reinvent-ml-mini-con - Speaker Deck

reinvent-ml-mini-con

by ryo nakamaru

Slide 1

Slide 1 text

AWS Ͱ࢝ΊΔ DeepLearning re:Invent 2016 Machine Learning Mini Con ࢀՃใࠂ JAWS-UG AI ࢧ෦ @ 2016.12.09

Slide 2

Slide 2 text

@pottava SUPINF Inc.

Slide 3

Slide 3 text

͍͖ͳΓͰ͕͢

Slide 4

Slide 4 text

FizzBuzz ஌͍ͬͯΔਓʙʁ

Slide 5

Slide 5 text

FizzBuzz ॻ͚Δਓʙʁ

Slide 6

Slide 6 text

ɹfor i in range(1,101): ɹ if i % 15 == 0: ɹ print 'FizzBuzz' ɹ elif i % 3 == 0: ɹ print 'Fizz' ɹ elif i % 5 == 0: ɹ print 'Buzz' ɹ else: ɹ print i ɹͰ͢Ͷɺྫ͑͹ɻ

Slide 7

Slide 7 text

Ͱ͸

Slide 8

Slide 8 text

ػցֶशͰ FizzBuzz ղ͚Δਓʙʁ

Slide 9

Slide 9 text

ʁʁʁ

Slide 10

Slide 10 text

ίϯϐϡʔλʹσʔλΛ౉ͯ͠ ύλʔϯΛݟ͚ͭͤ͞Δ

Slide 11

Slide 11 text

Ξϓϩʔν Ͳ͏ղ͔͘ ໋ྩత ɹfor i in range(1,101): ɹ if i % 15 == 0: ɹ print 'FizzBuzz' ɹ elif i % 3 == 0: ɹ print 'Fizz' ɹ elif i % 5 == 0: ɹ print 'Buzz' ɹ else: ɹ print i ػցֶश େྔͷσʔλΛ౉͠... ༧ଌਫ਼౓ΛߴΊΔ... parameter result 6 Fizz 7 7 10 Buzz 30 FizzBuzz …

Slide 12

Slide 12 text

Ҏ্ɺͱ͋ΔϫʔΫγϣοϓͰͷ ΞΠεϒϨΠΫͰͨ͠ɻ

Slide 13

Slide 13 text

ຊ୊

Slide 14

Slide 14 text

AWS Ͱ࢝ΊΔ DeepLearning

Slide 15

Slide 15 text

ࠓ೔ͷ࿩୊ ɾ࠷ۙ͋ͬͨ Amazon ػցֶशܥχϡʔεͱ DL ɾre:Invent Machine Learning Mini Con ใࠂ ɾMXNet ʹ͍ͭͯ ɾϫʔΫγϣοϓͷ༷ࢠͱ࣮ྫ

Slide 16

Slide 16 text

Amazon ػցֶशܥχϡʔε

Slide 17

Slide 17 text

ɾAmazon Echo ɾAmazon Go AI ΧϯύχʔɺAmazon.com

Slide 18

Slide 18 text

Amazon Echo ɾAmazonʮୈ࢛ͷऩӹͷபʯ͸ 2020 ೥·Ͱʹ 110 ԯυϧՔ͙ ɹਓ޻஌ೳΞγελϯτʮAlexaʯͱ ɹԻ੠ίϯτϩʔϥʔͷʮEchoʯ ɹhttp://thebridge.jp/2016/09/amazon-echo-alexa-add-11-billion-in-revenue-by-2020-2016-9-pickupnews ɾhttps://www.amazon.jobs/en/teams/alexa ɾre:Invent Ͱ͸ echo dot ͕ࢀՃऀʹ഑ΒΕ·ͨ͠

Slide 19

Slide 19 text

Amazon Echo

Slide 20

Slide 20 text

Amazon Echo Alexa ʹԻ੠Ͱ͓ئ͍ɾ࣭໰͢ΔͨΊͷσόΠεɻ ʢAlexa ͸ Amazon ͕։ൃͨ͠ AI ʣ ʮΞϨΫαɺUber ΛݺΜͰɻࠓ೔ͷఱؾ͸ʁ ʯ ʮΞϨΫαɺ͜ͷۂͷԋ૗ऀ͸୭ʁԻྔΛ্͛ͯʯ

Slide 21

Slide 21 text

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

Slide 22

Slide 22 text

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

Slide 23

Slide 23 text

ɾAmazon Echo ɾAmazon Go AI ΧϯύχʔɺAmazon.com

Slide 24

Slide 24 text

Amazon Go

Slide 25

Slide 25 text

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

Slide 26

Slide 26 text

Amazon Rekognition ଞࣾͰ΋੝Μͳ Computer vision API ͷҰछɻ

Slide 27

Slide 27 text

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

Slide 28

Slide 28 text

Amazon Rekognition ΍ͬͯΈͨ

Slide 29

Slide 29 text

Amazon Rekognition ฐࣾ༐ऀͷ ྨࣅ౓൑ఆɻ

Slide 30

Slide 30 text

re:Invent Machine Learning Mini Con

Slide 31

Slide 31 text

Machine Learning Mini Con ɾػցֶशܥͷηογϣϯ / ϫʔΫγϣοϓ ɾhttp://bit.ly/reinvent-2016-ml ɾࠓ೥͸ 17 ηογϣϯ ɾϫʔΫγϣοϓҎ֎͸ YouTube ͰݟΕ·͢

Slide 32

Slide 32 text

ೖ໳ฤ ɾMAC201: Amazon Mechanical Turk Λ࢖ͬͯҰൠతಛ௃Λ͔ͭΉ ɾMAC202: Alexa ʹ͓͚Δਂ૚ֶश ɾMAC203: Amazon Rekognition ͷ͝঺հ ɾMAC204: Amazon Polly ͷ͝঺հ ɾMAC205: Ϋϥ΢υΒ͘͠εέʔϧ͢Δਂ૚ֶश: ɹɹɹɹɹ AWS Ͱ Caffe ΛεέʔϧΞοϓͯ͠ϏσΦݕࡧΛվળ͢Δ ɾMAC206: ػցֶशͷݱঢ়

Slide 33

Slide 33 text

தڃฤ ɾMAC301: ਂ૚ֶशͰ޻৔ͷϓϩηεΛม͍͑ͯ͘ ɾMAC302: ෆಈ࢈Ͱͷઓུత༏ҐͷͨΊʹ Amazon ML, Redshift, S3 σʔλϨΠΫΛ׆༻͢Δ ɾMAC303: Amazon EMR ͱ Apache Spark ͰΫϥε෼ྨͱ ϨίϝϯσʔγϣϯΤϯδϯΛ։ൃ͢Δ ɾMAC304: Amazon Lex ͷ͝঺հ ɾMAC306: MXNet Λ࢖ͬͯϨίϝϯσʔγϣϯϞσϧΛߏங͢Δ ɾMAC306-R: MXNet Λ࢖ͬͨਂ૚ֶश

Slide 34

Slide 34 text

தڃฤ ɾMAC307: Predicting Customer Churn with Amazon ML ɾMAC308: ϫʔΫγϣοϓ: Amazon Lex, Amazon Polly ͦͯ͠ Amazon Rekognition Λ࢖ͬͨϋϯζΦϯ ɾMAC309: Amazon Polly ͱ Amazon Lex ͷ͝঺հ

Slide 35

Slide 35 text

্ڃฤ ɾMAC401: Scalable Deep Learning Using MXNet ɾMAC403: Automatic Grading of Diabetic Retinopathy ɹɹɹɹɹ through Deep Learning

Slide 36

Slide 36 text

ৄࡉ͸ YouTube ͱ Slideshare Ͱ

Slide 37

Slide 37 text

ϐοΫΞοϓ ɾMAC201: Amazon Mechanical Turk Λ࢖ͬͯҰൠతಛ௃Λ͔ͭΉ ɾMAC206: ػցֶशͷݱঢ় ɾMAC306: MXNet Λ࢖ͬͯϨίϝϯσʔγϣϯϞσϧΛߏங͢Δ ɾMAC401: Scalable Deep Learning Using MXNet

Slide 38

Slide 38 text

MAC201 Mechanical Turk Ͱػցֶश༻σʔλΛ࡞Δ ɾhttps://www.youtube.com/watch?v=vRtLdeNl7Tg ɾେྔͷɺߴ඼࣭ͳσʔληοτ͸ूΊʹ͍͘ ɾϝΧχΧϧλʔΫʹͦͷ࡞੒Λґཔ͢Δ

Slide 39

Slide 39 text

MAC206 Amazon ૑ۀ͔Β࠷৽ AI αʔϏε·Ͱ঺հ ɾhttps://www.youtube.com/watch?v=HqsUfyu0XJc ɾDeep Learning AMI, MXNet, Alexa ͳͲͳͲ.. ɾޙ൒͸ܯ࡯ʹαʔϏεఏڙ͢ΔϞτϩʔϥͷࣄྫ

Slide 40

Slide 40 text

MXNet

Slide 41

Slide 41 text

ֶशϑϨʔϜϫʔΫ ͲΕ͕͓޷ΈͰ͔͢ɾɾʁ MXNet / TensorFlow / Caffe / Chainerɻ ɾͲͷχϡʔϥϧωοτ࢖͏ͷʁCNNʁRNNʁ ɾGPU ࢖͏ͷʁCPU ͚ͩʁෳ਺ϊʔυ࢖͏ʁ ɾࠃ࢈ΛԠԉʁ

Slide 42

Slide 42 text

AWS ͸ MXNet Ұ୒ײ͋Δ ɾ͑ɺAmazon DSSTNE ɾɾ ɾͱ͸͍͑ଞͷ΋ݕ౼͍ͨ͠ํ͸ͪ͜Β ɹ CMP314: Bringing Deep Learning to the Cloud with Amazon EC2 https://www.youtube.com/watch?v=34Xorby_pyw

Slide 43

Slide 43 text

MAC306 Netﬂix ͷϨίϝϯυྫΛ௨ͯ͡ DL / MXNet Λৄઆ ɾhttps://www.youtube.com/watch?v=cftJAuwKWkA ɾDeep Learning ͷॳา͔Βɻͱͯ΋෼͔Γ΍͍͢ ɾGitHub ͷ MXNet ϦϙδτϦʹ͋ΔαϯϓϧΛσϞ https://github.com/dmlc/mxnet/tree/master/example/recommenders

Slide 44

Slide 44 text

ϫʔΫγϣοϓͷ༷ࢠͱ࣮ྫ

Slide 45

Slide 45 text

ϫʔΫγϣοϓʁ ϋϯζΦϯܗ͕ࣜଟ͍ɻάϧʔϓϫʔΫ΋͋ͬͨΓɻ ɾ࣮ࡍʹखΛಈ͔͢ͷͰͱͯ΋ཧղ͕ਐΉ ɾ·ΘΓͷࢀՃऀͱͷίϛϡχέʔγϣϯ .. !! ɾre:Invent ʹߦ͘ͳΒ௨ৗηογϣϯΑΓΦεεϝ

Slide 46

Slide 46 text

MAC401 ECS ্Ͱ MXNet ʹΑΔ DL ͷֶशɾਪ࿦Λମݧ ɾECS ͷ Runtask + CPU ͷΈ ɾGitHub ͷ awslabs ϦϙδτϦΛར༻ https://github.com/awslabs/ecs-deep-learning-workshop/

Slide 47

Slide 47 text

ࢼ͢ͷ͸ͱͯ΋؆୯ CloudFormation ʹΑΔ EC2 / ECS ౳ੜ੒ɻͦͷޙ.. ɾLab 3: ECS Ͱ MXNet ͷ Jupyter notebook ىಈ ɾLab 4: MXNet ʹΑΔը૾ͷΫϥε෼ྨ ɾLab 5: ECS λεΫͱͯ͠ը૾ΛΫϥε෼ྨ

Slide 48

Slide 48 text

Deep Learning AMI http://qiita.com/pottava/items/c79117089be2406b127f

Slide 49

Slide 49 text

͓஌Βͤ

Slide 50

Slide 50 text

དྷि͸ JAWS-UG ίϯςφࢧ෦

Slide 51

Slide 51 text

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

Slide 52

Slide 52 text

Amazon ECS ɾࠓ೥͸ ECS ͰδϣϒΛ૸ΒͤΔηογϣϯ͕ෳ਺ ɾMXNet on ECS ͷϫʔΫγϣοϓ͸ੈքͰ΋ਓؾ ɾECS Ϋϥελ্Ͱ MXNet ͷֶशɾਪ࿦

Slide 53

Slide 53 text

AWS Batch ɾECS ্ʹ HPC ۀքͷҙຯʹ͍ۙΫϥελΛߏஙɻ ɾδϣϒεέδϡʔϥ ≠ ίϯςφք۾ͷεέδϡʔϥ ɾECS ্ͳͷͰɺ࣮͸ίϯςφϕʔε ɾGlue ΍ EFS ͱͷ૊Έ߹Θͤ௒ॏཁ

Slide 54

Slide 54 text

͓ΘΓ

Slide 55

Slide 55 text

גࣜձࣾεϐϯϑ ΞΠσΟΞΛ͔ͨͪʹʂ +

Slide 56

Slide 56 text

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

Slide 57

Slide 57 text

͝૬ஊ͸͓ؾܰʹͪ͜Β·Ͱ.. 57