Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Build Image Classification service with Amazon ...

Avatar for Yuichiro Someya Yuichiro Someya
November 22, 2016
Programming
2.9k
4

Build Image Classification service with Amazon ECS and GPU instances

Avatar for Yuichiro Someya

Yuichiro Someya

November 22, 2016

More Decks by Yuichiro Someya

See All by Yuichiro Someya
にんげんがさき 基盤はあと / Developers over ML platform
Avatar for Yuichiro Someya ayemos
0
15k
機械学習をスモールスタートさせる方法 / small machine learning
Avatar for Yuichiro Someya ayemos
3
2.1k
アットホームな分析基盤の作り方 / Homemade Machine Learning Toolkits
Avatar for Yuichiro Someya ayemos
1
1k
サービス開発、機械学習、クラウド / the trinity of machine learning
Avatar for Yuichiro Someya ayemos
0
3.7k
成長を止めない機械学習のやり方 / Don't stop 'til you get enough (data).
Avatar for Yuichiro Someya ayemos
15
5.3k
AWS で加速する機械学習 / Accelerate Machine Learning with AWS
Avatar for Yuichiro Someya ayemos
1
370
クックパッドの機械学習基盤 2018 / Machine Learning Platform at Cookpad ~ 2018 ~
Avatar for Yuichiro Someya ayemos
15
21k
PyTorchとCaffe2とONNXと深層学習モデルのデプロイについて
Avatar for Yuichiro Someya ayemos
1
3.1k
クックパッドにおけるAWS GPUインスタンスの利用事例 / Powering by AWS GPU Instances in Cookpad Inc
Avatar for Yuichiro Someya ayemos
0
460

Other Decks in Programming

See All in Programming
LLMは4年分のCompose移行を再現できるのか？実プロダクト279件のXMLで探る自動化の境界線
Avatar for makun makun
0
260
Loosening the Reins: Go Generics Get More Flexible
Avatar for kuro kuro_kurorrr
0
360
FastAPI の並行処理モデルを完全に理解する
Avatar for Yuki Ishikawa hoto17296
9
3.4k
[PyCon KR 2026] More Variants, More Diversity for AI Accelerators
Avatar for Joongi Kim achimnol
0
130
Can LLMs Replicate 4 Years of Compose Migration? Exploring the boundaries of automation with 279 XML files from a real product
Avatar for makun makun
0
120
PHPプロジェクトの結合バランスを可視化する #php_night
Avatar for Takuma Kajikawa kajitack
0
190
自分的「カンファレンスの楽しみ方」
Avatar for syumai syumai
0
170
数年滞っていたダークモード対応をおよそ2週間で完了させる
Avatar for Kazuki Chigita chigichan24
0
210
Claude Codeを組織的に動かして月400PRを実現した話
Avatar for happy_ryo happy_ryo
0
240
Hello, Hiroshima Geospatial Data! — Exploring DoboX with Python
Avatar for ぴっかりん ra0kley
0
150
AI駆動開発にグラフDBを重ねてみた
Avatar for Satoshi Kaneyasu satoshi256kbyte
2
630
go-spidermonkeyでAIエージェントのCode Modeを実装する
Avatar for syumai syumai
3
1.5k

Featured

See All Featured
Chasing Engaging Ingredients in Design
Avatar for Sebastian Deterding codingconduct
0
290
How to build a perfect <img>
Avatar for Jono Alderson jonoalderson
1
5.9k
Refactoring Trust on Your Teams (GOTO; Chicago 2020)
Avatar for Rebecca Miller-Webster rmw
35
3.8k
Lightning Talk: Beautiful Slides for Beginners
Avatar for Ines Montani inesmontani
PRO
2
660
What the history of the web can teach us about the future of AI
Avatar for Ines Montani inesmontani
PRO
1
680
Into the Great Unknown - MozCon
Avatar for Noah Learner thekraken
41
2.7k
Marketing to machines
Avatar for Jono Alderson jonoalderson
1
5.7k
Data-driven link building: lessons from a $708K investment (BrightonSEO talk)
Avatar for Szymon Słowik szymonslowik
1
1.3k
How People are Using Generative and Agentic AI to Supercharge Their Products, Projects, Services and Value Streams Today
Avatar for Helen Beal helenjbeal
1
300
We Have a Design System, Now What?
Avatar for Morgane Peng morganepeng
55
8.3k
A Modern Web Designer's Workflow
Avatar for Chris Coyier chriscoyier
698
190k
Pawsitive SEO: Lessons from My Dog (and Many Mistakes) on Thriving as a Consultant in the Age of AI
Avatar for David Carrasco davidcarrasco
0
230

Transcript

  1. Build Image Classification service with AWS ECS and GPU instances

    Yuichiro Someya @ Cookpad

  2. • છ୩ ༔Ұ࿠ [Yuichiro Someya] • ౦޻େେֶӃ ܭࢉ޻ֶઐ߈ म࢜ •

    '16 ৽ଔ @ ΫοΫύου • github.com/ayemos • twitter.com/kumasan_com echo `whoami`

  3. • ྉཧࣸਅͷࣗಈऩूαʔϏεΛ • CaffeNet[1]Λݩʹ࡞ͬͨྉཧը૾ೝࣝϞσϧͱ • Amazon SQS/S3 Ͱߏங͞Εͨσʔλϑϩʔͱ • Amazon

    ECS (GPU instance) Λར༻ͯ͠ӡ༻͍ͯ͠Δ࿩ <>IUUQTHJUIVCDPN#7-$DB⒎F Agenda

  4. • ྉཧࣸਅͷࣗಈऩूαʔϏεΛ • CaffeNet[1]Λݩʹ࡞ͬͨྉཧը૾ೝࣝϞσϧͱ • Amazon SQS/S3 Ͱߏங͞Εͨσʔλϑϩʔͱ • Amazon

    ECS (GPU instance) Λར༻ͯ͠ӡ༻͍ͯ͠Δ࿩ <>IUUQTHJUIVCDPN#7-$DB⒎F Agenda

  5. ΫοΫύου • Ϩγϐ਺ɿ 250ສ඼Ҏ্ • ݄࣍ؒར༻ऀ਺ɿ 6,000ສਓҎ্

  6. • εϚϗ಺ͷࣸਅ͔Βྉཧ͚ͩΛࣗಈతʹऩू • Ұ෦ͷϢʔβʔ޲͚ʹݶఆతʹެ։த ྉཧ͖Ζ͘

  7. • CaffeNetΛ ྉཧʗඇྉཧ ൑ఆ޲͚ʹFine Tuningͨ͠Ϟσϧ • Caffe[1]Ͱֶश͞ΕͨϞσϧΛChainerͷCaffe emulatorͰಡΉ
 ref: http://docs.chainer.org/en/stable/reference/caffe.html

    • ෼ྨΧςΰϦΛ ྉཧʗඇྉཧ ʹมߋ͠ɺΫοΫύου্ͷ
 ྉཧࣸਅΛ࢖ֶͬͯश <>IUUQDB⒎FCFSLFMFZWJTJPOPSH CookpadNet

  8. • CookpadNet͸Ͳ͜Ͱ൑ఆΛߦ͍ɺͦͷ݁Ռ͸Ͳ͜ʹͲ͏఻͑Δ ͷ͔ʁ • ൑ఆϞσϧΛΫϥΠΞϯτʹஔ͍ͯ൑ఆ • ϞσϧαΠζ͕େ͖͍(100MB~)ͷͰɺݱ࣮తͰͳ͍ • (αΠζͷখ͍͞ϞσϧΛݚڀத) •

    ൑ఆΛߦ͏ίϯϙʔωϯτΛ֎෦ʹஔ͘ • HTTP Serverʁ σʔλϑϩʔʗϫʔΫϑϩʔ

  9. $MJFOU "OESPJE J04 "1*4FSWFS SVCZ $MBTTJpDBUJPO4FSWFS QZUIPO DIBJOFS

  10. $MJFOU "OESPJE J04 "1*4FSWFS SVCZ 1045DMBTTJGZ\QIPUPCJOBSZ^ $MBTTJpDBUJPO4FSWFS QZUIPO DIBJOFS

  11. $MJFOU "OESPJE J04 "1*4FSWFS SVCZ 1045DMBTTJGZ\QIPUPCJOBSZ^ $MBTTJpDBUJPO4FSWFS QZUIPO 1045DMBTTJGZ\QIPUPCJOBSZ^ DIBJOFS

  12. $MJFOU "OESPJE J04 "1*4FSWFS SVCZ 1045DMBTTJGZ\QIPUPCJOBSZ^ $MBTTJpDBUJPO4FSWFS QZUIPO 1045DMBTTJGZ\QIPUPCJOBSZ^ SFTVMU\JT@GPPECPPM^

    DIBJOFS

  13. $MJFOU "OESPJE J04 "1*4FSWFS SVCZ 1045DMBTTJGZ\QIPUPCJOBSZ^ $MBTTJpDBUJPO4FSWFS QZUIPO 1045DMBTTJGZ\QIPUPCJOBSZ^ SFTVMU\JT@GPPECPPM^

    DIBJOFS SFTVMU\JT@GPPECPPM^

  14. $MJFOU "OESPJE J04 "1*4FSWFS SVCZ 1045DMBTTJGZ\QIPUPCJOBSZ^ $MBTTJpDBUJPO4FSWFS QZUIPO 1045DMBTTJGZ\QIPUPCJOBSZ^ SFTVMU\JT@GPPECPPM^

    DIBJOFS ը૾ͷΞοϓϩʔυ ը૾ॲཧ ൑ఆ SFTVMU\JT@GPPECPPM^

  15. $MJFOU "OESPJE J04 "1*4FSWFS SVCZ 1045DMBTTJGZ\QIPUPCJOBSZ^ $MBTTJpDBUJPO4FSWFS QZUIPO 1045DMBTTJGZ\QIPUPCJOBSZ^ SFTVMU\JT@GPPECPPM^

    DIBJOFS ը૾ͷΞοϓϩʔυ ը૾ॲཧ ൑ఆ SFTVMU\JT@GPPECPPM^ >>> 300~500 ms <<<

  16. • ը૾ॲཧͱϞσϧʹinferenceʹֻ͕͔ͦͦ࣌ؒ͜͜Δ
 (300~500ms) • APIαʔόʔ͔Βಉظతʹୟ͚ͳ͍
 (Unicorn ͷ worker͕ਚ͖ͯ͠·͏) • Amazon

    S3, SQSΛར༻ͨ͠ඇಉظͳ൑ఆॲཧϫʔΫϑϩʔ σʔλϑϩʔʗϫʔΫϑϩʔ

  17. $MJFOU "OESPJE J04 "1*4FSWFS SVCZ $MBTTJpDBUJPO8PSLFS QZUIPO DIBJOFS "NB[PO4 4UPSBHF

    "NB[PO424 2VFVF %#

  18. $MJFOU "OESPJE J04 "1*4FSWFS SVCZ $MBTTJpDBUJPO8PSLFS QZUIPO DIBJOFS "NB[PO4 4UPSBHF

    <6QMPBEQIPUPUPDMBTTJGZ> "NB[PO424 2VFVF %#

  19. $MJFOU "OESPJE J04 "1*4FSWFS SVCZ $MBTTJpDBUJPO8PSLFS QZUIPO DIBJOFS "NB[PO4 4UPSBHF

    <6QMPBEQIPUPUPDMBTTJGZ> "NB[PO424 2VFVF FORVFVF
 \LFZ@PO@TTUSJOH^ %#

  20. $MJFOU "OESPJE J04 "1*4FSWFS SVCZ $MBTTJpDBUJPO8PSLFS QZUIPO DIBJOFS "NB[PO4 4UPSBHF

    <6QMPBEQIPUPUPDMBTTJGZ> "NB[PO424 2VFVF FORVFVF
 \LFZ@PO@TTUSJOH^ EFRVFVF
 \LFZ@PO@TTUSJOH^ %#

  21. $MJFOU "OESPJE J04 "1*4FSWFS SVCZ $MBTTJpDBUJPO8PSLFS QZUIPO DIBJOFS "NB[PO4 4UPSBHF

    <6QMPBEQIPUPUPDMBTTJGZ> "NB[PO424 2VFVF FORVFVF
 \LFZ@PO@TTUSJOH^ EFRVFVF
 \LFZ@PO@TTUSJOH^ <%PXOMPBE*NBHF> %#

  22. $MJFOU "OESPJE J04 "1*4FSWFS SVCZ $MBTTJpDBUJPO8PSLFS QZUIPO DIBJOFS "NB[PO4 4UPSBHF

    <6QMPBEQIPUPUPDMBTTJGZ> "NB[PO424 2VFVF FORVFVF
 \LFZ@PO@TTUSJOH^ EFRVFVF
 \LFZ@PO@TTUSJOH^ 1045SFTVMU
 \LFZ@PO@TTUSJOH SFTVMU\JT@GPPE <%PXOMPBE*NBHF> %#

  23. $MJFOU "OESPJE J04 "1*4FSWFS SVCZ $MBTTJpDBUJPO8PSLFS QZUIPO DIBJOFS "NB[PO4 4UPSBHF

    <6QMPBEQIPUPUPDMBTTJGZ> 1045JT@QIPUP\LFZ@PO@TTUSJOH^ "NB[PO424 2VFVF FORVFVF
 \LFZ@PO@TTUSJOH^ EFRVFVF
 \LFZ@PO@TTUSJOH^ 1045SFTVMU
 \LFZ@PO@TTUSJOH SFTVMU\JT@GPPE <%PXOMPBE*NBHF> %#

  24. $MJFOU "OESPJE J04 "1*4FSWFS SVCZ $MBTTJpDBUJPO8PSLFS QZUIPO DIBJOFS SFTVMU\JT@GPPECPPM^ "NB[PO4

    4UPSBHF <6QMPBEQIPUPUPDMBTTJGZ> 1045JT@QIPUP\LFZ@PO@TTUSJOH^ "NB[PO424 2VFVF FORVFVF
 \LFZ@PO@TTUSJOH^ EFRVFVF
 \LFZ@PO@TTUSJOH^ 1045SFTVMU
 \LFZ@PO@TTUSJOH SFTVMU\JT@GPPE <%PXOMPBE*NBHF> %#

  25. $MJFOU "OESPJE J04 "1*4FSWFS SVCZ $MBTTJpDBUJPO8PSLFS QZUIPO DIBJOFS SFTVMU\JT@GPPECPPM^ "NB[PO4

    4UPSBHF <6QMPBEQIPUPUPDMBTTJGZ> 1045JT@QIPUP\LFZ@PO@TTUSJOH^ "NB[PO424 2VFVF FORVFVF
 \LFZ@PO@TTUSJOH^ EFRVFVF
 \LFZ@PO@TTUSJOH^ 1045SFTVMU
 \LFZ@PO@TTUSJOH SFTVMU\JT@GPPECPPM^^ <%PXOMPBE*NBHF> ඇಉظʹ൑ఆॲཧ

  26. • ྉཧࣸਅͷࣗಈऩूαʔϏεΛ • CaffeNet[1]Λݩʹ࡞ͬͨྉཧը૾ೝࣝϞσϧͱ • Amazon SQS/S3 Ͱߏங͞Εͨσʔλϑϩʔͱ • Amazon

    ECS Λར༻ͯ͠ӡ༻͍ͯ͠Δ࿩ <>IUUQTHJUIVCDPN#7-$DB⒎F Agenda

  27. • ECS: Amazon EC2 Container Service • Docker ContainerΛEC2Ͱߏ੒͞ΕͨΫϥελʹ഑ஔ(Task) •

    github.com/eagletmt/hako • ECSͷߏ੒ΛyamlϑΝΠϧͰ؅ཧ ECSͱGPUͱDockerͱ…

  28. "8471$ # cookpadnet-worker.yml scheduler: type: ecs region: ap-northeast-1 cluster: hako-production-g2

    desired_count: 1 app: image: cookpadnet-worker-gpu cpu: 128 memory: 3072 memory_reservation: 2048 env: AWS_REGION: ap-northeast-1 COOKPADNET_ENV: production ... %PDLFS3FHJTUSZ ։ൃऀ EPDLFSQVTI IBLPEFQMPZ &$4 EPDLFSQVMM 5BTL DPPLQBEOFUXPSLFS

  29. "8471$ # cookpadnet-worker.yml scheduler: type: ecs region: ap-northeast-1 cluster: hako-production-g2

    desired_count: 1 app: image: cookpadnet-worker-gpu cpu: 128 memory: 3072 memory_reservation: 2048 env: AWS_REGION: ap-northeast-1 COOKPADNET_ENV: production ... %PDLFS3FHJTUSZ ։ൃऀ EPDLFSQVTI IBLPEFQMPZ &$4 EPDLFSQVMM 5BTL DPPLQBEOFUXPSLFS DockerԽ͞ΕͨWorkerΛ
 hakoͰσϓϩΠ & ߏ੒؅ཧ

  30. w XPSLFSͰ͸(16Λ࢖༻ w ಉՁ֨ଳͷ$16Πϯελϯεͱൺ΂ͯ ഒͷੑೳࠩ w %PDLFS (16 GPU

  31. • Driver͕ඞཁ • nvidia-driverͷkernel module • ಉ͡όʔδϣϯͷuser-level drivers • Docker

    Container͔ΒGPU devicesΛૢ࡞͢Δҝ
 Containerʹద੾ͳLinux Capabilityͷઃఆ͕ඞཁ Ծ૝Խ v.s. Χʔωϧ

  32. ubuntu EPDLFSDPOUBJOFS ཧ૝ OWJEJB(16 VTFSMFWFMESJWFS LFSOFMNPEVMFT

  33. ubuntu EPDLFSDPOUBJOFS ཧ૝ OWJEJB(16 VTFSMFWFMESJWFS LFSOFMNPEVMFT

  34. ubuntu EPDLFSDPOUBJOFS ཧ૝ OWJEJB(16 VTFSMFWFMESJWFS LFSOFMNPEVMFT ESJWFSךQBUIכ04ח״׶殯ז׷

  35. NVIDIA Docker • Docker CLIͷബ͍ϥούʔ • `docker run` ࣌ʹඞཁͳvolumeΛࣗಈతʹmount
 ͯ͘͠ΕΔ

  36. NVIDIA Docker • Docker CLIͷബ͍ϥούʔ • `docker run` ࣌ʹඞཁͳvolumeΛࣗಈతʹmount
 ͯ͘͠ΕΔ

    "NB[PO&$4דכ劢؟ه٦ز

  37. ubuntu EPDLFSDPOUBJOFS ཧ૝ OWJEJB(16 VTFSMFWFMESJWFS LFSOFMNPEVMFT

  38. ubuntu EPDLFSDPOUBJOFS ཧ૝ OWJEJB(16 VTFSMFWFMESJWFS LFSOFMNPEVMFT (ಉҰόʔδϣϯ)

  39. ubuntu EPDLFSDPOUBJOFS ཧ૝ OWJEJB(16 VTFSMFWFMESJWFS LFSOFMNPEVMFT 㣐⡤鍑寸 (ಉҰόʔδϣϯ)

  40. • Driver͕ඞཁ • nvidia-driverͷkernel module • ಉ͡όʔδϣϯͷuser-level drivers • Docker

    Container͔ΒGPU devicesΛૢ࡞͢Δҝ
 Containerʹద੾ͳLinux Capabilityͷઃఆ͕ඞཁ Ծ૝Խ v.s. Χʔωϧ

  41. • GPUσόΠε͸ಛघͳϑΝΠϧͱͯ͠ଘࡏ • ΞΫηε͢ΔͨΊʹಛఆͷCapabilityઃఆ͕ඞཁ Ծ૝Խ v.s. Χʔωϧ EPDLFSSVOa EFWJDFEFWOWJEJBEFWOWJEJBa EFWJDFEFWOWJEJBVWNEFWOWJEJBVWNa

    HQVXPSLFS

  42. • GPUσόΠε͸ಛघͳϑΝΠϧͱͯ͠ଘࡏ • ΞΫηε͢ΔͨΊʹಛఆͷCapabilityઃఆ͕ඞཁ Ծ૝Խ v.s. Χʔωϧ EPDLFSSVOa EFWJDFEFWOWJEJBEFWOWJEJBa EFWJDFEFWOWJEJBVWNEFWOWJEJBVWNa

    HQVXPSLFS &$4ͷ5BTLఆٛʹ͓͍ͯEFWJDFΦϓγϣϯ͸ະαϙʔτ

  43. • GPUσόΠε͸ಛघͳϑΝΠϧͱͯ͠ଘࡏ • ΞΫηε͢ΔͨΊʹಛఆͷCapabilityઃఆ͕ඞཁ Ծ૝Խ v.s. Χʔωϧ EPDLFSSVOa EFWJDFEFWOWJEJBEFWOWJEJBa EFWJDFEFWOWJEJBVWNEFWOWJEJBVWNa

    HQVXPSLFS &$4ͷ5BTLఆٛʹ͓͍ͯEFWJDFΦϓγϣϯ͸ະαϙʔτ

  44. • GPUσόΠε͸ಛघͳϑΝΠϧͱͯ͠ଘࡏ • ΞΫηε͢ΔͨΊʹಛఆͷCapabilityઃఆ͕ඞཁ Ծ૝Խ v.s. Χʔωϧ EPDLFSSVOQSJWJMFHFEHQVXPSLFS

  45. Ծ૝Խ v.s. Χʔωϧ EPDLFSSVOQSJWJMFHFEHQVXPSLFS • capability શ։์ • rootͰ࣮ߦ͞Ε͍ͯΔdockerd্ͷcontainerͷதͰrootΛ औ͍ͬͯΔͷͰ৭ʑग़དྷΔ

    EPDLFSSVOQSJWJMFHFEBMQJOFMBUFTUEBUFT • GPUσόΠε͸ಛघͳϑΝΠϧͱͯ͠ଘࡏ • ΞΫηε͢ΔͨΊʹಛఆͷCapabilityઃఆ͕ඞཁ

  46. • GPUσόΠε͸ಛघͳϑΝΠϧͱͯ͠ଘࡏ • ΞΫηε͢ΔͨΊʹಛఆͷCapabilityઃఆ͕ඞཁ Ծ૝Խ v.s. Χʔωϧ EPDLFSSVOQSJWJMFHFEHQVXPSLFS • rootҎ֎ͷϢʔβʔͰ࣮ߦ͢Δ͜ͱʹ͢Δ

    • DockerFile಺Ͱ `USER runner`

  47. • ྉཧࣸਅͷࣗಈऩूαʔϏεΛ • CaffeNet[1]Λݩʹ࡞ͬͨྉཧը૾ೝࣝϞσϧͱ • Amazon SQS/S3 Ͱߏங͞Εͨσʔλϑϩʔͱ • Amazon

    ECS (GPU instance) Λར༻ͯ͠ӡ༻͍ͯ͠Δ࿩ <>IUUQTHJUIVCDPN#7-$DB⒎F Agenda