20 minutes DeepLearning

by ryo nakamaru

Slide 1

Slide 1 text

20 ෼ͰΘ͔Δ DeepLearning AWS re:Invent ೺ѲͷҰॿʹ re:Port 2016 @ 2016.12.07

Slide 2

Slide 2 text

தؙ ྑ @pottava SUPINF Inc.

Slide 3

Slide 3 text

ૣ଎Ͱ͕͢ΫΠζͰ͢

Slide 4

Slide 4 text

re:Invent 2016 ϒϨΠΫΞ΢τηογϣϯ਺͸ʁ

Slide 5

Slide 5 text

393 ʢAWS re:Invent2016ʲอଘ൛ʳશηογϣϯΛ໢ཏʂඇެࣜ·ͱΊ #reinvent ʛ Developers.IOʣ

Slide 6

Slide 6 text

re:Invent 2015, 2014 ϒϨΠΫΞ΢τηογϣϯ਺͸ʁ

Slide 7

Slide 7 text

298, 214 ʢAWS re:Invent2015ʲอଘ൛ʳશηογϣϯΛ໢ཏʂඇެࣜ·ͱΊ #reinvent ʛ Developers.IOʣ ʢ2014೥౓͸ฐࣾͷϒϩάΑΓงғؾͰूܭʣ

Slide 8

Slide 8 text

Ͱ͸

Slide 9

Slide 9 text

re:Invent 2017 ϒϨΠΫΞ΢τηογϣϯ਺͸ʁ

Slide 10

Slide 10 text

214 → 298 → 393 → ?

Slide 11

Slide 11 text

͜ΕΛػցʹ༧ଌͤ͞Δ ͦͷϓϩηε͕ػցֶश

Slide 12

Slide 12 text

͜ΕΛػցʹ༧ଌͤ͞Δ ͦͷϓϩηε͕ػցֶश

Slide 13

Slide 13 text

ຊ୊

Slide 14

Slide 14 text

20 ෼ͰΘ͔Δ DeepLearning

Slide 15

Slide 15 text

20 ෼ͰΘ͔Δ DeepLearning ͞ΘΓ͕

Slide 16

Slide 16 text

ΰʔϧ ɾձࣾʹ re:Invent ͷ͔͍͍ͬ͜ϨϙʔτΛग़ͤΔ ɾࠓճͷ Deep Learning ܥαʔϏεͳΒԶʹฉ͚ ɾࣾ௕ɺ࣌୅͸ AI ͢Α

Slide 17

Slide 17 text

ΰʔϧ ɾձࣾʹ re:Invent ͷ͔͍͍ͬ͜ϨϙʔτΛग़ͤΔ ɾࠓճͷ Deep Learning ܥαʔϏεͳΒԶʹฉ͚ ɾࣾ௕ɺ࣌୅͸ AI ͢Α ɹ΄Μͱʹݴ͏ͱΠλ͍

Slide 18

Slide 18 text

ࠓ೔ͷྲྀΕ 1. Deep Learning ͬͯͳΜͳͷʁ 2. ͦͷ্Ͱ AWS ͷൃදͬͯͲ͏ͩͬͨͷʁ

Slide 19

Slide 19 text

Deep Learning ͱ͸

Slide 20

Slide 20 text

ͷ·͑ʹ

Slide 21

Slide 21 text

Machine Learning ͱ͸

Slide 22

Slide 22 text

ػցֶशͷେલఏ ػցֶशʹ͸ 2 ͭͷϑΣʔζ͕͋Γ·͢ɻ ͦΕ͸ ֶशɹͱɹ ਪ࿦

Slide 23

Slide 23 text

ֶशͱ͸ ΑΓ͏·͘໰୊ʹରॲͰ͖ΔΑ͏ʹͳΔաఔͷ͜ͱɻ

Slide 24

Slide 24 text

͡Ό͋ਪ࿦͸ʁ ֶΜͩ͜ͱΛ΋ͬͯ໰୊Λղ͘͜ͱɻ

Slide 25

Slide 25 text

͜Ε͸ֶश

Slide 26

Slide 26 text

͜Ε͸ਪ࿦

Slide 27

Slide 27 text

͜Ε͸ֶश

Slide 28

Slide 28 text

͜Ε͸ਪ࿦ དྷ೥ re:Invent ߦ͖͍ͨΜ͚ͩͲ ձࣾʹԿΛใࠂ͢Δͷ͕ਖ਼ղʁ

Slide 29

Slide 29 text

େࣄͳ͜ͱͳͷͰ΋͏Ұ౓ ػցֶशͷ 2 ͭͷϑΣʔζɿ ֶशɹͱɹ ਪ࿦

Slide 30

Slide 30 text

ͳΔ΄Ͳ

Slide 31

Slide 31 text

ػցֶशʢMachine Learningʣͱ͸ ػցʹֶशͤ͞ɺਪ࿦ͤ͞Δ͜ͱɻ ղ͖͍ͨ໰୊࣍ୈͰɺֶशɾਪ࿦ͷํ๏Λม͑Δɻ ɾճؼ ɾΫϥε෼ྨ ɾΫϥελϦϯά ɾ..

Slide 32

Slide 32 text

Deep Learning ͱ͸

Slide 33

Slide 33 text

DeepLearningͱ͸

Slide 34

Slide 34 text

DeepLearningͱ͸ ͳΔ΄Ͳɺ͸͍͸͍

Slide 35

Slide 35 text

Կ૚΋ͷਂ͍ϨΠϠʔʢχϡʔϥϧωοτϫʔΫʣΛ ࢖ͬͯɺֶशɾਪ࿦͢Δػցֶशͷ͜ͱɻ ϨΠϠʔ͕ਂ͍͔ΒσΟʔϓɻ ೔ຊޠͳΒɺਂ૚ֶशɻਂ͍૚ɻ DeepLearningͱ͸

Slide 36

Slide 36 text

ػցֶशʢMLʣͱਂ૚ֶशʢDLʣ ML DL ɾ౷ܭత෼ੳʢճؼɺΫϥε෼ྨɾɾʣ ɾਐԽతܭࢉ ɾχϡʔϥϧωοτϫʔΫ ɾ… ɾ৞ΈࠐΈχϡʔϥϧωοτϫʔΫʢCNNʣ ɾ࠶ؼܕχϡʔϥϧωοτϫʔΫʢRNNʣ ɾਂ૚ Q ωοτϫʔΫʢDQNʣ ɾ…

Slide 37

Slide 37 text

ɾAlphaGo ɾ?? ɾ?? DeepLearning Կ͕͍͢͝ͷʁ

Slide 38

Slide 38 text

AlphaGo ਓؒͷੈքҰʹίϯϐϡʔλ͕উͬͯ͠·ͬͨɾɾʂ ͍ͭ࠷ۙ·ͰɺAI ͸ع࢜ʹউͯͳ͍ɺͦͯ͠ϒϨΠΫεϧʔʹ͸ 10 ೥଴ͨͶ͹ͳΒͳ͍ͱݴΘΕ͍ͯͨͷʹɺͩɻ

Slide 39

Slide 39 text

ʢ࿦จ: http://www.nature.com/nature/journal/v529/n7587/abs/nature16961.html ʣ • 4ͭͷσΟʔϓχϡʔϥϧωοτϫʔΫΛར༻ • ͦΕͧΕ ॳڃऀpπ / தڃऀpσ / ্ڃऀpρ / ৘੎൑அvθ ͱ͍ͬͨҐஔ෇͚ • pπ ͸ਓؒͷ800ສͷ൫໘σʔλΛݩʹֶशɻਫ਼౓͸௿͍͕ߴ଎ʹղΛಘΔɻ • pσ ͸13૚ͷCNNɻ3,000ສ൫໘Λ50GPUͰ3.4ԯεςοϓɺ3िֶؒशɻϓϩͷࢦ ͠खΛ57.0%ͷਫ਼౓Ͱ༧૝Ͱ͖Δɻ • pρ ͸50GPUͰ1೔͔͚128ສճࣗݾରઓɻطଘιϑτʹ85%ͷѹ౗తউ཰ɻ • vθ ͸ pσ ͰϥϯμϜʹ3,000ສ൫໘Λੜ੒͠ɺpρ Ͱ1ԯ6,000ສճϩʔϧΞ΢τͨ͠ উ཰Λڭࢣσʔλͱ͠ɺ50GPUͰҰिؒ 5,000ສճ֬཰ޯ഑߱Լ๏Λ࣮ࢪɻ • ࣮ରઓͰ͸1,202CPU+176GPU͕࢖ΘΕɺpσͰ࣍ͷखબ୒ɺvθͰ൫໘ධՁɻ • উ཰͸͍͍͕ཧ٧ΊͷpρΑΓɺਓؒͷบΛֶΜͩpσΛ࢖͔ͬͨɻ • উҼ͸ pρ ͷపఈతͳڧԽֶशʹՃ͑ɺϞϯςΧϧϩ໦୳ࡧͱCNNͷ૊߹͔ͤɻ AlphaGo

Slide 40

Slide 40 text

ɾAlphaGo ɾAmazon Go ɾ?? DeepLearning Կ͕͍͢͝ͷʁ

Slide 41

Slide 41 text

Amazon Go 1. ೖళ࣌ɺήʔτʹεϚϗΞϓϦΛ͔͟͢ 2. ΄͍͠΋ͷΛόοάʹೖΕΔ 3. ͓ళΛग़Δ Coming early 2017 !! 2131 7th Ave Seattle, Washington

Slide 42

Slide 42 text

Amazon Go

Slide 43

Slide 43 text

Amazon Go ళ಄Ͱ͸ɺਪ࿦͑͞ಈ͚͹͍͍ɻ ͍΍ɺͦΜͳ͜ͱΑΓ ͏ͪͷళฮʹ΋͜Εಋೖͯ͠Έ͍ͨΜͰ͚͢Ͳɾɾʂ

Slide 44

Slide 44 text

Slide 45

Slide 45 text

ਂ૚ֶशϕʔεͷը૾ೝࣝ APIɻ ɾҰൠ෺ମ / ৘ܠݕग़ ɾද৘෼ੳ ɾإͷྨࣅ౓൑ఆ

Slide 46

Slide 46 text

Amazon Rekognition ଞࣾͰ΋੝Μͳ Computer vision API ͷҰछɻ

Slide 47

Slide 47 text

Amazon Rekognition ฐࣾ༐ऀͷ ྨࣅ౓൑ఆɻ

Slide 48

Slide 48 text

ଓ͖·ͯ͠

Slide 49

Slide 49 text

ɾAlphaGo ɾAmazon Go ɾAmazon Echo DeepLearning Կ͕͍͢͝ͷʁ

Slide 50

Slide 50 text

Amazon Echo Alexa ʹԻ੠Ͱ͓ئ͍ɾ࣭໰͢ΔͨΊͷσόΠεɻ ʢAlexa ͸ Amazon ͕։ൃͨ͠ AI ʣ ʮΞϨΫαɺిؾΛ͚ͭͯɺԻָ΋͔͚ͯʯ ʮΞϨΫαɺUber ΛݺΜͰɻࠓ೔ͷఱؾ͸ʁ ʯ

Slide 51

Slide 51 text

Amazon Echo

Slide 52

Slide 52 text

Amazon Echo 1. Ի੠Λฉ͖औΓ 2. ԿΒ͔ͷॲཧΛͯ͠ 3. Ի੠Λฦ͢

Slide 53

Slide 53 text

Amazon Echo ɾӡసதͷϋϯζϑϦʔΞϓϦ࡞Γ͍ͨ ɾԻ੠ग़ྗɺTwillio ʹ࢖͑ͦ͏ͳ ɾ..

Slide 54

Slide 54 text

Slide 55

Slide 55 text

Amazon Polly ςΩετΛԻ੠ʹม׵͢Δ APIɻ ɾAlexa ͷԠ౴ػೳʮग़ൃ͸͍ͭͰ͔͢ʁʯ ɾ೔ຊޠʹ΋ରԠͯ͠Δʂ ɾXML Ͱ౉ͤ͹Ի੠ͷΧελϚΠζ΋͔ΜͨΜ

Slide 56

Slide 56 text

Amazon Lex ର࿩ΠϯλʔϑΣΠε + ࣗવݴޠղੳ APIɻ ɾϢʔβͱର࿩͠ͳ͕Β࣍ͷॲཧΛ൑அ ɾ࢒೦ͳ͕ΒɺݱࡏӳޠͷΈͷରԠ

Slide 57

Slide 57 text

Amazon Echo 1. Ի੠Λฉ͖औΓ 2. ԿΒ͔ͷॲཧΛͯ͠ 3. Ի੠Λฦ͢ Amazon Lex Amazon Polly

Slide 58

Slide 58 text

ͪͳΈʹଞࣾͷঢ়گ

Slide 59

Slide 59 text

ɾVision API ɹը૾ೝࣝ APIɻ෺ମ / จࣈ / إ / ϩΰ / ϥϯυϚʔΫͷݕ஌ɻ ɾSpeech API ɹಉࣾͷར༻͢ΔԻ੠ʹΑΔςΩετೖྗΛ REST API Խɻ ɾNatural Language API ɹจষΛॻ͍ͨਓͷҙݟ΍ײ৘ɺݻ༗໊ࢺ΍࣌ؒදݱΛநग़ɻ ɾTranslate API ɹ਺ઍ΋ͷݴޠؒͰϦΞϧλΠϜʹςΩετΛ຋༁ɻ https://cloud.google.com/#products_machine-learning

Slide 60

Slide 60 text

ɾVision ɹը૾ೝࣝɾՃ޻ɺײ৘෼ੳɺإݕग़ɾಉҰ൑ఆɾྨࣅݕࡧɺಈը෼ੳͳͲɻ ɾSpeech ɹจষͱԻ੠ͷ૬ޓม׵ɺԻ੠ʹΑΔϢʔβࣝผɾೝূɻ ɾLanguage ɹݴޠɾτϐοΫɾΩʔϑϨʔζநग़ɺωΨϙδ൑ఆɺεϖϧνΣοΫɺ༧ଌɻ ɾKnowledge ɹ஌ࣝσʔλϕʔεΛ༻͍ͨ෼ੳɺ༧ଌɻ ɾSearch ɹWebɾಈըɾը૾ɾχϡʔεɾؔ࿈ϫʔυͷݕࡧɾऩूͳͲɻ https://azure.microsoft.com/ja-jp/services/cognitive-services/

Slide 61

Slide 61 text

ɾLanguage ɹࣗવݴޠॲཧʹΑΔςΩετղੳɻಛఆ৘ใɾΩʔϫʔυɾίϯηϓτநग़ɺ ɹΧςΰϦ෼ྨɺײ৘෼ੳɺωΨϙδ൑ఆɺެจղੳɾҙຯ෇͚Ԡ౴ɻ ɾSpeech ɹจষͱԻ੠ͷ૬ޓม׵ɻ ɾVision ɹը૾ೝࣝɻ ɾData Insights ɹχϡʔε΍ϒϩάهࣄͷݕࡧɺτϨϯυ෼ੳɻ ɹա৒ͳબ୒ࢶͷআ֎ɺෳ਺ͷໃ६͢ΔΰʔϧΛؚΉΑ͏ͳෳࡶͳܾఆͷ෼ੳɻ https://www.ibm.com/watson/developercloud/services-catalog.html

Slide 62

Slide 62 text

ͳΔ΄ͲͰ͢Ͷɻ

Slide 63

Slide 63 text

ͯ͞ AWS ͷ 3 ͭͷ৽αʔϏε

Slide 64

Slide 64 text

Rekognition / Lex / Polly ͷڞ௨఺ શ෦ɺਪ࿦ API

Slide 65

Slide 65 text

ਪ࿦ͷαʔϏεͱ͍͏͜ͱ͸ ֶश͸ AWS ೚ͤɻ͜Μͳ͜ͱʹ͸޲͍͍ͯͳ͍ ɾࣗࣾͷ঎඼Λը૾ݕग़͢Δ ɾࣗࣾͷΩϟϥΫλʔޠʁΛղऍͤ͞Δ ɾࣗࣾήʔϜ಺ͷຐ๏Λࣗવݴޠղੳͤ͞Δ

Slide 66

Slide 66 text

͑ʔ

Slide 67

Slide 67 text

ࣗࣾͷ΋ѻ͍͍ͨͳʔ

Slide 68

Slide 68 text

࢓ํͳ͍ɺֶश͠Α͏

Slide 69

Slide 69 text

Α͠ɺֶशͩ 0. ͦΕɺDeepLearning Ͱղ͘ͷ͕࠷దͳͷʁ 1. σʔλ͕ඞཁɻʢ2016 ೥ݱࡏɺҰൠతʹ͸ʣͨ͘͞Μ 2. ֶशΞϧΰϦζϜͷબ୒ 3. ܭࢉ͠·͘Δίϯϐϡʔλࢿݯ͕ඞཁ

Slide 70

Slide 70 text

Α͠ɺֶशͩ 0. ͦΕɺDL Ͱղ͘ͷ͕࠷దͳͷʁ౷ܭత෼ੳ͸ʁ 1. σʔλ͕ඞཁɻͨ͘͞Μ 2. ֶशΞϧΰϦζϜͷબ୒ 3. ܭࢉ͠·͘Δίϯϐϡʔλࢿݯ͕ඞཁ

Slide 71

Slide 71 text

σʔλΛूΊΑ͏ ʢ͜ͷલʹ͢΂͖ɺେࣄͳॾݕ౼͸ AWS αʔϏεؔ܎ͳ͍ͷͰஔ͍ͱ͍ͯʣ σʔλΛूΊͳ͚Ε͹࢝·Βͳ͍ʂ ɾKinesis Streams ɾCloudWatch Logs ɾS3 / RDS / DynamoDB

Slide 72

Slide 72 text

ڭࢣσʔλΛ࡞Ζ͏ ·ͩಈը / εϥΠυͳ͔͚ͬͨͲɾɾ ϝΧχΧϧλʔΫΛ࢖͏࿩͕͋Γ·ͨ͠ɻ

Slide 73

Slide 73 text

ֶशϑϨʔϜϫʔΫΛબ΅͏ ͲΕ͕͍͍ΜͩΖ͏ɾɾʁ MXNet / TensorFlow / Caffe / Chainerɻ ɾͲͷχϡʔϥϧωοτ࢖͏ͷʁCNNʁRNNʁ ɾGPU ࢖͏ͷʁCPU ͚ͩʁෳ਺ϊʔυల։ඞཁʁ ɾࠃ࢈Ԡԉͪ͠Ό͏ʁ ɾ..

Slide 74

Slide 74 text

AWS ͸ MXNet Ұ୒ͩͦ͏Ͱ͢ ɾ͑ɺAmazon DSSTNE ɾɾ ɾଞͷ΋ͷ΋ݕ౼͍ͨ͠ํ͸ͪ͜Β ɹ re:Invent 2016 Bringing Deep Learning to the Cloud with Amazon EC2 (CMP314)

Slide 75

Slide 75 text

Deep Dive ͍ͨ͠ํ͸

Slide 76

Slide 76 text

ࠓ೔ͷ࿩ΛΑΓਂ͘ɺAI ؔ࿈৘ใ͚ͩͷ 120 ෼ʂʂ http://jawsug-ai.connpass.com/

Slide 77

Slide 77 text

࠷ޙʹɺֶशΠϯϑϥબఆ

Slide 78

Slide 78 text

DL ޲͚ AWS ϚωʔδυαʔϏε దࡐదॴͰϐλΰϥεΠονɺࠓޙ͋Γ͑Δ ɾGlue ɾEMR ɾECS ɾBatch ɾ.. new! new!

Slide 79

Slide 79 text

͑ɺAmazon ML ͸ʁ ML DL ͜͜ͷྖҬͰ໰୊Λղ͘ͳΒ΋ͪΖΜ࢖͑·͢ʂ Amazon ML

Slide 80

Slide 80 text

Amazon ECS ɾre:Invent Ͱͷ MXNet ϫʔΫγϣοϓָ͔ͬͨ͠ ɹhttps://github.com/awslabs/ecs-deep-learning-workshop/ ɾECS Ϋϥελ্Ͱ MXNet ͷֶशɾਪ࿦ ɾDocker ίϯςφલఏʹ͸ͳΔ͕ ɾDocker ͸༷ʑϝϦοτ͕͋Γͭͭ ɾίϯςφΫϥελͱͯ͠ಈ͔͢ͳΒ࠷ߴ

Slide 81

Slide 81 text

AWS Batch ɾECS ্ʹ HPC ۀքͷҙຯʹ͍ۙΫϥελΛߏஙɻ ɾδϣϒεέδϡʔϥ ≠ ίϯςφք۾ͷεέδϡʔϥ ɾECS ্ͳͷͰɺίϯςφϕʔε ɾGlue ΍ EFS ͱͷ૊Έ߹Θ͕ͤॏཁͦ͏

Slide 82

Slide 82 text

ECS ͱ Batch ɾֶशϑΣʔζ͸ɺେྔͷσʔλͱେྔͷܭࢉ - EFS / EBS / Batch ͳͲʹΑΔδϣϒ؅ཧ ɾਪ࿦ϑΣʔζ͸ɺେྔͷϦΫΤετΛࡹ͘ - Մ༻ੑߴ͘εέʔϧՄೳͳৗ࣌ىಈ API αʔό - ECS ্ʹαʔϏεσϓϩΠ

Slide 83

Slide 83 text

Deep Dive ͢ΔͳΒ

Slide 84

Slide 84 text

ECS Λத৺ʹɺίϯςφ·ΘΓͷ࠷৽৘ใΛ͓ಧ͚ʂ http://jawsug-container.connpass.com/

Slide 85

Slide 85 text

http://prtimes.jp/main/html/rd/p/000000007.000007768.html Comfy for Docker ϓϩδΣΫτ΁ͷ Docker ಋೖɾ։ൃࢧԉɾӡ༻؂ࢹ୅ߦΛ͍ͨ͠·͢ɻ https://www.supinf.co.jp/service/dockersupport/

Slide 86

Slide 86 text

ͪͳΈʹଞࣾͷঢ়گ

Slide 87

Slide 87 text

GPU Λ࢖ͬͨϑϧϚωʔδυͳ TensorFlowɻ Google ͷάϩʔόϧ༧ଌϓϥοτϑΥʔϜͰ਺ઍϢʔβɺTB ڃσʔλͰ΋ɻ ΋ͪΖΜɺCloud Storage ΍ BigQuery ͱͷ౷߹΋؆୯ɻ TensorFlow ͕ۤखͳॲཧ͸ɾɾ

Slide 88

Slide 88 text

GPU΍InﬁnibandΛؚΉϚϧνΫϥ΢υHPCϋʔυ΢ΣΞ͕࢖͑Δɻ ΦϯϓϨϛεͱϋΠϒϦουԽͰ͖ɺ౷߹͞Εͨ160Λ௒͑Διϑτ΢ΣΞ ύοέʔδ͕؆୯ʹ࢖͑ɺ ैྔ՝ۚܕͰ଴ͪ࣌ؒͳ͘ HPCϦιʔεΛ͙͢ʹɻ