[PDF] Edge AI: Deep Learning techniques for Computer Vision applied to

View - Download Edge AI: Deep Learning techniques for Computer Vision applied to

Previous PDF

Next PDF

AlmaMater Studiorum·Universit

adiBologna

ScuoladiIngegneri aeArch itettura

CorsodiLaureaM agistrale inIngegneriaeSci enzeInformatiche

EdgeAI:

DeepLearningte chniquesforComputer

Visionappliedtoembe ddedsystems

ElaboratofinaleinMachineLearning

Relatore:

Prof.DavideMalto ni

Co-relatore:

dott.VincenzoLom onaco

Presentatada:

GiacomoBartoli

IISessi one

AnnoAccadem ico2018/19

Allamiafamiglia,

peravermisempre appoggiatoinogni miascelta.

When machines can see, doctors and nurses will have extra pairs of tireless eyes to help them to diagnose and take care of patients. Cars will run smarter and safer on the road. Robots, not just humans, will help us to brave the disaster zones to save the trapped and wounded. We will discover new species, better materials, and explore unseen frontiers with the help of the machines. Little by little, we're giving sight to the machines. First, we teach them to see. Then, they help us to see better. For the first time, human eyes won't be the only ones pondering and exploring our world. We will not only use the machines for their intelligence, we will also collaborate with them in ways that we cannot even imagine. This is my quest: to give computers visual intelligence and

to create a better future for the world. Fei-Fei Li, Head of Stanford AI Lab

Introduction

Morethanh alfacenturyago thecomputerw asi nvented.Sincethat daymanyfe lttheesse nceofthinking,theheartofin telligenc ewasfound. Reproducingintelligenceseemedtob epossiblebythewaycomputersworked. Allofas udden, itb ecamepossibletosimul atethi nking,problemsolv ing, evennatural language:ArtificialInt elligencewasborn.Thehumanb rainwas interpretedasacomputer.Weallagre eon thefac tthatcomputershaveb een oneoftheb iggesta chieveme ntsinhistory,butcom putershavenotfulfilled theexpec tationsofproducingintelligenceaswen ormally understandit.Here aresomec aseswhereAIse emedtobemoresm arterthanhumans : •In1997,w orld chesschampionGarry Kasparovplay edagainstDeep Blue,aprogrambu ilt byIBM.Asyoualreadypro bablyknow,Kas- parovwasde featedby themachine. •In2011,W atson, anothersoftwaremadeb yIBM,wonagainstahu- manplayer .ThegamewascalledJeop ardyandit was,essenti ally,a quizsho wwhereplayersw ereaskedquestions concerninganydomain: generalknowledge, history,politics,economicsetc. Boththese casesdonotn ecessarily implyth atacompu termustbesmart inorder toachieve these goals.Infact,achess algorithmcanperformve ry wellinsearch ingthr oughmanypositionsandpossibl emoveswithoutrequir- ingso muchintellige nce.Evenretrie vinginformationfromapre-builtarchive isnots omethingwe candefineassmart.Whilescie ntists andengin eerswere pushingthemse lvestobuildrobotscapableofdoingcommonplaceactivities i iiINTR ODUCTION (pathnavigation, obstacleavoidanc e,facerecognition) theyrealizedthat thesetaskswere extremelyhardf ormachines.Wh atiseasyforpeopleis extremelyhardforcomputersan d,viceversa,what isoft enhardforhumans iseasy forcomputer s.Bythemi d-1990s,researchersfromArtifici alInt elli- gencehadto admitthatp erhapstheidea ofcomputersas smartmachines wasmisgui ded.AsRolfPfeiferandChristianS che ierwroteint heirbook "UnderstandingIntelligence"[1],thebrai ndoesnotsimply"runprograms" butitdoes someth ingcomplete lydi erent.Whatthebraind oesisrei nforc- ingconnec tionsamongneurons,whichareacti vateddepending onspecific stimuli.Essentially,thebrai nlearnsfromexperience.Theideaofcr eatinga classofalgori th msthatlearnfromexperienceissumm edupbyasub fieldof ArtificialIntelligencecal ledMachineLearning.Themoreexamplesw egive them,themorethey learn and,subs equently,this knowledgecan beapplie d forin ferenceovernewexamples. Whenlittle childrenstillstru ggletotalk,parentsstillta lktothemas ifthe ycouldunderstand everything.This way,thebrainofthec hildrenis nourishedbyexamples .Furthermore,w henchildrenstartutteringtheir first wordsandtheyu seatermwr ong,thepare ntscorrec tth emwiththeri ght wordand,ag ain,childrenl earn.Thisiswhatin MachineLearningiscall ed SupervisedLearning:amodelisfe dbythousandsof examples andthen itisab letopr edict,act,c lassif y.However,ifthemodeliswro ngabouta predictionitcanbeautomatically correc ted,learn ingfro mitsownmistakes. Inth elastdecade,Machi neLearni ngtechniq ueshavebeenusedindi er- entfields ,rangingfromfinancet ohealthcareandevenma rketing.Amon gst alltheset echniques,theo nesadoptingaDeepLearningapproachw erere- vealedtooutperfo rmhuma nsintaskssuchasobjectdetecti on,imageclas- sificationandspeechrecognit ion. Thisthesisin troducesthebasiccon ceptsofMachineLearningandDeep Learning,andthendeepensthe convoluti onalmodel(CNN) .Thesecond chapterspecificallyi ntroducesDeepLearningarchitecture spresentinthe scientificliteratureforob jectrecognition.Thenweintroducetheconcept of

INTRODUCTIONiii

"EdgeAI",thatis thepossib ilitytobuil dlearn ingmodels capableofmak- inginfere ncelocally,withoutanydepend enceonexpensiveserverso rcloud services.Afirstcasestudywec ons iderisbasedo ntheGoogleA IYVision Kit,anintel ligen tcameraequippedwithagraphicboardtoo ptimizeCom- puterVisionalgor ithms.Thenwefo cusontwoapplications:wewanttotes t thepe rformancesofCORe50,adatasetforcontinuous obj ectrecognition, onemb eddedsystems.Thetechniquesdevelopedin thepreviousc hapters willthenbeus edtosolveac hallen gewithintheAu diAuton omousDriving Cup2018,where am obilecarequip pedwi thacamera,sensorsandagrap hi c boardmustrecog nizepedestri ansandstopbeforehittingthe m.

Introduzione

Pi`udimezzos ecolof ailcomputer fuinventato.Aparti redaqu el giorno,molte personein tuironochel'essenzadelragionamen to,ilcuore dell'intelligenzafutrovato.Sembravaatuttipossibil esfrut tareilmodoin cuiicomput ero peravanoperriprodurr eintelligenza.All'improvvis osembr`o possibilesimulareilpensier o,svolgereinmanieraautoma tizza taattivit`adi problemsolvingedanch eriprodurreillinguag gionat urale:l'Intelli genzaAr- tificialeeraappe nanata.Ilce rvelloumanoveniva interpre tatopropriocome uncompute r. Siamotuttid'ac cordos ulfattocheilcompute rsiastato unodeipi`u grandi traguardinella storia,tuttaviaic omputernonhannoraggiunto l'aspettativa diripr odurreintelligenzacos`ıcomeno ilaintendiamo.Analizziamoiseguenti casiincuil'AI sembra vaessere pi`uintel ligentediesseriumani: •Nel1997 ,GarryKasparov, campionemondia lidiscacchi,gioc`ocontro DeepBlue,unp rogrammacostrui todaIB M.Comenoto,Kasparovfu battutodalcalcolatore. •Nel2011 ,Watson,unaltro programmafattodaIBM, vinsec ontroun altrocampione.I lgiocoinquestionesichiamav aJeopar dyeconsisteva inunas pe ciediquizcondomanderelativeadogniposs ibileambito: culturagenerale,storia,polit ica,economiaeccecc. Entrambiicasinonrichiedono necessari ament eintel ligenzadapartedella macchina.Infatti,larice rcadipossibilicom binazioni emosse,comenelcaso degliscacchi,no nrichiedetroppaintellige nza.Ancher eperireinformazioni i iiINTRO DUZIONE dauna rchivio` equalcosachenonsi pu`oconsiderare intelligente.Me ntre scienziatiedingegnericoncentr avan oleloroforzenellosvi luppodirobotca- pacidiintra prender eattivit`aperumaniconsideratecomuni (muoversiinun ambiente,evitareostacoli,riconoscerefacce, a errareoggetti)si reserocon to chequesti taskeranoestremam entedi ciliperlemac chine.Ci` oche`e facile perleper sone`e moltodi cilepericompu tere, vic eversa,cioche`edi cile perleper sone`e facilepericompute r. Aca vallodeglianni90',ric ercatorinell'a mbitodell'In telligenzaAr tificale dovetteroammetterecheforsel'ideadic omputerintesocome unamacchina intelligenteerasbagliata.ComeRolfPfe ifereChris tianScheier scrivono nelloroli bro"Understan dingIntelligen ce"[1],ilcervellononmandasem- plicementeprogrammiinesecuzio nemafaqualcosadicompletamen tedi- verso.Ci`ocheilcer vellofa`eri nforzarele sinaps i,leconnessionitradi- versineuroniasec ondadeglistimolic her icevonodall'esterno. Pi`ustimoli riceviamo,pi`uleconne ssionisirinforzano.I np ocheparole,ilcervelloap- prendedall'espe rienza.L'ideadicreareunaclassedialgoritmicheimpara dall'esperienza`eriassuntainunasottobrancade ll'In telligenzaArtificiale chiamataMachineLearnin g.Pi`uesempidiamo inpastoaquestialgoritmi conoscenzaperfareinferenzasun uoviesempi . Quandoibambini,anco rap iccoli,stentanoaparl are,igenitoriparla noloro comesepotesser oin tendereognicosa.Inqu estamaniera,ilcervellodei bambini`enutrito inmanieracontin uadiesempi. In pi`u,quandoilbam bino imparaleprime paroleele usain manieras bagliatailge nitorelo corregge. Questo`eci`ochein Machin eLearningvi enechiama tocome Apprendimento Supervisionato:unmodello`ealimen tatodamigliaia diesempie,di con- seguenza,diventa ingradodipredire,agire,cl assificare.Tu ttavia,seil mod- ellosisbaglia inu nodiquestitaskpu `oessereaut omaticamentecor retto.Si apprendedaipro prierrori. Negliultimid iecianniletecniche diMachineLearn ingsonostate appli cate aipi`u svariatiamb iti,dallafinanzaallamedic inafinoalmarketing.Tra tutte

INTRODUZIONEiii

questetecniche,que llebasatesulDeepLearnin ghannodimostratodi avere performancemiglioridegliessere umaniintaskcomerilevazi onedioggetti, classificazionediimmaginiericonoscimentode lparla to. L'elaboratoinquestioneintro ducei concettibasedell'a pprendimentoauto- maticoedelDe epLe arning,perp oiapprofondireilmodellocon voluzionale (CNN).Ilsecond ocapit oloesponeeconfrontal earchitetturediDeepLear n- ingpres entiinletteraturaperilriconoscime ntodi oggetti.Siintroduce poiilconc ettodi "EdgeAI",ovverolapossibi lit`ad icostruiremode llidi apprendimentoingradodifareinferenzaloca lmente ,senzaalcunadi pen- denzadaservizi cloudo servercostosi.Ilc asodistudio` ebas atosulGoogle AIYVisio nKit,unacamera intelligente dotatadi unaschedagraficaper l'ottimizzazionedialgoritmidiComputerVision.Losco pofinal e`eduplice: dauna partesi voglionote stareleperf ormancediCORe50, datasetper ilric onoscimentocontinuodioggetti,susistemiembe dded.Inseguito,le tecnichesviluppateneicapitolipre cedentis arannoutilizzatepe rrisolvere unachalle ngeall'internodell'AudiAuton omousDrivingCup2018,doveuna macchinadotatadicamera,s ensori esc hedagraficadeveric onoscereipedoni efermarsi.

Contents

Introductioni

Introduzionei

1Ma chineLearning1

MachineLearning1

1.1Machine LearningData.... ... ... .. ... ... ... .2

1.2Machine LearningProblems.... ... ... .. ... ... .2

1.3Intro ductiontoDeepLearning..... ... ... .. ... ..4

1.3.1ArtificialNe uron................... ..6

1.3.2ArtificialNe uralNetwork............. ....8

1.3.3Convolutio nalNeuralNetwork..............10

1.4T rainingdetails.... ... ... ... ... .. ... ... ..13

1.4.1Stoc hasticGradientDiscent... ...... ...... 13

1.4.2SoftMaxas functionforthe outputlevel ... .. ... 14

1.4.3Cross-Entrop yaslossfunction..... ... .. ... .14

1.4.4Regularization ... ... ... ... ... .. ... ... 15

1.4.5Momentum ...... ... ... ... .. ... ... .16

1.4.6LearningRate ... ... ... ... .. ... ... ... 16

2Co mputerVision19

2.1DeepAr chitecturesf orObjectDetection............20

2.1.1R-CNN...... ........... ... ... ... 21

v viINTRODUZIONE

2.1.2FastR-CNN.. ........... ...........22

2.1.3FasterR-CNN ................. ......24

2.1.4YOLO ...... ... ... ... .. ... ... ... .26

2.1.5SSDMultib ox ......... ... ... ... ... .. 28

2.1.6Performance evaluation.... ...... .. ... ... 29

3T ensorflowforObjectDetectio n35

3.1HighLe velAPIswit hKeras........ ......... ..36

3.2HighLe velAPIswit hEagerExecutio n........ .....39

3.3LowLe velAPIs:Te nsors,Graphsan dSessions... ......41

3.4Tensor flowObjectDetectionAPIs..... ...........45

3.4.1Training amodelandexport afrozengraph ..... .45

3.4.2Outof thebo xinference. ... ........ ... ..47

4Edg eAI51

4.1Mobil eNets...................... ... ... 52

4.2Acases tud y:GoogleAIYVi sionKit........... ...55

4.2.1VisionBonne t............. ..........56

4.3Gettin gstarted:AIYdemos.... ...............57

4.4Customdeplo y.. ...... ... .. ... ... ... ... .59

4.4.1Creating thedataset. ... ... ... ... ... .. .61

4.4.2Setting thetrainingpip eline.. ... ...... .. ..64

4.4.3Training phaseandevaluation. ... ... ...... .67

4.4.4Exporting thefrozengraph... ... ... .. ... .69

4.4.5Co ding..... ... ... ... ... .. ... ... ..69

4.4.6Runningthe modelontheVisionKit.. ........74

4.4.7Runningth emodelasAndroidapp... ........75

5CO Re50attheEdge77

5.1Intro duction........ ... ... .. ... ... ... ... 77

5.2Parsing data,settingtheTFpipeline ... ... ... .. ...78

5.3Perfo rmanceevaluation.............. ........79

vii

5.4Comparing MobilenetvsSSD Mobilenet.......... ..82

6A udiAutonom ousDrivingCup85

6.1TheV ehicle. ......... ... .. ... ... ... ... .86

6.2TheChallenge ... ... ... ... ... .. ... ... ... .87

6.3Artific ialIntelligencedrivingta sks................88

6.3.1Zebracrossing task.. ... ... .. ... ... ... .88

6.3.2Crossingtask ... ... ... ... .. ... ... ... 89

6.3.3Adultversus child...... ...............89

6.3.4YieldingtoE mergencyVehicles.... ...... ...90

6.4DeepLearning techniquesfor drivingtasks. ..... ... ..90

6.5Performa nceevaluation..... ...... ... ... .. ... 91

6.6Testin gthemodelontheADTF.... ...... .......92

6.6.1T estinginferencespeed.. ... ...... .. ... ..93

6.6.2Deployont heAudicar......... ...... ...96

Conclusions99

Appendix99

Aem beddedssdmobilenetpipeleine101

BScri ptforconver tinglabel s107

CScr iptforconve rtingPNGim agestoJPEG109

DScri ptforcreat ingCORe50cs vfile111

ESpl ittingclassificationandde tectionerrors115

Bibliografia117

ListofFigures

1.1Exampleof overfitting ... ........ ... ... ... ..4

1.2ComparingMac hineLearningand DeepLearningperformances5

1.3Artifici alNeuron................ .........6

1.4StandardLogistic Function. .. ...... ... ... ... ..6

1.5Hyperbo licTangent................ ........7

1.6Artific ialNeuralNetwork......... ............8

1.7VisualC ortexSystem... .................. ..10

1.8Conv olutionoperationwithfilter.... ..... ... ... .11

1.9Poolin goperationusingmax,2x2 filterandstride=2.....12

1.10ReLU,Rec tifiedLinearUn it..................12

1.11Convolu tionalNeuralNetwork.................13

1.12SGD andLearningRate ... ... ... ... ... .. ... .14

1.13Thee

ectofmomentum. ... ...... ... ... ... .. 16

2.1ImageClassification, Localization, DetectionandSegmen tation20

2.2Selective Search... ...... ... ... ... ... ... .. 21

2.3R-CNN.. ............ .. ... ... ... ... ... 22

2.4FastR- CNNArchitectur e.......... ..........23

2.5RoI Pooling Layer...... ......... .. ... ... ..23

2.6Faste rR-CNNArchitecture.. ........... ......25

2.7Regio nProposalNetwork... ..................25

2.8YOLO,i magedividedi ntoboundingbox es...........26

2.9YOLO,c lassprobabil ities......... ...........27

ix xLI STOFFIGURE S

2.10YOLO,boun dingboxandco nditional.............. 27

2.11YOLO,thre sholddetect ion....................28

2.12YOLOarchi tecture.. ......................28

2.13SSDarc hitecture.. ...... ... ... .. ... ... ... 29

2.14Precision-Recallcurv efor theclassdog.. ..... ... ... 30

2.15Intersec tionovertheUnion.................... 31

2.16IoUstop sign.. ... ... ... .. ... ... ... ... ..31

2.17Precision/Recall curvef orIoU.... ...... .. ... ... 32

3.1Object DetectionusingTen sorflowAPIs............50

4.1EdgeAIw orkflow...... .. .................52

4.2Stand ardConvolutionFilter..... ..............53

4.3Depth wiseConvolutionalFilter.... .............53

4.4Pointw iseConvolutionalFilter..... .............53

4.5Mobile NetArchitecture.......... ............54

4.6Mobile NetCheckpoints.......... ............54

4.7Googl eAIYVisionKit...... ...... ..........55

4.8VisionB onnet......... .................. 56

4.9Myria d2VisionProcessingUn it. .............. .57

4.10JoyDetec tordem o...................... ..58

4.11Image ClassificationCamera demo..... ... ... .. ..59

4.12Pikac hu......... ... ... ... ... ... ... .. .60

4.13CSVlab elsand bbox... ... ......... ... .. ... 62

4.14Pikachud etectoronGoogleAIYVisi onkit...........75

4.15Pikachud etectorasAndroidapp.. ..............76

5.1Screenshotof pre-trainedchec kpoints ... ...........79

5.2SSDMobilenetaccuracy....... ......... .. ... 80

5.3SSDMobilenetloss........ ...... ... .. ... ..80

5.4CORe50detection .. ... ... ... ... ... ... ... .. 81

5.5SSDMobilenetdetection..... .............. ..81

xi

6.1AADCveh icle,c omputerandsensors...... ........87

6.2Zebra crossing.. ... ... ... ... ... .. ... ... ..88

6.3Crossing. .. ... ... ... ... ... ... ... ... .. .89

6.4Adultve rsuschild... ................. .....89

6.5Yieldi ngtoemergencyvehicles ..... .............90

6.6Precision. ... ... ... ... .. ... ... ... ... ... 91

6.7Precisionp ercategory ...... ...... .. ... ... ..92

6.8Detectingadult, childand emergencycar.. ...... .. ..92

6.9Testing multipledetections.. ...... ...... ... .. .93

ListofTables

2.1Object Detectionperformances... ... ..... ... ... 32

4.1Vision Bonnetconstrain ts............... .....65

4.25-Fold CrossValidation.. ...... ..... ... ... ... 68

4.35-Fold CrossValidationre sults........... .......68

4.45-Fold CrossValidationre sults........... .......76

5.1CORe50b enchmarksfor classification....... .. .....82

6.1Audi's carinferencetime ........ .............96

xiii

Chapter1

MachineLearning

MachineLearningisasub jectlocatedattheinterse ctiona mon gStatistics, DataAnalysis ,PatternRecognitionandAr tificialIntelligence.Asa lready explainedintheintrodu ction,th emainideai sthatoffeedingamodelwith manyexamplesand, lateron,applyinginference .Th us,machines "mag- ically"learnfromdata .Thepointistode finethecorrec tmodelforthe learningphase.Moreformal ly,wecansaythat "acomputerp rogramissaid tolearn fromexperie nceEwithre specttosomeclassesoftasksT andpe rfor- mancemeasureP ifitsperformanceattas ksinT ,asme asure dbyP,improves withexperience E"[2].Adatasetisnormallysp lit intoth ree part s:thefirst part,calledthet rainingset,isusedf orthetrai ningphase;thesecondo ne, thetes tset,isaimedatthe evaluation,whilethelastone ,the validation set,issui tablefor tuninghyperparameters.Th etraini ngphaseischarac- terizedbythelearningproce ssandknowle dgeacquis itionofthemode l.The subsequentapplicationofsaidkn owledgeisthetestingphase. If,during thetrainingphas e,dataare labelledthenitiscalledS upervised Learning.Thegoalistofindafun ctionthat mapstheinpu tda taonit s correspondingclasses.Incasetherearenolabel s,which makes the problem harderthanbefore,i tiscalledUns upervisedLearning.The train ingset can evenbepar tiallyl abelled:thisiswhatSemi-superv isedLearningmeans.The lastparadigmofMac hineLearning isReinf orcementLearning:an agenttake s 1

21. MachineL earning

actioninteractingw iththeenvironment.Toeachactionth erecorrespon dsa reward.Thegoaloftheag entistom aximizethesum ofthe totalrewa rds.

Thisapproachh asshowntobeincredibl ye

ectivewhenanagentmust learn behaviours:controltheory,simul ation,gaming.

1.1Machine LearningData

MachineLearningalgori thmscanhandledi

erentkinds ofdata: •Numerical:thesearevaluesass ociatedwithmea surablech aracteristics. Thereisanordera mongth emandt heycanbebot hdiscreteorcont in- uous.Theyc anberepresente dasvectors inamultidimens ionalspace. Ex:findi ngtheheight,weightorfoot sizeofagivenp erson. •Categorical:valuesassociatedwithqua litativecharacteri stics.Binary valuesareconside redcateg oricalaswell.Ex:findingthesexo rblood typeofgive napers on. •Sequences:thesedataexpressarelations hipbetweentime andspac e. Whatreally mattersistheirpos itionintoasequenceand there ference withpredecessorsan dsuccessors.Ex:asequenceofwor ds,streams of data.

1.2MachineL earningProblems

MachineLearningtech niquescanbeappliedforfacingseveral problems, whichare: •Classification:startingfromalabelleddataset ,theaimistocorrectly classifynewpatter nsasbelon gingtotheirclasses.Ex:givent hew eight andheight ofaperson,isthat amaleor female?Traditional approaches toclas sificationtasksare:supportedvectormachi nes(SVM),decisi on trees,perceptron,neare stneighborsandothermodels.

1.2MachineL earningProblems3

•Clustering:groupingdatathatshar ethesamecharacteristics.Dat a arenotlab elled.Thepurp oseistominimizeint ra-clusterdistanceand maximizeinter-c lusterdistance.Clusteringalgorithmsare:K-means,

K-median,X-means,etc.

•Regression:thetask ofappro ximatinga mappingfunctionfrom input variablestoacontinuousou tputvar iab le.Evenifitmayseemcloseto classification,themaindi erenceisthefact thatregressi onconsiders continuousvariablesasoutput,wh ileclassificationworksonly with discretevalues. •DimensionalityReduction:mappingaspac eK n toK m wherem41. MachineL earning

Figure1.1:Examp leofoverfitti ng

therelat ionshipbetweendataandlabels.Figure1. 1showsanexampleof overfitting. MachineLearningalgor ithmshaveaperformanceli mit.Whenthislimit isreac hed,evenifdatainputisin creased,perf ormanc esdonotimpro ve. Toov ercomethislimititisneces sarytointro ducemorepowerfu ltech niques basedonaDeep Learningapproac h.In figure1.2itispossibleto seethe di erencebetweenM achineLearningandDeepLearningalgorithmsinterms ofperf ormances.

1.3Introduct iontoDeepLearning

Deeplearnin gisthefieldofresearchinMa chineL earningth atisbase d ondi erentlevelsofre presentation,corr espond ingtohierarchiesofcharac- teristicsoffactors,where high-leve lconceptsaredefinedonthe basisoflow-quotesdbs_dbs18.pdfusesText_24

[PDF] Ma Cuisine - Magimix

[PDF] livre-maitre

[PDF] Télécharger le petit livre 2017 - Fédération Nationale des Chasseurs

[PDF] Livre Phare 4eme | PDF, DOCX, EPUB and other eBooks formats

[PDF] LISTE DES MANUELS SCOLAIRES 1ère S - Lycée international

[PDF] Livre du professeur - Le livre du prof

[PDF] Manuels scolaires Classe de Terminale ANNEE SCOLAIRE 2016

[PDF] Les 10 étapes du succès par la PNL - PNL Audio Institut

[PDF] Histoire de la science politique dans ses rapports - Gallica - BnF

[PDF] Manuel de jardinage pour débutants - aupetitcolibri - Free

[PDF] Le jardin potager dans les zones tropicales - Agriculture et

[PDF] PRATIQUE DU FRANÇAIS

[PDF] Programme du CAPLP externe d économie et gestion de la session

[PDF] concours officier de police l essentiel - Bookelis

[PDF] Le Livre et la lecture en Algérie - unesdoc - Unesco