아이펠 양재 해커톤 MLOps

Transcript

1. ӣঙ੢ MLOps Kubernetesܳ ੉ਊೠ উ੿੸ੋ ࢲ࠺झ ਍৔җ प೷ ߂ ߓನ ੗زച
2. ݾର • ޖ঺ਸ ೞҊ੗ ೞחо? • যځೠ ࢶఖ૑о ੓঻חо? • पण
3. ޖ঺ਸ ೞҊ੗ ೞחо? ੷ח ױࣽ൤ ࠙ܨ݅ ೞҊ रਸࡺੋؘਃ??
4. MNIST ࠙ܨܳ ਤೠ ࢲ࠺झ • https://github.com/Ssuwani/mnist-classi fi cation-service Training Trained Model Serving API Web Application
5. MNIST ࠙ܨܳ ਤೠ ࢲ࠺झ ੉ۧѱ рױೞѱ ೡ ࣻ ੓חؘ? • https://github.com/Ssuwani/mnist-classi fi cation-service 👈 ৈӝоࢲ पण ೧ࠁӝ http://182.161.136.200:3000/
6. MNIST ࠙ܨܳ ਤೠ ࢲ࠺झ ࠗೞ పझ౟ܳ ਤೠ Tool, Locustܳ ੉ਊೠ ౟ې೗ పझ౟ • https://locust.io/ 👈 ੉Ѣب рױ൤ पण೧ࠁ੗ self.client.post(‘/prediction’ , fi les={“ fi le": img})
7. MNIST ࠙ܨܳ ਤೠ ࢲ࠺झ ࠗೞ పझ౟ܳ ਤೠ Tool, Locustܳ ੉ਊೠ ౟ې೗ పझ౟ • https://locust.io/ 👈 ੉Ѣب рױ൤ पण೧ࠁ੗ self.client.post(‘/prediction’ , fi les={“ fi le": img}) ഻ঁ ؊ ݆਷ ࢎਊ੗о ੽ࣘೞח ࢲ࠺झۄݶ ੉ۧѱ ߓನ೧ب ҡଳ਷ Ѧөਃ??
8. ௢ߡ֎౭झ ஶప੉ցചػ গ೒ܻா੉࣌੄ ੗ز ٣೒۽੉, झாੌ݂ ١ਸ ઁҕೞח ҙܻदझమ • ц੗ӝ ౟ې೗੉ ૐоೞݶ য૵૑? AutoScaler • ౟ې೗ী ٮۄ ੸੺ೠ ݅ఀ੄ ߓನܳ ೡ ࣻ ੓ب۾ ب৬ળ׮ • ࢜۽਍ ߡ੹ਸ ߓನೞҊ रয? ׮নೠ ߓನ ߑߨٜ(܀݂ সؘ੉౟)
9. Work on Kubernetes Auto Scaler • Horizontal Pod AutoScaler -> प೯غח Pod੄ ࣻܳ טܽ׮ • Vertical Pod AutoScaler -> Podী ೡ׼غח ܻࣗझܳ טܽ׮ • Cloud AutoScaler -> प೯غח Node ࣻܳ טܽ׮ (௿ۄ਋٘ ੋ೐ۄ৬ োز)
10. Work on Kubernetes Horizontal Pod AutoScaler(HPA)
11. Work on Kubernetes ܀݂ সؘ੉౟ • ܀݂ সؘ੉౟ • ࠶ܖ/Ӓܽ ߓನ • ஠աܻ ӝߨ 👈 ࢜۽਍ ߡ੹ਸ ߓನೞҊ रযਃ!
12. Work on Kubernetes ܀݂ সؘ੉౟
13. Work on Kubernetes ܀݂ সؘ੉౟
14. Work on Kubernetes ܀݂ সؘ੉౟
15. Work on Kubernetes ܀݂ সؘ੉౟
16. Work on Kubernetes ܀݂ সؘ੉౟
17. Work on Kubernetes Why Docker ݅ড AIFFEL Үਭࢤ 100ݺ੄ ѐߊജ҃ਸ ૒੽ ࣇ౴೧઻ঠ ೠ׮ݶ.. Fundamental Exploration GoingDeeper
18. Work on Kubernetes Why Docker 👈 ഑द ೧ழహ ೞदݶࢲ ѐߊജ҃ ࢸ੿ ٸޙী ൨٘द૑ ঋਵ࣑աਃ?? ݅ড ղо ۽ஸীࢲ ز੘೮؍ ௏٘о ௢ߡ֎౭झীࢲ ز੘ೞ૑ ঋח׮ݶ..? Github Push Kubernetes Hyper Parameter Tuning
19. CI & CD & CT ௏٘ ۨ߰ীࢲ੄ CT • CI(Continuous Integration) ׮ࣻ੄ ѐߊ੗о ੘ࢿ, ࣻ੿ೠ ࣗझ௏٘ܳ ૑ࣘ੸ਵ۽ ా ೤, పझ౟ೞח Ѫ • CD(Continuous Delivery/Deployment)ۆ ѐߊ, ా೤, ߓನ, ܾܻૉ, పझ౟ܳ ੗زച ೞৈ ૑ࣘ੸ਵ۽ ߓನೞח Ѫ • CT(Continuous Training)ۆ Production ױ҅ীࢲ ࢜۽਍ ؘ੉ఠ۽ ೟णೞৈ ࢜۽਍ ݽ ؛ਸ ߓನೞח Ѫ
20. CI & CD & CT ௏٘ ۨ߰ীࢲ੄ CT • CI(Continuous Integration) ׮ࣻ੄ ѐߊ੗о ੘ࢿ, ࣻ੿ೠ ࣗझ௏٘ܳ ૑ࣘ੸ਵ۽ ా ೤, పझ౟ೞח Ѫ • CD(Continuous Delivery/Deployment)ۆ ѐߊ, ా೤, ߓನ, ܾܻૉ, పझ౟ܳ ੗زച ೞৈ ૑ࣘ੸ਵ۽ ߓನೞח Ѫ • CT(Continuous Training)ۆ Production ױ҅ীࢲ ࢜۽਍ ؘ੉ఠ۽ ೟णೞৈ ࢜۽਍ ݽ ؛ਸ ߓನೞח Ѫ
21. MLOps? MLOps
22. MLOps? MLOps Machine Learning Model Operationalization Management Machine Learning Operations
23. ݠन۞׬ ݽ؛ਸ উ੿੸੉Ҋ ബਯ੸ਵ۽ ߓನ ߂ ਬ૑ ҙܻೞח Ѫਸ ݾ಴۽ ೞח ੌ۲੄ ҙ೯ ਤఃೖ٣ই - https://en.wikipedia.org/wiki/MLOps
24. MLOps? ठۄ੉٘ ୹୊ http://ndcreplay.nexon.com/NDC2017/sessions/NDC2017_0044.html (ӣకഅ ש)
25. MLOps? https://cloud.google.com/architecture/mlops-continuous-delivery-and-automation-pipelines-in-machine-learning
26. ݠन۞׬ ݽ؛ਸ উ੿੸੉Ҋ ബਯ੸ਵ۽ ߓನ ߂ ਬ૑ ҙܻೞח Ѫਸ ݾ಴۽ ೞח ੌ۲੄ ҙ೯ ਤఃೖ٣ই - https://en.wikipedia.org/wiki/MLOps
27. MLOps? https://cloud.google.com/architecture/mlops-continuous-delivery-and-automation-pipelines-in-machine-learning
28. MLOps Pipeline ҳ୷ೞӝ (ѹ਋?) mnistܳ ߓನೞӝ ਤೠ ৈ੿
29. ੉؀۽݅ ೞݶ غѷחؘ? https://cloud.google.com/architecture/mlops-continuous-delivery-and-automation-pipelines-in-machine-learning
30. ੉۠ ోਸ ॳݶ غѷ֎? https://towardsai.net/p/machine-learning/mlops-demysti fi ed
31. ੉۠ ోب ॳݶ જѷ֎??? https://jumpcap.com/insights/mlops-the-key-to-scaling-machine-learning-in-an-enterprise
32. ੉۠ ోب ॳݶ…? https://neptune.ai/mlops-tool-stack
34. য۰ਛ؍ ੼ • ࢤࣗೠ ѐ֛ٜ • ցޖ ݆਷ ࢶఖ૑ • ই૒ ৮߷ೞ૑ ঋ਷ بҳٜ Container Orchestration Google Cloud Service Container Kubernetes Docker GitOps Continues Integration Continues Training Continues Deployment Model Serving Model Serving MLOps
35. য۰ਛ؍ ੼ • ࢤࣗೠ ѐ֛ٜ • ցޖ ݆਷ ࢶఖ૑ • ই૒ ৮߷ೞ૑ ঋ਷ بҳٜ
36. য۰ਛ؍ ੼ • ࢤࣗೠ ѐ֛ٜ • ցޖ ݆਷ ࢶఖ૑ • ই૒ ৮߷ೞ૑ ঋ਷ بҳٜ KF Serving
37. য۰ਛ؍ ੼ • ࢤࣗೠ ѐ֛ٜ • ցޖ ݆਷ ࢶఖ૑ • ই૒ ৮߷ೞ૑ ঋ਷ بҳٜ
 (਋ܻীѲ য়൤۰ જ׮?) Contribution ӝഥо ੓ਸ૑ب..? ੷ח पಁ೮૑݅…
38. mnist۽ MLOps ೧ࠁӝ
39. ݾ಴ • Experiment Pipeline ҳ୷ • Hyper parameter Tuning • Experiment Tracking • Deploy Pipeline ҳ୷ • Best Hyper Parameterܳ ੉ਊೠ ೟ण • ೟णػ ݽ؛ ߓನ
40. Stack Github Action docker build & push Github Packages kustomize set image Argo CD Goole Cloud Storage Bento ML Slack App Kubernetes Github Data Version Control
41. Experiment Pipeline Github Actions Github Push Docker Build & Push Kustomize Set Image Version ArgoCD Github Package Kubernetes Google Storage Data Version Control Load Data Hyper Parameter Tuning Training Job Check Code Conventions 👀 Watching Bot
43. Experiment Pipeline Github Actions Github Push Docker Build & Push Kustomize Set Image Version ArgoCD Github Package Kubernetes Google Storage Data Version Control Load Data Hyper Parameter Tuning Training Job Check Code Conventions 👀 Watching Bot
44. https://kangwoo.kr/2020/03/21/kube fl ow-katib-metrics-collector/
45. Flow Experiment Pipeline Github Hyper Parameter Tuning Experiment Tracking ML Researcher Experiment Tracking Slack Noti fi cation
46. Deploy Pipeline Github Actions Kustomize Set Image Version & Hyper Parameter ArgoCD Github Kubernetes Google Storage Load Data Service API Save Model Web App Request Response Training Job 🤔 Deploy Model Github Actions Packing Bento ml Docker Build & Push Kustomize Set Image Version Treining-Job Deploy, Service
47. Flow Deploy Pipeline Model Registry Model Serving Experiment Tracking Model Training Slack Action API Service
48. Flow MLOps Experiment Pipeline Deploy Pipeline Model Monitoring Data
49. Flow MLOps Experiment Pipeline Deploy Pipeline Model Monitoring Data
50. दো
54. ؊ ࠁৈ઱Ҋ र૑݅ ׸૑ ޅೠ Ѫ • DVCܳ ੉ਊೠؘ੉ఠ ߡ੷׬ • Node pool ܻ࠙ • Katib प೷
55. ഥҊ
56. જও؍ ੼ • ౠ൤ MLOps౱ীѱ GCP ૑ਗ਷ ցޖ ੸੺೮׮. • ഒ੗೮׮ݶ ನӝೞӝ ٯ જ਷ ੌٜ੉ ݆ওחؘ যڌѱٚ ೧Ѿ೮׮. • दبೞҊ पಁ೮؍ ҃਋ب ݆ওӝ ٸޙী ૑Әࠁ׮ ѐࢶೡ ࢎ೦ٜ੉ ݆׮.
57. ইए਍ ੼ • Ai ff el җ੿઺ী ҕࠗ೮؍ ղਊٜ੉ ൞޷೧ઉр׮.
 ( mnist߆ী ӝর੉ উա…) • Feature store, Performence monitoring ١١ MLOps੄ ׮ܲ ࠗ࠙ਸ ҃೷೧ࠁ૑ ޅ ೮׮ • ѐ֛ਸ ঌও׮ݶ औѱ ೧Ѿ೮ਸ ޙઁܳ ೠଵزউ Ҋ޹ೞҌ೮׮.
59. Ѿҗޛ • https://github.com/KIMnJANG • http://35.238.80.73/