You are on page 1of 13

Experiment, Collaboration, Deployment

for ML & AI Workflow

Let’s simplify it…

- Kuldeep Singh
About Me
Mr Kuldeep Singh is a seasoned AI practitioner and ML architect, skilled in
Data Science, DevOps Implementation and Cloud Infrastructure setup with
Agile mindset. With 12+ years of experience, he brings a unique blend of
expertise that assist him to integrate the business domain with traditional as
well as emerging new IT landscape. He has been involved with the world's
leading consulting and tech firms, and holds an MBA from IMT along with
PGP in Big Data and Machine Learning from Great Lakes.

...Paired with very lovingly wife and father of two daughters.

https://www.linkedin.com/in/iamkuldeepsingh

k.duhan@gmail.com
Agenda
● Why?
● DVC
● Docker
● Kubernetes
● ML Workflow Integration
● Use Cases
● Reference and Resources
● QnA
Why
● Can AI produce outcome only by running code in notebook ?
● Do you think DS workflow is very easy to reuse ?
● Do you have consistent experience from local to large scale deployment ?
● Can you quickly integrate it with IT landscape for production usage ?
● In short... can you build, experiment, deploy in production scale easily ?
DVC
● Build to make ML models shareable and reproducible
● Designed to handle large files. Data sets, ML models, metrics as well as code and
version control it
● Full code and date provenance help track the complete evolution which guarantees
reproducibility and make the switch easy
● Takes advantage of Git branches to try different ideas along with metrics and ML
pipelines
● Use push/pull commands to move consistent bundles of ML models, data and code
into production.
● Connect multiple steps into a DAG, creates pipeline which helps in remove friction
to get code into production
Features
● Git compatible
● Storage agnostic
● Reproducible
● Low friction branching
● Metric tracking
● ML pipeline framework
● Language and framework agnostic
● Track failures

Image Source DVC


Workflow
Docker

● Ensure all dependencies are on one place


● Provides consistent experience
● Easy to debug
● Immutable in nature
● Quick to experiment
● Large community base

Image Source Docker


Kubernetes

● Defacto for Container deployment


● Unified interface for development and deployment
● Abstract Infrastructure details
● Scaling is very smooth
● Very popular in Tech community

Image Source TheNewStack


ML Workflow Integration
Sample integration

Image Source Kubeflow


Use cases
● Save and reproduce your experiments
● Version control models and data
● Establish workflow for deployment, collaboration and deployment
● Consistency and Scalability
● Integration with IT ecosystem
● One language to speak for Software Engg, Data Engg and Data Scientist
Reference and Resources

● DVC Get Started https://dvc.org/doc/start


● DVC Blog https://dvc.org/blog
● Docker Get Started https://docs.docker.com/get-started/
● MiniKube Setup https://minikube.sigs.k8s.io/docs/start/
● Overall Kubernetes https://kubernetes.io/docs/tutorials/
QnA

Thank You

You might also like