SLIDE 1
S91030 - Hybrid Machine Learning with the Kubeflow Pipelines and RAPIDS
Sina Chavoshi
The right approach for the right problem
Building blocks Platform Solutions
S91030 - Hybrid Machine Learning with the Kubeflow Pipelines and - - PowerPoint PPT Presentation
S91030 - Hybrid Machine Learning with the Kubeflow Pipelines and - - PowerPoint PPT Presentation
S91030 - Hybrid Machine Learning with the Kubeflow Pipelines and RAPIDS Sina Chavoshi Cloud AI Strategy: The right approach for the right problem Building blocks Platform Solutions Cloud AI Strategy: The right approach for the right
SLIDE 2
SLIDE 3
The right approach for the right problem
Building blocks Platform Solutions
Cloud AI Strategy:
SLIDE 4
Building Blocks
Sight Language Conversation
SLIDE 5
The right approach for the right problem
Building blocks Platform Solutions
Cloud AI Strategy:
SLIDE 6
Solutions / Contact Center
Customer Phone Chat Contact Center Provider Contact Center Interface Virtual Agent Agent Assist Knowledge Base (PDF/HTML) Backend Fulfillment Virtual Agent Agent Google Cloud Contact Center AI
SLIDE 7
The right approach for the right problem
Building blocks Platform Solutions
Cloud AI Strategy:
SLIDE 8
Cloud AI Platform
Data pipeline
Cloud Dataprep BigQuery Cloud Dataflow Cloud Dataproc
Model development
Cloud ML Engine
Model deployment and management
Cloud ML Engine Cloud Kubernetes Engine
Tools
Jupyter Notebooks
Services
ASL
Community
Kubeflow
SLIDE 9
Building & deploying real-life ML applications is hard and costly because of lack of tooling that covers end-to-end ML development & deployment.
SLIDE 10
In addition to the actual ML...
ML Code
SLIDE 11
You have to worry about so much more.
Configuration Data Collection Data Verification Feature Extraction Process Management Tools Analysis Tools Machine Resource Management Serving Infrastructure Monitoring
ML Code Source: Sculley et al.: Hidden Technical Debt in Machine Learning Systems
SLIDE 12
02 03
AI problems today
Problems Solutions
Deployment Brittle, opinionated infrastructure that is hard to productionize and breaks between cloud and on-prem Talent Machine Learning expertise is scarce Collaboration Difficult to find, leverage existing solutions
Reusable pipelines
01 02 03
SLIDE 13
01: Kubeflow
Scalable ML services on Kubernetes
Easy to get started
- Out-of-box support for top frameworks
– pytorch, caffe, tf and xgboost
- Kubernetes manages dependencies, resources
Swappable & scalable
- Library of ML services
- GPU support
- Massive scale
Meet customer where they are
- GCP
- On-prem with Cisco
Cloud On-prem
Training
ML microservices
Predict Training Predict
SLIDE 14
Product Overview
RAPIDS
SLIDE 15
THE BIG PROBLEM IN DATA SCIENCE
All Data ETL
Manage Data
Structured Data Store Data Preparation
Training
Model Training Visualization
Evaluate
Scoring
Deploy Slow Training Times for Data Scientists
SLIDE 16
RAPIDS — OPEN GPU DATA SCIENCE
Software Stack Python
Data Preparation
cuDF
Graph Analytics
cuGRAPH
Model Training
cuML
CUDA PYTHON APACHE ARROW on GPU Memory DASK/SPARK DEEP LEARNING FRAMEWORKS CUDNN RAPIDS CUML CUDF CUGRAPH
SLIDE 17
BENCHMARKS
cuML — XGBoost End-to-End cuIO/cuDF — Load and Data Preparation Benchmark
200GB CSV dataset; Data preparation includes joins, variable transformations.
CPU Cluster Configuration
CPU nodes (61 GiB of memory, 8 vCPUs, 64-bit platform), Apache Spark
DGX Cluster Configuration
5x DGX-1 on InfiniBand network Time in seconds — Shorter is better
cuIO / cuDF (Load and Data Preparation) Data Conversion XGBoost
SLIDE 18
AI Hub & Pipelines: Fast & simple adoption of AI
- 5. Publish
Upload & share pipelines running best within your org or publicly.
- 1. Search & Discover
Find best-of-breed solutions on the AI Hub which leverage Cloud AI solutions
- 2. Deploy
Quick 1-click implementation of ML pipelines onto Google Cloud Platform .
- 4. Run in production
Deploy customized pipelines in production.
- 3. Customize
Experiment and adjustment
- ut-of-the-box pipelines to custom use
cases. Network effect
The Flywheel of AI Adoption
SLIDE 19
02: Reusable Pipelines
Enable developers to build custom ML applications by easily “stitching” and connecting various components.
- Reuse instead of reimplement or reinvent
- Discover, learn and replicate successful pipelines
SLIDE 20
What constitutes a Kubeflow Pipeline
- Containerized implementations of ML Tasks
○ Containers provide portability, repeatability and encapsulation ○ A task can be single node or *distributed* ○ A containerized task can invoke other services
- Specification of the sequence of steps
○ Specified via Python SDK
- Input Parameters
○ A “Job” = Pipeline invoked w/ specific parameters
SLIDE 21
03: AI Hub at a glance
All AI content in one place Quick discovery of plug & play AI pipelines & other content built by teams across Google and by partners and customers. Fast & simple implementation of AI on GCP One-click deployment of AI pipelines via Kubeflow on GCP as the go-to platform for AI + hybrid & on premise. Enterprise-grade internal & external sharing Foster reuse by sharing deployable AI pipelines & other content privately within organizations & publicly. 1 2 3
SLIDE 22
Mission
The one place for everything AI, from experimentation to production.
SLIDE 23
Public and private AI Hub
By Google Unique AI assets by Google By partners Created, shared & monetized by anyone By customers Content shared securely within and with other organizations
Public content + Private content
AutoML, TPUs, Cloud AI Platform, etc.
SLIDE 24
Kubeflow Pipelines enable
Workflow
- rchestration
Rapid reliable experimentation Share, re-use & compose
SLIDE 25
Demo
SLIDE 26
Visual depiction of pipeline topology
SLIDE 27
View all current and historical runs, grouped as “Experiments”
SLIDE 28
Rich visualizations of metrics
SLIDE 29
Clone an existing pipeline
SLIDE 30
Access to all config params, inputs and outputs for each run
SLIDE 31
Update parameters and submit
SLIDE 32
Easy comparison of Runs
SLIDE 33
Easy comparison of Runs
SLIDE 34
SLIDE 35