Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Speaker Deck
PRO
Sign in
Sign up for free
メルカリの写真検索を支えるバックエンド CCSE 2019 version
Hirofumi Nakagawa/中河 宏文
July 13, 2019
Programming
0
170
メルカリの写真検索を支えるバックエンド CCSE 2019 version
過去発表分との差分としてEdge関連のデモが追加されています。
デモ動画:
https://youtu.be/UsNMDuZyaJU
Hirofumi Nakagawa/中河 宏文
July 13, 2019
Tweet
Share
More Decks by Hirofumi Nakagawa/中河 宏文
See All by Hirofumi Nakagawa/中河 宏文
IoTデバイスでMLモデルを動かす技術
hnakagawa
0
39
Kanazawa_AI.pdf
hnakagawa
0
89
メルカリ写真検索における Amazon EKS の活用事例と プロダクトにおけるEdgeAI technologyの展望
hnakagawa
5
7.3k
メルカリ写真検索における Amazon EKS の活用事例
hnakagawa
6
29k
メルカリの写真検索を支えるバックエンド
hnakagawa
1
890
Mercari ML Platform
hnakagawa
1
16k
mlct.pdf
hnakagawa
2
1.6k
機械学習によるマーケット健全化施策を支える技術
hnakagawa
0
140
メルカリのマーケット健全化施策を支えるML基盤
hnakagawa
10
8.2k
Other Decks in Programming
See All in Programming
Cybozu GoogleI/O 2022 LT会 - Input for all screens
jaewgwon
0
150
io22 extended What's new in app performance
veronikapj
0
320
こそこそアジャイル導入しようぜ!
ichimichi
0
1k
How we run a Realtime Puzzle Fighting Game on AWS Serverless
falken
0
240
Mobile Product Engineering
championswimmer
0
280
Swift6のprotocol
omochi
3
310
Keeping your team in top shape with the Gradle Enterprise API
runningcode
3
120
Improving Developer Experience Through Tools and Techniques 2022
krzysztofzablocki
0
200
A Philosophy of Software Design 後半
yosuke_furukawa
PRO
8
2.5k
Android Tools & Performance
takahirom
1
410
Overview of The Modern Data Stack / モダンデータスタック概論
satoshihirose
6
3.2k
1時間半で克服するJavaScriptの非同期処理/async_javascript_kokufuku
marchin1989
2
590
Featured
See All Featured
The Power of CSS Pseudo Elements
geoffreycrofte
46
3.9k
How New CSS Is Changing Everything About Graphic Design on the Web
jensimmons
213
11k
Art Directing for the Web. Five minutes with CSS Template Areas
malarkey
196
9.4k
GitHub's CSS Performance
jonrohan
1020
420k
Done Done
chrislema
174
14k
Designing Dashboards & Data Visualisations in Web Apps
destraynor
224
49k
The Straight Up "How To Draw Better" Workshop
denniskardys
225
120k
Designing with Data
zakiwarfel
91
3.9k
How GitHub Uses GitHub to Build GitHub
holman
465
280k
Creatively Recalculating Your Daily Design Routine
revolveconf
207
10k
Cheating the UX When There Is Nothing More to Optimize - PixelPioneers
stephaniewalter
269
11k
4 Signs Your Business is Dying
shpigford
169
20k
Transcript
Confidential - Do Not Share 1 ϝϧΧϦͷࣸਅݕࡧΛࢧ͑ΔόοΫΤϯυ גࣜձࣾϝϧΧϦ தՏ จ
Confidential - Do Not Share 2 • 20177݄ೖࣾ • ॴଐSRE→AI/MLνʔϜ
• σόΠευϥΠό։ൃ͔ΒϑϩϯτΤ ϯυ։ൃ·ͰΔԿͰ Twitter: hnakagawa14 GitHub: hnakagawa தՏ จ (Hirofumi Nakagawa) ࣗݾհ 2
Confidential - Do Not Share 3 Introduction
Confidential - Do Not Share 4 What is ࣸਅݕࡧ •
ࣸਅݕࡧͱɺ͍ΘΏΔը૾ ݕࡧػೳ • ΞϓϦ͔ΒࣸਅΛݩʹΛ ݕࡧ͢Δ • ໊ΛΒͳͯ͘ը૾͔ ΒΛݕࡧͰ͖Δ ಈըϦϯΫ: https://youtu.be/kTni8EvOCgI
Confidential - Do Not Share 5 جຊతͳࣸਅݕࡧͷΈ Deep Neural Networks
(DNN)Λ༻ͯ͠ը૾ ͔ΒಛϕΫτϧΛऔಘ औಘͨ͠ಛϕΫτϧΛ Approximate Nearest Neighbor Index(ANN Index) ʹՃͯ͠ը૾indexΛߏங ݕࡧ࣌ʹಉ͘͡ը૾͔Β DNNΛհͯ͠ಛϕΫτϧΛऔ ಘ͠ɺANN Index͔Βݕࡧ 2 3 1
Confidential - Do Not Share 6 What is Kubernetes •
KubernetesʢҎԼk8sʣͱ Φʔϓϯιʔεͷίϯςφɾ ΦʔέετϨʔγϣϯγες Ϝ • k8sʹCustom Resource DefinitionͱݺΕΔಠࣗͷ ϦιʔεΛఆٛͰ͖Δػೳ͕ ͋Γɺ։ൃऀͦͷػೳΛհ ͯ͠k8sͷػೳΛ֦ுͰ͖Δ
Confidential - Do Not Share 7 ML Platform Lykeion •
Training/Serving CRD & ΧελϜ ίϯτϩʔϥ • ίϯςφϕʔεɾύΠϓϥϯ • Training/Serving ίϯςφΠϝʔ δɾϏϧμʔ • ϞσϧɾϨϙδτϦ ࣸਅݕࡧLykeionͱݺΕΔͷ ML Platform্ʹߏங͞Ε͓ͯΓɺԼ هͷػೳPlatformଆͷػೳΛ༻͠ ͍ͯΔ
Confidential - Do Not Share 8 Architectureɹ
Confidential - Do Not Share 9 Architecture֓ཁਤ
Confidential - Do Not Share 10 1.TrainingɾϦιʔεͷ࡞
Confidential - Do Not Share 11 TrainingɾϦιʔεͷ࡞ • Training custom
resourceΛ CronJob͕࡞ • ΧελϜɾίϯτϩʔϥ͕ CRDɾϦιʔεͰઃఆ͞Εͨ ίϯςφϕʔεɾύΠϓϥΠ ϯΛ࣮ߦ • ࣮ߦ͢Δόον୯Ґͱͯ͠ Hourly, Daily, Monthly͕ଘࡏ
Confidential - Do Not Share 12 • ֤ఔΛݸผͷίϯςφɾΠϝʔδͰ࣮ߦ • ϥΠϒϥϦͷґଘؔͳͲڥφΠʔϒͳMLύΠϓϥΠϯͷΛղܾ
• ύΠϓϥΠϯDAGYAMLͰهड़ ֤ఔͷೖग़ྗPersistent VolumeʢҎ ԼPVʣΛհ͢ ίϯςφϕʔεɾύΠϓϥΠϯ
Confidential - Do Not Share 13 TrainingɾϦιʔεͷ࡞ • શͯͷόον࣮ߦใ͕ CRDɾϦιʔεͱͯ͠k8s্ʹ
Δ • ಉ͡ॲཧΛ࠶࣮ߦग़དྷΔͨ Ίɺόονͷ࠶࣮ߦΛ͏ো ෮چ࡞ۀ͕༰қ
Confidential - Do Not Share 14 2.ը૾ͷμϯϩʔυ
Confidential - Do Not Share 15 • S3্ʹଘࡏ͢ΔϝϧΧϦɾΠϝʔδετΞ͔Βը૾Λμϯϩʔυ • ύΠϓϥΠϯ্ͬͱ͕͔͔࣌ؒΔఔʢը૾͕େͳͨΊ)
• ͦͷͨΊPVʹҰఆظؒΩϟογϡ͢ΔࣄʹΑͬͯ࠶ΠϯσοΫε͕ඞཁ ͳ࣌ʹૉૣ͘ύΠϓϥΠϯΛճͤΔΑ͏ʹ͍ͯ͠Δ ը૾ͷμϯϩʔυ
Confidential - Do Not Share 16 3. ΞηοτͷΞοϓϩʔυ
Confidential - Do Not Share 17 • ύΠϓϥΠϯͷՌͰ͋ΔಛϕΫτϧͱANN IndexΛϞσϧɾϨϙδτ Ϧʹอଘ
• શͯͷՌόʔδϣϯཧ͞Εͨঢ়ଶͰอଘ͞ΕΔ • ϞσϧɾϨϙδτϦGCS্ʹߏங ΞηοτͷΞοϓϩʔυ
Confidential - Do Not Share 18 4.ServingΠϝʔδͷϏϧυ
Confidential - Do Not Share 19 1. ϞσϧɾϨϙδτϦΛImage BuilderͱݺΕΔdaemon͕ࢹ 2.
৽͍͠Serving͖͢Ϧιʔε͕Ճ͞ΕΔͱࣗಈͰServingίϯςφɾΠϝ ʔδΛϏϧυ 3. ίϯςφɾΠϝʔδશͯͷANN IndexͷαʔϏϯάʹඞཁͳϦιʔεΛ શؚͯΜͰ͍Δ 4. Ϗϧυ͞ΕͨίϯςφɾΠϝʔδΛίϯςφɾϨδετϦʹϓογϡ ServingΠϝʔδͷϏϧυ
Confidential - Do Not Share 20 5.ServingϦιʔεͷ࡞
Confidential - Do Not Share 21 ServingɾϦιʔεͷ࡞ • Image BuilderίϯςφɾΠϝʔ
δΛϏϧυͨ͋͠ͱɺServingΧε λϜɾϦιʔεΛ࡞ • ServingΧελϜɾίϯτϩʔϥ CRDɾϦιʔεͷઃఆΛݩʹඞཁ ͳDeploymentɺServiceΛ࡞ • ຊγεςϜͰߏங͞ΕͨANN IndexΛݸผͷIndexαʔϏεͱ͠ ͯσϓϩΠ
Confidential - Do Not Share 22 6.ServiceσΟεΧόϦ
Confidential - Do Not Share 23 αʔϏεɾσΟεΧόϦ • Ϋϥελ্ʹଘࡏ͢ΔIndexαʔϏ εΛk8sΛհͯࣗ͠ಈతʹऔಘ͢
Δ • ͳΔ͘େ͖ͳཻͷIndexΛ ༻͢ΔΑ͏ɺҟͳΔظؒɾཻͷ IndexαʔϏε(Hourly, Daily, Monthly) ΛࣗಈతʹΈ߹ΘͤΔ • REST <-> IndexαʔϏεؒͷϓϩ τίϧgRPCΛ༻
Confidential - Do Not Share 24 7.֓ཁਤৼΓฦΓ
Confidential - Do Not Share 25 Conclusionɹ
Confidential - Do Not Share 26 1. ίϯςφɾϕʔεͷ࠶ݱੑͷߴ͍γεςϜ 2. k8sͷCRD/ΧελϜɾίϯτϩʔϥαʔϏεɾσΟεΧόϦͷػೳΛ׆༻
3. Batch Execution as Custom ResourceɺML PlatformͰ࣮ݱ͞Ε͍ͯΔػ ೳΛ༻͠ɺϩόετͳγεςϜΛߏங 4. ΫϥυɾΠϯϑϥΛk8sͰநԽ͢ΔࣄʹΑͬͯɺ֤Ϋϥυɾϕϯμͷྑ ͍ͱ͜औΓΛ͍ͯ͠Δ ࣸਅݕࡧͷόοΫΤϯυɾΠϯϑϥ
Confidential - Do Not Share 27 Next Futureɹ
Confidential - Do Not Share 28 Realtime image search •
EdgeଆͰobject detection & feature extractionΛߦ͏ • feature vectorΛΫΤϦʹANN Index service͔ΒྨࣅΛऔಘ • Graph RuntimeTF Lite + custom operatorΛ༻
Confidential - Do Not Share 29 Realtime image search demo
• ݕࡧʹඞཁͳਪॲཧͷେ෦ΛEdgeଆͰߦ ͍ͬͯΔ • ϦΞϧλΠϜͳΠϯλϥΫγϣϯΛ࣮ݱ • UX্େ͖ͳϝϦοτ͕༗Δ
Confidential - Do Not Share 30 Thank you all for
coming todayɹ