make-your-deaplearning-apps-on-aws-201703

JAWS DAYS 2017 での登壇資料です。

Transcript

1. #jawsdays #jd2017_d Deep Learning on AWS Ͱ ࣗ෼͚ͩͷֶशϞσϧΛ࡞Ζ͏ʂ by JAWS-UG AI ࢧ෦
2. #jawsdays #jd2017_d தؙ ྑ @pottava SUPINF Inc. (AWS Certified) SA, DevOps Engineer Pro ❤ Amazon ECS, AWS Batch, IAM
3. #jawsdays #jd2017_d גࣜձࣾεϐϯϑ
4. #jawsdays #jd2017_d ͜ΜͳձࣾͰ͢ • AWS Λத৺ʹɺΫϥ΢υωΟςΟϒͳ։ൃɾӡ༻อकͷडୗ • Docker Ͱͷ։ൃɾӡ༻ ͕େಘҙ • GPU, Jupyter, Mesos.. Պֶٕज़ܭࢉ؀ڥͷߏஙɾอकडୗ • ͦͷଞ Web αʔϏεɾωΟςΟϒΞϓϦɾۀ຿γεςϜɾɾ
5. #jawsdays #jd2017_d ૣ଎ɺσΟʔϓϥʔχϯά
6. #jawsdays #jd2017_d σϞ
7. #jawsdays #jd2017_d AWS ϗϫΠτϖʔύʔ͘Μ • σΟʔϓϥʔχϯάΛར༻ʢLSTM with MXNetʣ • AWS നॻɺPDF 27 ϑΝΠϧΛֶशσʔλʹར༻ • ग़ͩ͠ͷ਺୯ޠ͔ΒɺͦΕΒ͍͠ޙଓͷจষΛࣗಈੜ੒
8. #jawsdays #jd2017_d σϞͷྲྀΕ 1. ·ͣҰ౓ʢֶशෆे෼ͳʣਪ࿦͕ಈ͘͜ͱΛ֬ೝ 2. ͦͷޙվΊͯɺσʔλͷऔಘɾՃ޻ 3. ֶश 4. ৽൛ਪ࿦αʔϏεσϓϩΠ
9. #jawsdays #jd2017_d ਪ࿦ͷ༷ࢠ ʮAmazon EC2 isʯΛ༩͑ͨ࣌ͷɺֶशෆे෼ͳਪ࿦ྫ ৽͍͠ݴޠ͕ੜ·Ε·ͨ͠ɻ
10. #jawsdays #jd2017_d ਪ࿦ͷ༷ࢠ ʮAmazon EC2 isʯͰɺͦΕͳΓʹֶशͨ͠ޙͷਪ࿦ྫ ݴ͍͍ͨ͜ͱ͸Θ͔Γ·ͤΜ͕ɺ୯ޠ΍จ๏͸ਵ෼੒௕͠·ͨ͠ɻ
11. #jawsdays #jd2017_d γεςϜߏ੒ AWS Batch S3 2.ֶशδϣϒ౤ೖ ΤϯδχΞ SpotFleet ECR DeepLearning AMI v2 1.ֶशσʔλ౤ೖ 3.δϣϒεέδϡʔϧ 4.σʔλऔಘ & ֶश 5.݁ՌϞσϧΛอଘ ҰൠϢʔβ AWS Lambda 7.Ϟσϧऔಘ & ਪ࿦ APIGateway 6.ਪ࿦ϦΫΤετ 8.݁ՌԠ౴ ECS EC2 g2.2xlarge EC2 g2.2xlarge
12. #jawsdays #jd2017_d ΄΅શ෦ެ։͠·ͨ͠ • PDF 1 ຕఔ౓ͳΒ 10 ෼Ͱࢼͤ·͢ • AWS ΁ͷσϓϩΠπʔϧ܈͸Ұ෦ܝࡌ ʢશ෦ࢿྉԽ͢Δͷ͸೉ّͳ΋ͷͰ..ʣ https://github.com/pottava/mxnet-char-lstm
13. #jawsdays #jd2017_d ͳΔ΄ͲΘ͔ΒΜ • ެ։͞Εͨࢿྉ͚ͩͰ͸Α͘෼͔Βͳ͍ɻ͸͍ɾɾ • Qiita Ͱ΋৘ใൃ৴͠·͢ • σΟʔϓϥʔχϯάͰ΋ AWS ͷઃܭͰ΋ • Twitter ͳͲͰฉ͍ͯ΋Β͑Ε͹༨ྗ͋Δͱ͖౴͑·͢ :) • ͓ؾܰʹ @pottava ·Ͱ
14. #jawsdays #jd2017_d ৼΓฦΓ
15. #jawsdays #jd2017_d Ͳ͏࡞ͬͨͷ͔ 1. PDF Λద౰ʹμ΢ϯϩʔυʢσʔλऩूɺ5 ෼ʣ 2. ֶश͠΍͍͢Α͏ʹσʔλΛՃ޻ʢલॲཧɺ30 ෼ʣ 3. LSTM ωοτϫʔΫʹ౤ೖʢֶशɺ5 ࣌ؒʣ 4. REST API Λ࣮૷ʢਪ࿦ɺ45 ෼ʣ ʢ + ຊࢫͰͳ͍ AWS ͷઃܭɺϑϨʔϜϫʔΫͷΩϟονΞοϓʹ૬Ԡͷ࣌ؒʣ
16. #jawsdays #jd2017_d ͳͥΫϥ΢υͰʁ • ֶशσʔλ΍࡞੒ͨ͠Ϟσϧ͸ൺֱతେ͖ͳϑΝΠϧ • ֶशɺྫ͑͹Ұൠ޲͚ PC ͩͱͦΕ͚ͩͰ਺೔͔͔Δ • Ұ࣌తʹɺ๲େͰߴੑೳͳܭࢉೳྗ͕࢖͑Δ • ਪ࿦αʔϏεͷσϓϩΠɾӡ༻΋ͱͯ΋؆୯
17. #jawsdays #jd2017_d Կ͕Ͳ͏ػցֶशͩͬͨͷʁ ❌ʮ͜ͷ୯ޠ͕ೖྗ͞ΕͨΒɺ͜͏ग़ྗͯ͠ʯ ⭕ʮେྔʹಡΜͩจষΛੜ͔ͯ͠ɺ࣍ͷ୯ޠΛ༧ଌͯ͠Ͷʯ
18. #jawsdays #jd2017_d Կ͕Ͳ͏σΟʔϓͩͬͨͷʁ จষͷࣗಈੜ੒ʹ࢖ͬͨػցֶशͷख๏͕ LSTM ͱ͍͏ ϦΧϨϯτχϡʔϥϧωοτϫʔΫ͔ͩͬͨΒɻ ʁʁʁ
19. #jawsdays #jd2017_d σΟʔϓϥʔχϯά ʢਂ૚ֶश / DLʣ.
20. #jawsdays #jd2017_d σΟʔϓϥʔχϯάͱ͸ • ଟ૚ߏ଄ͷχϡʔϥϧωοτϫʔΫΛ༻͍ͨػցֶशͷ͜ͱ • χϡʔϥϧωοτϫʔΫͱ͍͏ಠಛͳܗͰอ࣋͞Εͨ • ҎԼͷ a ͱ b ΛܾΊΔ͜ͱֶ͕श ɹy = ax + b • ͦΕΛ΋ͱʹ x ͕͖ͨΒ y ͱฦ͢ͷ͕ɺਪ࿦
21. #jawsdays #jd2017_d σΟʔϓϥʔχϯάͱ͸ • ࣮ࡍ͸๲େͳม਺ͦΕͧΕʹ࠷ద஋Λ୳͠ʮϞσϧʯΛੜ੒ • Ұͭͷ૚ͷɺҰͭͷϊʔυ͚ͩͰ΋ม਺͸ͨ͘͞Μ ɹy = f ( w1x1 + w2x2 + w3x3 + … + wnxn + b )
22. #jawsdays #jd2017_d Α͘ݟΔ DL ͷྫ • Amazon Rekognition ͷҰൠ෺ମݕग़͸ɺ͖ͬͱ CNNɻ • ৞ΈࠐΈχϡʔϥϧωοτϫʔΫʢCNNʣ ෳ਺ͷ૚͕͋Γ·͢Ͷʙ
23. #jawsdays #jd2017_d Ͱ͸
24. #jawsdays #jd2017_d Ͳ͏΍ͬͯϞσϧΛ࡞Δͷ͔
25. #jawsdays #jd2017_d ࢼߦࡨޡ͢ΔͨΊͷ
26. #jawsdays #jd2017_d Ϋϥ΢υͷ౷߹ݚڀ؀ڥ
27. #jawsdays #jd2017_d Jupyter as a Service • ࢼߦࡨޡɺݚڀͷڞ༗ͷελϯμʔτͱ͍͑͹ Jupyter • ࠷ۙྲྀߦͷػೳ֦ுɿΫϥ΢υͷσʔλετΞͱͷ࿈ܞ • σʔλͷநग़ɺม׵ɾՃ޻ɺͦͷऔΓࠐΈʢETLʣ • AWS ͸Ұาग़஗Ε͍ͯΔ΋ͷͷɾɾ
28. #jawsdays #jd2017_d Google Cloud Datalab
29. #jawsdays #jd2017_d Google Cloud Datalab • TensorFlow ✖ BigQuery ͷྫ • ͦͷૢ࡞ UI ͱͯ͠ Jupyter ɹϕʔεͷ Cloud Datalab • ϩʔΧϧ൛·Ͱ͋Δɾɾ
30. #jawsdays #jd2017_d Azure ML Studio
31. #jawsdays #jd2017_d ͱʹ͔͘ੑೳॏࢹ or ࣗ༝ʹ ϑϨʔϜϫʔΫΛબͼ͍ͨͳΒ
32. #jawsdays #jd2017_d GPU ΍ Infiniband ΛؚΉ ϚϧνΫϥ΢υ HPC ؀ڥ ΦϯϓϨͱͷϋΠϒϦοτ΋ɻ ࢖͑ΔϑϨʔϜϫʔΫଟ਺ʂʂ
33. #jawsdays #jd2017_d ฒྻԋࢉʹΑΔֶश
34. #jawsdays #jd2017_d ֶशϑΣʔζ͸ HPC • ػցֶशͰͷֶश͸ɺ๲େͳύϥϝλΛܭࢉ͢Δ͜ͱ • ຊ֨తʹͳΔఔɺखڐͷ൚༻୺຤Ͱ͸౸ఈܭࢉͰ͖ͳ͍
35. #jawsdays #jd2017_d όονίϯϐϡʔςΟϯά • ແݶʹεέʔϧ͢Δߴੑೳ / GPU ౥ࡌͷΫϥελ܈ • ࣮ߦॱংʢґଘʣͷ͋ΔδϣϒΛฒྻ෼ࢄॲཧ • ͦΕͰ΋ऴΘΒͳֶ͍शɾɾ • AWS ͸͜ͷྖҬ͕͍·೤͍ʂʂ
36. #jawsdays #jd2017_d AWS Batch • Ϛωʔδυ όονίϯϐϡʔςΟϯά؀ڥ • Ϋϥελ؅ཧɺΩϡʔɺεέδϡʔϥΛ AWS ʹ͓೚ͤ • Ωϡʔ͕ਂ͘ͳΔͱࣗಈͰεέʔϧʂʂ • SpotFleet ͱ౷߹͞Ε͍ͯΔʂ • Docker ϕʔεʹͳ͓ͬͯΓɺLinux ܥͷॲཧ͸ͳΜͰ΋ಈ͘
37. #jawsdays #jd2017_d AWS Batch • ࠓճͷΑ͏ʹɺGPU Λ࢖͍͍ͨ / ಠࣗ AMI Λ࢖͍͍ͨ • ΋͘͠͸ t2 ܥΠϯελϯεͰࢼ͍ͨ͠ • ͦΜͳ࣌͸ʮUnmanaged ؀ڥʯΛબ୒͢Ε͹ OKʂ • ASG ΍ SpotFleet ͷ্Ͱ ECS Λಈ͔ͤ͹े෼շదͰ͢
38. #jawsdays #jd2017_d AWS Batch ʢҙ֎ʁͳར༻ྫʣ S3 ʹ͋͛ͨେྔը૾ Rekognition Ͱͷ Ұൠ෺ମݕग़Λ AWS Batch ͔Β࣮ߦɻ
39. #jawsdays #jd2017_d ϋϯζΦϯ͋Γ㽂 ΈΜͳେ޷͖ɺCOBOL ͷఆྫόον Λ AWS Batch Ͱ
40. #jawsdays #jd2017_d αʔϏε඼࣭͕ॏཁɺਪ࿦
41. #jawsdays #jd2017_d ਪ࿦ϑΣʔζ͸ 24 / 365 • εϚϗɺIoT ػثͰਪ࿦Ͱ͖ΔΑ͏ʹͳ͖ͬͯͨݱ୅ • Մ༻ੑͱϨΠςϯγ͕ॏཁɺҰൠͷαʔϏεͱಉ͡ • EC2 Ͱ API Λఏڙ͢Δ͜ͱ΋΋ͪΖΜͰ͖Δ͕ • ͦΕɺLambda Ͱ΋ಈ͘Α
42. #jawsdays #jd2017_d AWS Lambda http://aws.typepad.com/sajp/2017/01/seamlessly- scale-predictions-with-aws-lambda-and-mxnet.html
43. #jawsdays #jd2017_d AMI for DeepLearning !! • ͍΍ɺ΍ͬͺΓֶशɾਪ࿦ͱ EC2 ͕͍͍ͱ͍͏ํɺ࿕ใͰ͢ • Batch ΍ ECS Ͱ΋༗༻ɺશ෦ೖΓ AMI ʂʂ • ࠷৽ͷ TensorFlow, MXNet, ɹCaffe, CNTK, Theano, Torchɻɻ ɹAnaconda ΋ೖͬͯ·͢ʂ
44. #jawsdays #jd2017_d DeepLearning AMI • ࠓճͷσϞͰ΋࢖͍·ͨ͠ • GPU Λ࢖͏࣌ͷ೉ؔɺNVIDIA υϥΠό΋ηοτΞοϓࡁ • docker & ecs-agent Λ user-data ͰΠϯετʔϧ͢Ε͹ ɹAWS Batch ΍ Amazon ECS Ͱ΋؆୯ʹಈ࡞͠·͢
45. #jawsdays #jd2017_d JAWS-UG AI ࢧ෦
46. #jawsdays #jd2017_d ίϯςϯπ • AWS Ͱ AI αʔϏεΛ࣮૷ɾӡ༻͢ΔͨΊͷ ɹҰൠతͳٕज़৘ใɺ஌ݟɺࣄྫڞ༗ͷ৔ • ͢Ͱʹ׆༻͍ͯ͠Δํ • ಋೖΛݕ౼͍ͯ͠Δํ • ԿͦΕ͓͍͍͠ͷʁͳํʢ։࠵͝ͱʹ೉қ౓͕ଟগҧ͍·͢ʣ
47. #jawsdays #jd2017_d ࠓ݄຤εϖγϟϧճ։࠵ʂ • 3 / 31 (ۚ) 19:00 ʙ • HPC ઐ໳ࢧ෦ͱͷ߹ಉΠϕϯτ • AWS ຊ͔ࣾΒήετ 2 ໊ʂʂ
48. #jawsdays #jd2017_d ӡӦϝϯόʔ
49. #jawsdays #jd2017_d Thank you for coming! ࢀߟจݙ: • AWS Batch – ؆୯ʹ࢖͑ͯޮ཰తͳόονίϯϐϡʔςΟϯάػೳ – AWS https://aws.amazon.com/jp/batch/ • AWS Black Belt Online SeminarʮAWS Batchʯͷࢿྉ͓ΑͼQAެ։ http://aws.typepad.com/sajp/2017/02/aws-black-belt-online-seminar-aws- batch.html#QCPzBdn.twitter_tweet_count_m • re:Invent 2016: AWS Big Data & Machine Learning Sessionsɻ https://aws.amazon.com/blogs/big-data/reinvent-2016-aws-big-data- machine-learning-sessions/