continuous machine learning - ai ukraine · easily deploy and run datacenter-wide app services such...

Continuous Machine Continuous Machine LearningLearning

AI Ukraine’16AI Ukraine’16

Kostiantyn Bokhan, PhDKostiantyn Bokhan, PhDProject Lead at Samsung R&D UkraineProject Lead at Samsung R&D UkraineKharkiv, October 2016Kharkiv, October 2016

●ML dev. workflows ML dev. workflows

●ML dev. issuesML dev. issues

●ML dev. solutions ML dev. solutions

●Continuous machine learning (CML)Continuous machine learning (CML)

●Aspects of CMLAspects of CML

●CML infrastructureCML infrastructure

●CML – deliveryCML – delivery

Agenda Agenda AI Ukraine’16AI Ukraine’16

K.BokhanK.Bokhan 2

ML dev. workflows ML dev. workflows AI Ukraine’16AI Ukraine’16

K.BokhanK.Bokhan 3

RequirementsRequirements PlanningPlanning

Gathering Datasets

Gathering Datasets

Featuredesign Featuredesign

Model TrainingModel

TrainingModel

ValidationModel

ValidationModel TestingModel Testing

Train framework (Python/R/Matlab)

TestsTests FEFE Modeltools

Modeltools ClassifierClassifier UIUI

QAQA MarketMarket

Application (C++/Java)


K.BokhanK.Bokhan 4

Define the Problem

Define the Problem

Gather DatasetsGather

Datasets

Clean DatasetsClean

Datasets

Visualize,Explore

Visualize,Explore

Measure,EvaluateMeasure,Evaluate

Hypothesize,Model

Hypothesize,Model

DeployDeploy

Data selection

Data selection

Data transformation

Data transformation

Feature design


K.BokhanK.Bokhan 5

Datasets

Train data

Test data

labels

FeatureExtraction

Trainingthe

model

Evalthe

modelModelfeatures

Training

Input Data

FeatureExtractio

n PredictLabels

Modelfeatures

Predicting

ML dev. issues ML dev. issues AI Ukraine’16AI Ukraine’16

K.BokhanK.Bokhan 6

SmallBig

Bigger

Complex

Simple

Complex

DataSize

Modelcomplexity

1996 2006 2016


K.BokhanK.Bokhan 7

Perf

orm

ance

Perf

orm

ance

Machine Learningon PC


K.BokhanK.Bokhan 8

Perf

orm

ance

Perf

orm

ance


My other computer is Amazon EC2

Machine Learning on AWS


K.BokhanK.Bokhan 9

Perf

orm

ance

Perf

orm

ance


Machine Learning on dedicated cluster

My other computer is Amazon EC2

Machine Learning on AWS


K.BokhanK.Bokhan 10

Unce

rtain

tyU

nce

rtain

ty


K.BokhanK.Bokhan 11

Vari

ety

Vari

ety


K.BokhanK.Bokhan 12

Relia

bili

tR

elia

bili

tyy


K.BokhanK.Bokhan 13

Reso

urc

e

Reso

urc

e

man

ag

em

en

tm

an

ag

em

en

t

ML dev. solutions ML dev. solutions AI Ukraine’16AI Ukraine’16

K.BokhanK.Bokhan 17

+

Performance Performance


K.BokhanK.Bokhan 18

+

Performance + Reliability Performance + Reliability

Mesos


K.BokhanK.Bokhan 19

Variety Variety


K.BokhanK.Bokhan

Resource management Resource management

20

Memory CPU Storage GPU

Kernel →

Init.rd → Marathon

Mesos

dcos CLI Singularity Aurora

MPI

CUDA

Theano


K.BokhanK.Bokhan 22

Source: https://dcos.io/docs/1.8/overview/architecture/

Mesosphere Enterprise DC/OS is an enterprise grade datacenter-scale operating system, providing a single platform for running containers, big data, and distributed apps in production.

Services & Applications

Easily deploy and run datacenter-wide app services such Docker, Cassandra, Spark pooled on a single platform

DC/OS Powered by Apache Mesos

Runtime, tools and best practices built-in to simplify operations and deliver a production self-healing infrastructure

Run Anywhere

Bare-metal, virtual, cloud or hybrid – DC/OS runs on it all – only requirement is a modern Linux distro, Windows support coming soon :)


K.BokhanK.Bokhan 23

Source: https://dcos.io/docs/1.8/overview/architecture/


K.BokhanK.Bokhan 24

Apache Mesos is the open-source distributed systems kernel at the heart of the Mesosphere DC/OS. It abstracts the entire datacenter into a single pool of computing resources, simplifying running distributed systems at scale.

Sources: http://nvidia.com, http://blog.arungupta.me/docker-apache-mesos-marathon/

CML Resources CML Resources AI Ukraine’16AI Ukraine’16

K.BokhanK.Bokhan 25

Virtual nodesBare-metal nodes

with GPUTest devices

Hybrid


K.BokhanK.Bokhan

Uncertainty Uncertainty

26

Continuous Machine Continuous Machine LearningLearning

We can’t remove uncertainty but we can automate routines We can’t remove uncertainty but we can automate routines especially delivery and integrationespecially delivery and integration

Aspects of CML Aspects of CML AI Ukraine’16AI Ukraine’16

K.BokhanK.Bokhan 28

Continuous

Continuous

Development

Continuous

Integration Continuous

Deployment

Continuous Delivery

Continuous Everything

CML Infrastructure CML Infrastructure AI Ukraine’16AI Ukraine’16

K.BokhanK.Bokhan 31

Train Validation Test

Build Test Verification

GIT Jenkins Dockerregistry

Developers

Data scientists

DevelopersDevelopers

MesosMaster

StandbyMaster

StandbyMaster

MesosAgent

Spark job

Batchdocker

jobSparkCUDA

job

MesosAgent

SparkDocker

job

MesosAgent

Spark job

CUDA job

MesosAgent

Singularity Marathon

Singularity Marathon

Pool ofDevices

http://nvidia.com/

CML – deploy CML – deploy AI Ukraine’16AI Ukraine’16

K.BokhanK.Bokhan 32


K.BokhanK.Bokhan 33


K.BokhanK.Bokhan 34


K.BokhanK.Bokhan 35

Questions?Questions?

AI Ukraine’16AI Ukraine’16

K.BokhanK.Bokhan Samsung R&D Institute UkraineSamsung R&D Institute Ukraine36

continuous machine learning - ai ukraine · easily deploy and run datacenter-wide app services such...

Documents