【初心者向け勉強会#9】MLOpsの基本 ~構築から運用まで~ / MLOps Basics: From Development to Operations

by Ogata Katsuya

Slide 1

Slide 1 text

ॳ৺ऀ޲͚ษڧձ #9 MLOps Author: GitHub@ogatakatsuya  Lisence: CC BY-SA 4.0

Slide 2

Slide 2 text

ࣗݾ঺հ • ໊લ: ॹํ ࠀ࠸ (͓͕ͨ ͔ͭ΍) • ग़਎: ٶ࡚ݝখྛࢢ • ॴଐ: େࡕେֶ ৘ใՊֶݚڀՊ B4 • झຯ: ւ֎ཱྀߦ / ϙʔΧʔ / ొࢁ • X: @ogata_katsuya • Homepage: www.ogatakatsuya.com 2 ஍ݩͱ࣮ՈͱʹΌΜ͜

Slide 3

Slide 3 text

ຊߨ࠲ʹ͍ͭͯ 3 • ର৅ • MLOpsʹ͍ͭͯ஌Γ͍ͨ • ϓϩμΫτʹMLΛऔΓೖΕ͍ͨ • ಺༰ • MLOpsͱ͸ʁ • MLOpsʹ͍ͭͯͷجૅతͳ಺༰

Slide 4

Slide 4 text

͓඼ॻ͖ Agenda 1. MLOpsͱ͸ʁ (10෼) 2. MLOpsͷߏ੒ཁૉ (25෼) 3. LLMOpsͱ͸ʁ(5෼) 4. MLOpsΛମݧͯ͠ΈΑ͏ (30෼) 4

Slide 5

Slide 5 text

1. MLOpsͱ͸ʁ

Slide 6

Slide 6 text

MLOpsͱ͸ʁ Machine Learning × Operations 6 Machine Learning Operations × DevOps for Machine Learning    The techniques for operating the system   with machine learning modules.

Slide 7

Slide 7 text

ػցֶशͷྺ࢙ 2010೥͝Ζ͔Βίϯελϯτʹ஫໨͞Ε࢝ΊΔ 7

Slide 8

Slide 8 text

ػցֶशͷྺ࢙ 2010೥͝Ζ͔Βίϯελϯτʹ஫໨͞Ε࢝ΊΔ 8 DeepLearning  (2006) AlexNetͷ஀ੜ  (2012) ResNetͷ஀ੜ  (2015) AlphaGo  (2016) Transformer  (2017) GPT3  (2020)

Slide 9

Slide 9 text

ML͔ΒLLM΁ ML͕མͪண͍ͨͷͰ͸ͳ͘LLM͕஫໨͞Ε͍ͯΔ 9 ੨ઢ: ػցֶश  ੺ઢ: LLM

Slide 10

Slide 10 text

MLOpsͷྺ࢙ 2020೥͝Ζ͔Β஫໨͞Ε͍ͯΔ👀 10 ੨ઢ: MLOps  ੺ઢ: DevOps ʮHidden Technical Debt in   Machine Learning Systemsʯ  ͷެ։ (2015) Google Cloud NextͰ  MLOpsͱ͍͏ݴ༿͕ॳΊͯ࢖ΘΕΔ

Slide 11

Slide 11 text

MLOpsͷྺ࢙ MLOpsͷࢥ૝త஀ੜ 11 Hidden Technical Debt in Machine Learning Systems (NeurlIPS2015) ΑΓҾ༻

Slide 12

Slide 12 text

MLOpsͷྺ࢙ MLOps = DevOps for ML 12 Geminiʹੜ੒ͯ͠΋ΒͬͨMLOpsͷΠϝʔδ

Slide 13

Slide 13 text

MLOpsͱ͸ʁ MLͱϓϩμΫτΛܨ͙Ξμϓλ 13 • MLΛ࣮ϓϩμΫτͰσϦόϦʔ͢ΔͨΊͷٕज़ • MLͱϓϩμΫτ΁ͷσϦόϦʔͱͷ੹຿Λ෼཭͢Δ • MLνʔϜ͸ϞσϧͷվળʹूதͰ͖ΔΑ͏ʹ͢Δ • ͦͷதͰ΋MLʹΑΔٕज़తෛ࠴ΛஷΊ͍͔ͯͳ͍ͨΊͷٕज़

Slide 14

Slide 14 text

2. MLOpsΛߏ੒͢Δཁૉ

Slide 15

Slide 15 text

MLOpsͷߏ੒ཁૉ σʔλͷલॲཧ͔ΒϞσϧ։ൃɾσϦόϦʔ·Ͱ ࣮ݧج൫ ֶशύΠϓϥΠϯ ਪ࿦ج൫ σʔλɾϞσϧͷόʔδϣϯ؅ཧ ܧଓతֶश ؂ࢹج൫

Slide 16

Slide 16 text

MLOpsͷߏ੒ཁૉ ͬ͘͘͢͟͝ΓͱMLOpsͷશମ૾ 16 Preprocessing Model development Training Inference Data/Model Registry (versioning) loop Experimentation Watch

Slide 17

Slide 17 text

࣮ݧج൫ MLͷ࠶ݱੑͱ։ൃੜ࢈ੑΛߴΊΔ • ૊৫ͷதͷݸʑਓ͕ͦΕͧΕͷ࣮ݧ؀ڥΛ࣋ͭͱ໰୊͕ى͜Δ • ։ൃ؀ڥͷࠩʹΑΔ؀ڥߏஙίετͷ૿Ճ • ύοέʔδͷόʔδϣϯ͕߹Θͳ͍ • ࣮ݧΛߦͬͨ࣌ͷύϥϝʔλ͕Θ͔Βͳ͍ • ࠶ݱੑͷͳ͍ίʔυɾϞσϧ͕ྔ࢈͞ΕΔ 17

Slide 18

Slide 18 text

࣮ݧج൫ MLͷ࠶ݱੑͱ։ൃੜ࢈ੑΛߴΊΔ • ։ൃ؀ڥΛվળ͢Δ • ύοέʔδϚωʔδϟʔͷಋೖ (uv, poetry, etc.) • ؀ڥͷίϯςφԽ • ࣮ݧ༻ͷ؀ڥΛ༻ҙ͢Δ (Jupyter notebookͳͲͷNaaS) • ࣮ݧ؅ཧπʔϧͷಋೖ (Weights and Bias, TensorBoard, etc.) 18

Slide 19

Slide 19 text

ֶशύΠϓϥΠϯ લॲཧ͔ΒֶशɾσϦόϦʔ·ͰͷϑϩʔΛ੔උ͢Δ • ֶशύΠϓϥΠϯ • σʔλऔಘ -> લॲཧ -> ֶश -> ݕূ -> σϓϩΠ • ͜ͷύΠϓϥΠϯΛ͋Β͔͡Ίߏ੒͓ͯ͘͜͠ͱʹΑΓɺ҆શʹ͔ͭߴ଎ʹ ߴ඼࣭ͳMLϞδϡʔϧΛఏڙͰ͖Δ • MLνʔϜ͸Ϟσϧͷ։ൃͷΈʹूதͰ͖Δ 19

Slide 20

Slide 20 text

ਪ࿦ج൫ ຊ൪ͷτϥϑΟοΫʹ଱͑͏Δج൫Λ༻ҙ͢Δ • ML Workload͸ҰൠతͳCPUॲཧͱҟͳΓɺ஫ҙ఺͕ଟ͍ • CPUͱ͸ผʹGPUͷ࢖༻཰Λߟ͑Δඞཁ͕͋Δ • 1ϦΫΤετ͋ͨΓͷϦιʔεΛ઎༗͢Δ͕࣌ؒ௕͍ • 2ͭͷਪ࿦ํ๏͕͋Δ • όονਪ࿦ • ΦϯϥΠϯਪ࿦ 20

Slide 21

Slide 21 text

ਪ࿦ج൫ ΦϯϥΠϯਪ࿦ / όονਪ࿦ 21 Users Server Data Store Server ML Model ML Model Online inference Batch inference

Slide 22

Slide 22 text

ਪ࿦ج൫ ͦΕͧΕͷਪ࿦ํ๏ͷPros/Cons • ΦϯϥΠϯਪ࿦ • Pros: ϦΞϧλΠϜʹ༧ଌ݁ՌΛฦ͢͜ͱ͕Ͱ͖Δ • Cons: αʔόʔΛৗʹཱͯͯɺϦΫΤετΛࡹ͚Δঢ়ଶʹ͓ͯ͘͠ඞཁ͕͋Δ • όονਪ࿦ • Pros: ߴεϧʔϓοτͰඅ༻ରޮՌ͕ߴ͍ • Cons: ϢʔβʔͷϦΫΤετ΍Πϕϯτʹରͯ͠ଈ࠲ʹਪ࿦Ͱ͖ͳ͍ 22

Slide 23

Slide 23 text

ਪ࿦ج൫ ͦΕͧΕͷਪ࿦ํ๏ͷϢʔεέʔε • ΦϯϥΠϯਪ࿦ • ΫϨδοτΧʔυͷෆਖ਼༧ଌ • ޿ࠂαʔόʔͷ͓͢͢Ί༧ଌ • όονਪ࿦ • γϣοϐϯάαΠτͷ͓͢͢Ί༧ଌ • إೝূͷຒΊࠐΈੜ੒ (ϦΫΤετ͕དྷͨΒ͢Ͱʹ͋Δ΋ͷͱݟൺ΂Δ͚ͩ) • RAGͷߏங (υΩϡϝϯτͱΠϯσοΫεͷಉظ͸εέδϡʔϧ͢Δ) 23

Slide 24

Slide 24 text

σʔλɾϞσϧͷόʔδϣϯ؅ཧ σʔλͱϞσϧ͕MLͷ؊ • ҰൠతͳϓϩμΫτ։ൃͱMLΛؚΉϓϩμΫτ։ൃͷҧ͍ • σʔλͱϞσϧ͕͋Δ͔Ͳ͏͔ • ίʔυͷόʔδϣϯ؅ཧ͸GitHubͰߦ͏͜ͱ͕Ͱ͖Δ • MLͷڍಈΛܾఆ͚ͮΔͷ͸ʁʁ • ܇࿅ʹ࢖༻͞ΕΔσʔλͱ࣮ࡍʹਪ࿦͢ΔϞσϧ • ͜ΕΒ΋όʔδϣϯ؅ཧ͓͔ͯ͠ͳ͍ͱσάϨͰ͖ͳ͘ͳͬͯ͠·͏ 24

Slide 25

Slide 25 text

σʔλɾϞσϧͷόʔδϣϯ؅ཧ σʔλͱϞσϧ͕MLͷ؊ • جຊతʹ͸όʔδϣϯʹ૬౰͢Δ΋ͷ • v1,v2ͱ͔ɺGitͰ͍͏ͱ͜Ζͷcommit hashͱσʔλɾϞσϧΛؔ࿈෇͚ͯετϨʔδ ʹอଘͰ͖Ε͹ྑ͍ͱࢥ͏ • MLʹಛԽͨ͠πʔϧ͸ҎԼ • DVC (Data Version Control) • Gitͱಉ͡ײ͡ͷίϚϯυͰετϨʔδʹpushͰ͖Δ • MLFlow 25

Slide 26

Slide 26 text

ܧଓతֶश σʔλͷมԽʹؤ݈ͳϞσϧΛܧଓతʹ࡞Δ • CI/CD (Continuous Integration, Continuous Deployment/Delivery)ͱಉ͡Α͏ ʹɺCT (Continuous Training) ͱݺ͹ΕͨΓ͢Δ • ͳͥ͜ͷ࢓૊Έ͕ඞཁ͔ʁ • ࣮ϓϩμΫτͰ͸σʔληοτͷ܏޲͸มΘΓ͏Δ • e.g. ϢʔβʔΠϕϯτ/ ੈͷதͷ৘੎ / ग़඼͞ΕΔ঎඼ɾΧςΰϦ • ͜ΕΒͷมԽʹ௥ਵ͢ΔͨΊʹఆظతʹֶशΛ͢Δඞཁ͕͋Δ 26

Slide 27

Slide 27 text

؂ࢹج൫ ෆ҆ఆͳMLͷෆ۩߹ʹؾ͚ͮΔج൫Λ࡞Δ • MLϞσϧ͸֬཰తͳڍಈΛ͢Δ΋ͷ΋͋Δ • શͯͷೖྗͱग़ྗΛ׬ᘳʹςετ͢Δ͜ͱ͸Ͱ͖ͳ͍ • ઌड़ͷ௨Γɺεϧʔϓοτ΋ෆ҆ఆʹͳΓ͕ͪ • ML͸ෆ۩߹Ͱ΋໌ࣔతͳΤϥʔΛు͍ͯ͘Εͳ͍ • ͙͢ʹෆ۩߹΍όάʹؾ͚ͮΔΑ͏ʹ؂ࢹج൫Λ͓࣋ͬͯ͘͜ͱ͕ॏཁ 27

Slide 28

Slide 28 text

؂ࢹج൫ ओͳ؂ࢹ߲໨ • Ϟσϧͷ༧ଌ඼࣭ͷϝτϦΫε • σʔλυϦϑτ • όΠΞεͱެฏੑ • ӡ༻ɾιϑτ΢ΣΞϝτϦΫε 28 Ҿ༻: https://www.evidentlyai.com/ml-in-production/model-monitoring

Slide 29

Slide 29 text

؂ࢹج൫ σʔλυϦϑτͱ͸ 29 Ҿ༻: https://www.evidentlyai.com/ml-in-production/model-monitoring

Slide 30

Slide 30 text

૊৫ͱMLOps ͦͷMLOpsຊ౰ʹඞཁͰ͔͢ʁ • ͜͜Ͱڍ߲͛ͨ໨͸΄ΜͷҰ෦͔ͭඇৗʹந৅త • ߦ͏΂͖MLOpsͷ߲໨͸MLϞσϧͷಛ௃΍૊৫ͷߏ଄ɾن໛ʹେ͖͘ґΔ • ʮ͏Θ͊ʔɺMLϞσϧΛ૊ΈࠐΉ͔ΒMLOps΋શ෦΍Βͳ͖Ό💦ʯ • ͪΐͬͱ౿ΈͱͲ·ͬͯߟ͑Δ΂͖ 30

Slide 31

Slide 31 text

૊৫ͱMLOps ͦͷMLOpsຊ౰ʹඞཁͰ͔͢ʁ • MLOpsʹ΋؅ཧίετ͕͔͔ΔͨΊɺಋೖ͸ஈ֊త͔ͭ৻ॏʹߦ͏ • ૊৫ͷن໛ͱϏδωεతͳ؍఺ΛؑΈͯຊ౰ʹඞཁ͔ߟ͑Δ • ·ͣ࠷ॳ͸खಈͰ܇࿅Λߦ͍ɺຖճखಈͰσϓϩΠͰ໰୊ͳ͍ • ࠓճͷ৔߹ɺϋοΧιϯͰ͸࣮ӡ༻ʹ͍ͭͯͦ͜·Ͱߟ͑Δඞཁ͸ͳ͍ • ͔͠͠ɺ࣮ࡍͷӡ༻Λݟӽ࣮ͯ͠૷͢Δ͜ͱ͸େมษڧʹͳΔ • ྫ͑͹ɺֶशύΠϓϥΠϯΛ࡞੒ͨ͜͠ͱͰΑΓՁ஋ఏڙʹ஫ྗͰ͖ΔΑ͏ʹͳͬͨΒ࠷ߴ🙆 31

Slide 32

Slide 32 text

૊৫ͱMLOps ͦͷMLOpsຊ౰ʹඞཁͰ͔͢ʁ • ૊৫΋ͦ͏Ͱ͕͢ɺMLϞσϧ΍ϏδωεϞσϧʹΑͬͯMLOps΋͞·͟· • ຊ౰ʹੜ͖ΔMLOpsͷφϨοδ͸اۀʹ͋Γ·͢ • େ͖ͳαʔϏε͔ͭେ͖ͳϞσϧΛӡ༻͠ͳ͍ͱΘ͔Βͳ͍τΠϧ͕͋Δ͸ͣ • ຊߨ࠲ͰऔΓѻͬͨൣғ͸ɺͦͷͨΊͷجຊͷʮΩʯͰ͢ • ςοΫϒϩάΛಡΜͩΓษڧձʹࢀՃ͢Δͷ͕ͱͯ΋ྑ͍ͱࢥ͍·͢ 32

Slide 33

Slide 33 text

3. LLMOpsͱ͸ʁ

Slide 34

Slide 34 text

LLMOpsͱ͸ʁ Large Language Model × Operations 34 Large Language Model Operations × DevOps for Large Language Model

Slide 35

Slide 35 text

LLMOpsͱ͸ʁ ஫໨͞Ε࢝Ί͍ͯΔLLMOps 35 ੨ઢ: MLOps  ੺ઢ: LLMOps

Slide 36

Slide 36 text

LLMOpsͱ͸ʁ Ops͕ੜ·ΕΔॠؒ • *Ops͕ੜ·ΕΔॠؒ • *Λ෼ׂͯ͠ద੾ʹ؅ཧ͠ͳ͍ͱ͍͚ͳ͍՝୊͕ग़͖ͯͨ࣌ • طଘͷ*Opsͱ͸ҟͳΔੑ࣭Λ͍࣋ͬͯΔ࣌ • LLM͸֬཰తͳϞσϧͰ͋Δͱಉ࣌ʹɺϓϩόΠμʹڧ͘ґଘ͢Δ • ඼࣭ͷ୲อ͕ඇৗʹॏཁʹͳͬͯ͘Δ 36

Slide 37

Slide 37 text

LLMOpsͱ͸ʁ ͓͢͢Ίͷษڧձ 37 MLOps/LLMOps/AgentOpsษڧձ • MLOps/LLMOpsͷੜ͖ͨφϨοδ͸اۀͷ ࣮ϓϩμΫτʹଟ෼ʹଘࡏ͢Δ • ಛʹLLOps, AgentOpsʹؔͯ͠͸୭΋ਖ਼ղΛ ͓࣋ͬͯΒͣɺ໛ࡧஈ֊ • (࠶ܝ) ςοΫϒϩά΍ษڧձʹࢀՃ͍ͯ͠ ͖·͠ΐ͏ • ٯʹݴ͑͹ίϛϡχςΟʹد༩͢Δνϟϯε

Slide 38

Slide 38 text

LLMOpsͱ͸ʁ ࣄྫ঺հ (ςετ/඼࣭อূ) 38 • גࣜձࣾIVRy͞Μ • LLMΛ༻͍ͨࣗಈి࿩ରԠαʔϏε • ϓϩόΠμʹڧ͘ґଘ͍ͯ͠ΔҎ্ɺ ϑΥʔϧόοΫઓུΛߟ͑Δඞཁ͕͋Δ • Ϟσϧ͕ߋ৽͞Εͨ࣌ɾϓϩϯϓτΛमਖ਼ ͨ࣌͠ͷ඼࣭Λ୲อ͢ΔςετΛͲ͏͢Δ ͔ • LLM as a Judge ࣮ӡ༻ͰֶΜͩԻ੠ର࿩γεςϜͷධՁͱςετ LLM APIΛ2೥ؒຊ൪ӡ༻ͯۤ͠࿑ͨ͠࿩

Slide 39

Slide 39 text

LLMOpsͱ͸ʁ ࣄྫ঺հ (؂ࢹج൫) 39 • גࣜձࣾαΠόʔΤʔδΣϯτ͞Μ • LLMͷ඼࣭͸མͪΔՄೳੑ͕ଟ෼ʹ͋Δ • ໌ࣔతͳΤϥʔ΋ͳ͘མͪΔͷͰɺؔ͠ج൫͕ॏཁ • LLMͷ୅දతͳ؂ࢹج൫ • LangFuse • PromptOpsతͳϓϩϯϓτͷόʔδϣχϯά • LLM as a judgeΛ༻͍ͨࣗಈςετ • LLMͷೖྗɾग़ྗͷ؂ࢹ Langfuseͷߏங

Slide 40

Slide 40 text

LLMOpsͱ͸ʁ ࣄྫ঺հ (ࣗಈԽ) 40 • גࣜձࣾαΠόʔΤʔδΣϯτ͞Μ • גࣜձࣾLayerX͞Μ • ϓϩϯϓτͷࣗಈ࠷దԽ • ͜Ε΋LLMOpsͷҰछͩͱࢥ͏ • ϓϩϯϓτΛίωίω͢Δͷ͸ଐਓతʹͳΓ͕͔ͪͭɺ ܾ·ͬͨܕ͕͋ΔͷͰࣗಈԽͰ͖ͦ͏ • LLM͸ϓϩϯϓτͱ͍͏Ϣʔβʔ͕ૢ࡞Ͱ͖Δॊೈͳύ ϥϝʔλ͕͋ΔͷͰ܇࿅ͳ͠ͰύʔιφϥΠζͰ͖Δ • In-Context Learning LLMͷϕϯνϚʔΫείΞΛ7෼ɺ100ԁͰ͋͛Δ AI Agent࣌୅ʹ͓͚Δʮ࢖͑͹࢖͏΄Ͳݡ͘ͳΔAIػೳʯͷ։ൃ

Slide 41

Slide 41 text

(ؓ࿩ٳ୊) LLMΞϓϦέʔγϣϯͷצॴ ϑϥΠϗΠʔϧ 41 • ϑϥΠϗΠʔϧ • AI͕Ϣʔβʔͷߦಈ΍ᅂ޷Λֶश͢Δ͜ ͱʹΑΓɺ࢖͑͹࢖͏͚ͩݡ͘ͳ͍ͬͯ ͘޷॥؀ • ઌड़ͷ௨ΓɺϓϩϯϓτΛ༻͍Ε͹͍ΖΜ ͳ͜ͱ͕Ͱ͖ͦ͏ • DeNA͞ΜͷYouTube΍ϓϩμΫτ͕  ษڧʹͳΓ·͢ Edge

Slide 42

Slide 42 text

4. MLOpsΛମݧͯ͠ΈΑ͏

Slide 43

Slide 43 text

ϋϯζΦϯͷείʔϓ ࣮ݧج൫ ֶशύΠϓϥΠϯ ਪ࿦ج൫ σʔλɾϞσϧͷόʔδϣϯ؅ཧ ܧଓతֶश ؂ࢹج൫

Slide 44

Slide 44 text

ϋϯζΦϯͷείʔϓ ࣮ݧج൫ ֶशύΠϓϥΠϯ ਪ࿦ج൫ σʔλɾϞσϧͷόʔδϣϯ؅ཧ ܧଓతֶश ؂ࢹج൫

Slide 45

Slide 45 text

ࠓճ࡞੒͢ΔύΠϓϥΠϯ YOLOΛ༻͍ͨը૾ೝࣝ 45 VertexAI Cloud Storage Cloud Build Artifact Registry Pipelines trigger Training Deployment Cloud Run Custom Jobs GitHub trigger Weights & Bias

Slide 46

Slide 46 text

Vertex AIͱ͸ʁ اۀ޲͚ͷ౷߹ܕAI/ػցֶशϓϥοτϑΥʔϜ • ओͳػೳ • Gemini API • ADKΛ༻͍ͨAI Agentͷߏங • Vertex AI Search (RAGͷΠϯσοΫεͱݕࡧ) • Ϟσϧ։ൃ • τϨʔχϯάɺσʔληοτɺFeature Storeɺςετ • τϨʔχϯάͰ͸ಠࣗͷύΠϓϥΠϯߏங͕Մೳ • ϞσϧͷσϓϩΠϝϯτ (ΤϯυϙΠϯτ) 46

Slide 47

Slide 47 text

Vertex AIͱ͸ʁ اۀ޲͚ͷ౷߹ܕAI/ػցֶशϓϥοτϑΥʔϜ 47 VertexAI Pipelines Training Deployment Custom Jobs • ύΠϓϥϯػೳΛ༻͍ͯ  ֶशύΠϓϥΠϯΛ੔උ • Custom jobΛఆٛͰ͖ΔͷͰ  ਪ࿦αʔόʔ΁ͷϞσϧͷσϓϩΠ·Ͱ • +α • σϓϩΠલʹϞσϧͷݕূΛߦ͏ • σʔλͷલॲཧΛߦ͏

Slide 48

Slide 48 text

όʔδϣχϯάͱ࣮ݧ؅ཧ Cloud Storage / Weights and Bias 48 Cloud Storage Weights & Bias • Cloud StorageΛ༻͍ͯσʔληοτɾϞσϧͷ  όʔδϣχϯάΛߦ͏ • σʔληοτ͸ࠓճ͸खಈͰόʔδϣχϯάϞσϧ͸ Cloud BuildͷBuildIDͱରԠ͚ͮͯόʔδϣϯχϯά • +α • DVCΛಋೖͯ͠ΈΔ • Vertex AIͷσʔληοτɾFeature StoreػೳΛ  ࢖ͬͯΈΔ

Slide 49

Slide 49 text

Vertex AIͱ͸ʁ GitHub͔ΒͷࣗಈτϦΨʔ / CloudRun΁ͷσϓϩΠ 49 • GitHub΁ͷ܇࿅ઃఆϑΝΠϧͷpushΛτϦΨʔʹֶशύΠϓ ϥΠϯશମΛτϦΨʔ • CloudBuildΛ༻͍ͯ܇࿅ͱਪ࿦༻ͷdocker imageΛϏϧυ • ਪ࿦ج൫ʹ͸Cloud RunΛ࠾༻ • ࠷ۙ GPU on Cloud Run͕GAʹ • +α • GitHub ActionsͰtest/linter/formatterΛೖΕͯΈΔ • Vertex AI EndpointsΛ࢖ͬͯΈΔ Cloud Run Cloud Build GitHub trigger

Slide 50

Slide 50 text

ϋϯζΦϯ YOLOΛ࢖ͬͨΩϟϥΫλʔͷը૾ೝࣝ 50 • ͪ͜ΒͷϨϙδτϦΛ༻͍ͯϋϯζΦϯΛߦ͍·͢ • https://github.com/ogatakatsuya/mlops-distribution • ຊ౰͸ҰॹʹਐΊ͍ͯ͘༧ఆͰ͕ͨ͠ɺ  ܇࿅ʹࢥͬͨΑΓ͕͔͔࣌ؒΔͷͰɺ࣮ࡍʹ  खΛಈ͔͔͢Ͳ͏͔͸ࣗݾ൑அͰ͓ئ͍͠·͢🙏 • ࢖͍෺ʹͳΔ͘Β͍܇࿅͠Α͏ͱࢥ͏ͱ  ͔ͳΓ͕͔͔࣌ؒΓ·͢⚠

Slide 51

Slide 51 text

ϋοΧιϯʹ޲͚ͯ 51 • MLΛϓϩμΫτʹ૊ΈࠐΉͱ΍ΕΔ͜ͱ͕޿͕Γ·͢ • ಛʹίϯϐϡʔλϏδϣϯܥ(YOLO΋)͸΍ΕΔ͜ͱ͕ଟ͍Ͱ͢ • ߟ͑Δ͜ͱ͸ • σʔληοτΛߏஙͰ͖Δ͔(܇࿅͠ͳ͍ͳΒ͍Βͳ͍) • ӡ༻Ͱ͖Δ͔ (CPU্Ͱಈ͘ͳΒͳ͓ྑ͍) • ࣮૷ͷΠϝʔδ͕͔ͭ͘ • MLOpsతͳ؍఺ΛೖΕΔ͔Ͳ͏͔͸৻ॏʹɻɻ • Ͱ΋ҙࣝͰ͖ͯΔͱධՁ͕ྑͦ͞͏ʁ • ϓϨθϯʹ͔ͬ͠Γ૊ΈࠐΈ·͠ΐ͏

Slide 52

Slide 52 text

;Γ͔͑Γ 52 • MLOpsͱ͸ʁ • Machine Learning Operations / DevOps for ML • MLOpsͷߏ੒ཁૉ • ֶशύΠϓϥΠϯɾόʔδϣχϯάɾ؂ࢹج൫ɾܧଓతֶशɾɾɾ • LLMOpsͱ͸ʁ • LLM Operations / DevOps for LLM • ϋοΧιϯؤு͍ͬͯͩ͘͞💪

Slide 53

Slide 53 text

ࢀߟจݙ • CyberAgent AIࣄۀຊ෦2024೥౓MLOpsݚमجૅฤ • MLOps࣮ફΨΠυ -ຊ൪ӡ༻Λݟਾ͑ͨ։ൃઓུ- 53