Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
mlct.pdf
Search
Hirofumi Nakagawa/中河 宏文
July 23, 2018
Programming
2
2k
mlct.pdf
Hirofumi Nakagawa/中河 宏文
July 23, 2018
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
140
Kanazawa_AI.pdf
hnakagawa
0
160
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
8.7k
メルカリの写真検索を支えるバックエンド CCSE 2019 version
hnakagawa
0
280
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
1.1k
Mercari ML Platform
hnakagawa
1
17k
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
220
メルカリのマーケット健全化施策を支えるML基盤
hnakagawa
10
8.9k
Other Decks in Programming
See All in Programming
今年のアップデートで振り返るCDKセキュリティのシフトレフト/2024-cdk-security-shift-left
tomoki10
0
360
Stackless и stackful? Корутины и асинхронность в Go
lamodatech
0
1.3k
Внедряем бюджетирование, или Как сделать хорошо?
lamodatech
0
940
見えないメモリを観測する: PHP 8.4 `pg_result_memory_size()` とSQL結果のメモリ管理
kentaroutakeda
0
930
AWS re:Invent 2024個人的まとめ
satoshi256kbyte
0
100
DevinとCursorから学ぶAIエージェントメモリーの設計とMoatの考え方
itarutomy
0
140
PSR-15 はあなたのための ものではない? - phpcon2024
myamagishi
0
400
Jaspr Dart Web Framework 박제창 @Devfest 2024
itsmedreamwalker
0
150
ErdMap: Thinking about a map for Rails applications
makicamel
1
600
Fibonacci Function Gallery - Part 2
philipschwarz
PRO
0
210
Оптимизируем производительность блока Казначейство
lamodatech
0
950
Rubyでつくるパケットキャプチャツール
ydah
0
170
Featured
See All Featured
Navigating Team Friction
lara
183
15k
GraphQLとの向き合い方2022年版
quramy
44
13k
Thoughts on Productivity
jonyablonski
68
4.4k
Building Flexible Design Systems
yeseniaperezcruz
328
38k
Imperfection Machines: The Place of Print at Facebook
scottboms
267
13k
Mobile First: as difficult as doing things right
swwweet
222
9k
Building a Modern Day E-commerce SEO Strategy
aleyda
38
7k
The Pragmatic Product Professional
lauravandoore
32
6.4k
How GitHub (no longer) Works
holman
312
140k
Making the Leap to Tech Lead
cromwellryan
133
9k
GraphQLの誤解/rethinking-graphql
sonatard
68
10k
Fontdeck: Realign not Redesign
paulrobertlloyd
82
5.3k
Transcript
ϝϧΧϦͷMLج൫ MLCT vol.5 hnakagawa
ࣗݾհ • Hirofumi Nakagawa (hnakagawa) • 20177݄ೖࣾ • ॴଐSRE •
σόΠευϥΠό։ൃ͔Βϑϩϯ τΤϯυ։ൃ·ͰΔԿͰ • NOT σʔλαΠΤϯςΟετ • https://github.com/hnakagawa
͓ࣄ • ML Platform։ൃ • σʔλαΠΤϯςΟετͱSREͷεΩϧΪϟο ϓΛຒΊΔ • ML Reliability,
SysML?, MLOps? • SREͷཱ͔ΒMLγεςϜͷࣗಈԽΛߦ͏
ML Platform • ͷML Platform • kubernetesϕʔε • طଘͷML FrameworkΛ༻͠
؆୯ʹTraining/ServingΛߦ͏ ڥΛఏڙ
ͦͷ͏ͪOSSͰެ։༧ఆ(ଟ
ϝϧΧϦͷMLར༻ࣄྫ • ײಈग़ • ҧग़ݕ • Ձ֨αδΣετ • ΤΠταδΣετ ʑ…
̍ઍສpredictionΛߦ͍ͬͯΔ
ML Platform Architecture ,VCFSOFUFT $POUSPMMFS $-* $MVTUFS8PSLGMPX %BTICPBSE 4UPSBHF(BUFXBZ .FUSJDT
3VOOFS $PNQPOFOU .FSDBSJ.- $PNQPOFOU &YUFSOBM .JEEMFXBSF
Model Training & Serving Workflow
.-1MBUGPSN USBJOJOHDMVTUFS Workflow for Production $* .-1MBUGPSN TFSWJOHDMVTUFSGPSUFTU .PEFM3FHJTUSZ +PC
+PC ɾɾ 3&45 "1* 4USFBNJOH 5'4FSW JOH ɾɾɾ
.-1MBUGPSN USBJOJOHDMVTUFS Training Workflow $* .PEFM3FHJTUSZ +PC +PC ɾɾɾ 1.
GitHubͷpushΛτϦΨʹtrainingΛىಈ 2. Training͞ΕͨModelModel Registry ্͕Δ
Serving Workflow .-1MBUGPSN TFSWJOHDMVTUFSGPSUFTU .PEFM3FHJTUSZ ɾɾ 3&45 "1* 4USFBNJOH 5'
4FSWJOH 1. Model RegistryΛࢹͯࣗ͠ಈͰModel ΛServing 2. Serving&Test͕ޭ͢Δͱຊ൪༻k8s manifestΛग़ྗ
Container Workflow %BUB4PVSDF *NBHF 5FYUɹ 1SFQSPDFT TJOH *NBHF &TUJNBUPS *NBHF
17 17 1JDUVSF 1SFQSPDFT TJOH *NBHF 17 It’s own implementation
Model Serving APIͷߏྫ 5FOTPS'MPX 4FSWJOH 5' .PEFM 5' .PEFM 'MBTL
4, .PEFM 4, .PEFM 4, .PEFM gRPC .FSDBSJ"1* REST FlaskͰલॲཧΛߦ͍ ཪͷTensorFlow Servingʹ͍͛ͯΔ
Model Serving API Streaming ver ͷߏྫ 5FOTPS'MPX 4FSWJOH 5' .PEFM
5' .PEFM .-1MBUGPSN 'SBNFXPSL PS "QBDIF#FBN 4, .PEFM 4, .PEFM 4, .PEFM gRPC PubSub
ModelͱίϯςφɾΠϝʔδ • ڊେͳML ModelΛίϯςφɾΠϝʔδʹؚΊ Δ͔൱͔ • ؚΊͳ͍ͷͰ͋ΕԿॲʹஔ͢Δ͔ • ϙʔλϏϦςΟੑͱϩʔυ࣌ؒͷτϨʔυΦϑ •
ྑ͍ΞΠσΟΞ͕͋Εڭ͑ͯԼ͍͞…
௨ৗͷAPIͱಛੑ͕ҧ͏ • ѻ͏ϦιʔεɺModelαΠζ͕େ͖͘ͳΔ ߹͕ଟ͍(ඦMBʙGB) • CPUɾϝϞϦϦιʔεͷফඅ͕ܹ͍͠ • ߹ʹΑͬͯGPU͏
ϝϞϦফඅ • ҧݕγεςϜͷPython࣮෦࣮ߦ࣌ ʹ2GBϝϞϦΛফඅ͢Δˠࠓޙ͞Βʹ૿͑ Δ༧ఆ͋Δ • Scikit-learnͰهड़͞Εͨલॲཧ෦͕େ͖͘ ͳΓ͕ͪ
Pythonͱฒྻੑ • વThread͕͑ͳ͍(GILͷͨΊ) • ϓϩηεຖʹModelΛϩʔυ͢Δͱඞཁͳϝ ϞϦαΠζ͕େ͖͘ͳΔˠ Blue-Green DeployͷোʹͳΔ
ਖ਼PythonͰͷServing Πϯϑϥతʹਏ͍ࣄ͕ଟ͍…
ϝϞϦΛݡ͘͏ • fork͢ΔલʹmodelΛϩʔυ͠Copy on Write Λޮ͔͢ • k8sͷone process per
containerηΦϦ͋ ͑ͯഁ͍ͬͯΔ
Copy On Writeͷ෮श ϝϞϦ ϓϩηε ࢠϓϩηε 2.fork 1BHF" 1.allocation ಉ͡ྖҬΛࢀর
ϓϩηε͕ϝϞϦͷ༰Λ ॻ͖͑Δͱ… ϝϞϦ ϓϩηε ࢠϓϩηε 1BHF" 1BHF# OS͕ผͷྖҬΛAllocationͯ͠ݩσʔλΛίϐʔ͢Δ ผͷྖҬΛࢀর
Current Issues
ߴͳܧଓతϝϯςφϯε͕ඞཁ • MLػೳσʔλͷ͕มΘͬͨΓɺ༧֎ ͷ͕ൃੜͨ͠Γͯ͠ɺͦΕΒʹରԠ͠ଓ ͚Δඞཁ͕͋Δ MLػೳϦϦʔεޙେ͖ͳ ίετ͕͔͔Γଓ͚Δ
େ෯ͳࣗಈԽ͕ඞਢ
In Progress
ߴͳࣗಈԽ • ࣾͷσʔλ͔ΒFeature Extraction͢Δ࣮ ΛίϯϙʔωϯτԽ • ಛఆͷΛղܾ͢ΔϞσϧߏஙΛ͋Δఔ ࣗಈԽ • ϦϦʔεޙͷRe-TrainingɺHyper
parameter optimizationɺDeployΛࣗಈԽ
AutoFlow 'FBUVSF&YUSBDUJPO $PNQPOFOUT $MBTTJGJDBUJPO $PNQPOFOUT $PODBUFOBUJPO $PNQPOFOUT .PEFM #VJMEFS $PNQPOFOUT
3FHJTUSZ Ϋϥελ্ͰϞσϧͷࣗಈߏஙͱϋΠύʔύϥ ϝʔλͷࣗಈௐΛߦ͏
AutoServing %FQMPZ ϦϦʔεޙͷਫ਼ࢹɾRe-TrainingɾRe-Deploy ΛࣗಈͰߦ͏ .POJUPSJOH &WBMVBUJPO )ZQFS QBSBNFUFS PQUJNJ[BUJPO 3F5SBJOJOH
·ͱΊ • MLʹগ͠௨ৗͱҧ͏Πϯϑϥ͕ඞཁʹͳΔ ˠ·ͩϕετɾϓϥΫςΟε͔Βͳ͍ • ͦͦMLͳػೳΛຊ֨ӡ༻͠Α͏ͱ͢Δ ͱɺେ෯ͳࣗಈԽɾΈԽΛਐΊͳ͍ͱ্ ख͘ߦ͔ͳ͍
͝ਗ਼ௌ͋Γ͕ͱ͏͍͟͝·ͨ͠!!