aws-sageMaker

Transcript

1. SageMaker - Building & Deploying an ML Model Ofir Naor, Data Science & Backend Team Leader Amenity Analytics
2. 1 / 35 Agenda • Intro to Amenity Analytics • Studio - an NLP cloud IDE • SageMaker - build & deploy ML model
3. Finance Documents NLP Models & Analysis Structured Extractions API & Dashboard results SUCCESS Intro to Amenity Analytics 2 / 35
4. 4 3 / 35 Goals 100x Faster 100x Clients 100x More data
5. 5 / 35 ALL IN ON SERVERLESS! NewArc
6. 5 / 35 Only few weeks old .. Amenity Studio - IDE for NLP
7. 5 / 35 .. and already recognizer linguistic patterns Amenity Studio - IDE for NLP
8. 5 / 35 This is Amenity Studio. It got lemma_by_sample(‘love’) SageMaker - Amenity Studio
9. 5 / 35 Let’s add a new cool function! synonym_by_context(‘love’, sentence) SageMaker - Amenity Studio
10. 5 / 35 lemma_by_group(‘love’) SageMaker - Amenity Studio synonym_by_context (‘love’, sentence) affection appreciation devotion emotion fondness friendship ... sex it depends - I love Apple - like fond of appreciate … no sex I made love with an apple. sex.
11. 5 / 35 Looking for context? SageMaker - BERTTTT
12. 5 / 35 Meet BERT SageMaker - BERTTTT
13. 5 / 35 BERT has changed the world... of NLP Bidirectional Encoder Representations from Transformers SageMaker - BERTTTT
14. 5 / 35 BERT Training SageMaker - BERTTTT Input: the man went to the [MASK1] . he bought a [MASK2] of milk. Labels: [MASK1] = store; [MASK2] = gallon
15. 5 / 35 BERT Training SageMaker - BERTTTT Input: I [MASK] Apple. Predict: like, appreciate, ate (if uncased) Input: I made [MASK] with an apple. Predict: sex
16. 5 / 35 Let’s learn how to deploy it Using SageMaker SageMaker - BERTTTT
17. 5 / 35 SAGEMAKER AT AMENITY - STATUS SageMaker - Status
18. 5 / 35 We don’t use it. SageMaker - Status
19. 5 / 35 We host SAGEMAKER meetup What shall we present? SageMaker - Presentation
20. 5 / 35 Lets try SAGEMAKER again SageMaker - Presentation
21. 5 / 35 We got data scientists? Researchers? SageMaker - Who we got?
22. 5 / 35 .. and it makes everyone around them MAD SageMaker - Who we got?
23. 5 / 35 Is SageMaker the solution? Lets PoC SageMaker - Presentation
24. 5 / 35 Deploying a custom container in SAGEMAKER SageMaker - Custom Containers
25. 5 / 35 Does it worth it? (effort, deployment time, devops, constraints?) SageMaker - Custom Containers
26. 5 / 35 SageMaker - Researchers.. One-Class © Framework
27. 5 / 35 SageMaker - Researchers.. One-Class © Framework
28. 5 / 35 Use a Framework. SageMaker - DevOps.. They say.
29. 5 / 35 SageMaker - Data scientist.. Based on Amazing Technology Deep Copy Deep Paste Deployable on AWS
30. 5 / 35 SageMaker - Data scientist.. It’s not the actual model run that is complicated I run a deep learning model !
31. 5 / 35 SageMaker - Data scientist.. .. but the pre-process ..
32. 5 / 35 SageMaker - Data scientist.. .. and the post-process
33. 5 / 35 SageMaker - FullStack, AWS certified, Data scientist.. .. and the deployment step .. and the deployment steps (that’s how I felt - good & safe at start, then it hurt)
34. 5 / 35 • Git clone • Then another sub-clone • docker build • another docker build • setup.py dependencies SageMaker - FullStack, AWS certified, Data scientist.. From what I remember
35. 5 / 35 • … fix dependencies • deps also in Dockerfile • rebuild docker • not that one - the first one • now build bdist_wheel • now rebuild docker From what I remember SageMaker - FullStack, AWS certified, Data scientist..
36. 5 / 35 • deploy try#1 (8 minutes) • fix • repeat • … after ~10 deployments • integration test passes From what I remember SageMaker - FullStack, AWS certified, Data scientist..
37. 5 / 35 Now CDK, automation, CI/CD …. Does it worth it? From what I remember SageMaker - FullStack, AWS certified, Data scientist..
38. 5 / 35 As a DataScientist / researcher If that’s the given way, only way SageMaker - Summary Why would someone choose to serve on SageMaker endpoint over EC2 / ECS / Fargate ?
39. 5 / 35 As Devops - If it combines well with the whole training / serving lifecycle .. Or as the 1st iteration SageMaker - Summary Why would someone choose to serve on SageMaker endpoint over EC2 / ECS / Fargate ?
40. 9 / 35
41. 5 / 35 Console Vs. Script Vs. Cloudformation • • • • • • • • • • • • •
42. 5 / 35 AWS Cloud Development Kit (AWS CDK)
43. 5 / 35 AWS Cloud Development Kit (AWS CDK)
44. 5 / 35 AWS Cloud Development Kit (AWS CDK)
45. 5 / 35 AWS Cloud Development Kit (AWS CDK)
46. 5 / 35 AWS Cloud Development Kit (AWS CDK)
47. 5 / 35 AWS Cloud Development Kit (AWS CDK)
48. 5 / 35 AWS Cloud Development Kit (AWS CDK)
49. 5 / 35 AWS Cloud Development Kit (AWS CDK)
50. 5 / 35 AWS Cloud Development Kit (AWS CDK)
51. 5 / 35 References and credits • https://aws.amazon.com/blogs/developer/aws-cdk-developer-preview/ • https://github.com/peerjako-aws/cdk-mlops • • • https://carbon.now.sh •