacotextadapurposethatyoucauderstadituitively,ratherthastartigwithalgebraicsymbolmaipulatio.
ThesoftwareyouwillbeusigIthiscourse,you'llbeusigPyTorchadfastai.
We'vecompletedhudredsofmachielearigprojectsusigdozesofdifferetpackages,admaydifferetprogrammiglaguages.Atfast.ai,wehavewrittecoursesusigmostofthemaideeplearigadmachielearigpackagesusedtoday.WespetoverathousadhourstestigPyTorchbeforedecidigthatwewoulduseitforfuturecourses,softwaredevelopmet,adresearch.PyTorchisowtheworld'sfastest-growigdeepleariglibraryadisalreadyusedformostresearchpapersattopcofereces.
PyTorchworksbestasalow-levelfoudatiolibrary,providigthebasicoperatiosforhigher-levelfuctioality.Thefastailibraryisthemostpopularlibraryforaddigthishigher-levelfuctioalityotopofPyTorch.Ithiscourse,aswegodeeperaddeeperitothefoudatiosofdeeplearig,wewillalsogodeeperaddeeperitothelayersoffastai.Thiscoursecoversversio2ofthefastailibrary,whichisafrom-scratchrewriteprovidigmayuiquefeatures.
WhatyouwilllearAfterfiishigthiscourseyouwillkow:
Howtotraimodelsthatachievestate-of-the-artresultsi:Computervisio,icludigimageclassificatio(e.g., classifyigpetphotosbybreed),adimagelocalizatioaddetectio(e.g., fidigwheretheaimalsiaimageare)Naturallaguageprocessig(NLP),icludigdocumetclassificatio(e.g., moviereviewsetimetaalysis)adlaguagemodeligTabulardata(e.g., salespredictio)withcategoricaldata,cotiuousdata,admixeddata,icludigtimeseriesCollaborativefilterig(e.g., movierecommedatio)Howtoturyourmodelsitowebapplicatios,addeploythemWhyadhowdeeplearigmodelswork,adhowtousethatkowledgetoimprovetheaccuracy,speed,adreliabilityofyourmodelsThelatestdeeplearigtechiquesthatreallymatteripracticeHowtoimplemetstochasticgradietdescetadacompletetraiigloopfromscratchHowtothikabouttheethicalimplicatiosofyourwork,tohelpesurethatyou'remakigtheworldabetterplaceadthatyourworkis'tmisusedforharmHerearesomeofthetechiquescovered(do'tworryifoeofthesewordsmeaaythigtoyouyet--you'lllearthemallsoo):
RadomforestsadgradietboostigAffiefuctiosadoliearitiesParametersadactivatiosRadomiitializatioadtrasferlearigSGD,Mometum,Adam,adotheroptimizersCovolutiosBatchormalizatioDropoutDataaugmetatioWeightdecayImageclassificatioadregressioEtityadwordembeddigsRecurreteuraletworks(RNNs)SegmetatioAdmuchmore
评论