ML IN PRODUCTION : AN UBISOFT'S USE CASE

1

Data Platform Group

SPEAKER

• I am Jean-Michel Daignan

​

• Data scientist at Ubisoft Montreal since 2018

​

• Education: Master's degree from Polytech Clermont Ferrand in Physics and engineering

​

• Worked previously at CEA, EDF

​

Details

2

DaignanJM

​

​

jeanmicheldaignan/

​

​

the-odd-dataguy.com/

​

​

​

ORIGIN

LOCATIONS

STAFF

Born in 1986 in France

Installed in 1997 in Montreal

More than 45+ studios in the world

21000 talents worldwide

4000 employees in Montreal

COMPANY

3

UBISOFT LICENSES

4

RAINBOW SIX SIEGE

JUST DANCE

THE DIVISION

FOR HONOR

GHOST RECON

WATCH DOGS

ROCKSMITH

ANNO

BRAWLHALLA

RAYMAN

MARIO + RABBIDS

THE CREW

ASSASSIN’S CREED

IMMORTAL FENYX RISING

ROLLER CHAMPIONS

FAR CRY

Ubisoft : ML reality

Merlin : Ubisoft's ML platform

A journey to MLOps

AGENDA

Takeaways and next steps

Part II

UBISOFT : ML REALITY

6

VIDEO GAMES IS A BIG INDUSTRY

7

56%

VIDEO GAMES

$175 BILLION

+23% vs. 2019

Entertainment estimated revenue in 2020

$299Bn

*Unprecedented Results Due To Covid-19 **Subscription Video On Demand

Source: Newzoo, Ubisoft CMK estimates, NPD, GFK, GSD, App Annie

INDUSTRY IS EVOLVING

8

8

56%

“My competitive set is much bigger than my direct competitors in Sony and Microsoft. I compete for time”

Reggie Fils-Aime

Former president of Nintendo America

INDUSTRY IS EVOLVING

9

9

56%

UBISOFT AND MACHINE LEARNING

10

ML practitioners at Ubisoft

+200*

UBISOFT AND MACHINE LEARNING

11

PROCEDURAL GENERATION OF CONTENT

​

• Player Analytic France in Paris

​

• Rework the conception of flow of levels based on bots interaction and human feedback on the generated level

​

• Collaboration cross teams ML and game developers

​

​

Details

12

DEEP RL FOR GAME DEVELOPMENT

• La Forge MTL

​

• Explore development of bots for map exploration (with ability to jump)

​

• Build smarter non playable character to make the experience more fun

​

Details

13

Ubisoft R&D blog

ZOOBUILDER

• Ubisoft China AI lab (La Forge China)

​

• Ability to make motion capture of wild animal based only on image

​

• Leverage game engine to create training data

​

​

Details

14

ROAD TO PRODUCTION

15

ROAD TO PRODUCTION

16

GDC vault

PAST

PRESENT

FUTURE

Multiple initiatives started in studios around the world

Standardization was needed to reach stability

Propagate standardization along the Ubisoft group and ML entities

ROAD TO PRODUCTION

17

MERLIN : UBISOFT'S ML PLATFORM

18

I KNOW !?

19

NVIDIA MERLIN

• Solution developed by NVIDIA

​

• Help for the deployment of deep learning recommender systems

​

• From ETL to ML boosted with GPU

​

​

Details

20

Nvidia

SHOPIFY MERLIN

• Internal ML platform at Shopify

​

• Focus on scalability, fast iterations and flexibility

​

• Cool tech used and very inspiring

​

​

Details

21

Shopify engineering

UBISOFT MERLIN

• Internal ML platform at Ubisoft

​

• Project/team started in July 2018

​

• Inspired by the Uber’s Michelangelo ML platform

​

• Part of the Ubisoft Data Office

​

​

Details

22

NAMING CHOICE

23

NAMING CHOICE

24

FOR WHOM ?

WHAT ?

WHY ?

Ubisoft DS / ML eng and online programmer

A machine learning platform to build, deploy and operate ML applications easily and at scale

To democratize and simplify the use of Machine Learning within Ubisoft

WHAT IS MERLIN ?

25

THE TEAM

26

Data science

Front-end

Back-end

Management

MERLIN’S COMPONENTS

27

Ubisoft data ecosystem

“Feature” store

Experimentation space

Production space

Serving

“FEATURE” STORE

• Ingestion of features in AWS

​

• Ability to set incremental or full synchronization

​

• Shared data storage

​

• Integrate in internal ETL scheduler

Details

Technologies used/available

28

EXPERIMENTATION SPACE

• Build on top of Jupyter

​

• Different setups (Spark cluster or single GPU machine)

​

• Default ML environment (ability to make manual upgrade)

​

• Direct connection to Merlin data storage

Details

Technologies used/available

29

PRODUCTION SPACE

• Place to deploy code

​

• Scheduling mades with airflow

​

• Deployment from a Gitlab repository

​

• Same machine than in exploration

Details

Technologies used/available

30

PRODUCTION SPACE

31

SERVING

• Ability to load predictions to be served (batch mode)

​

• Direct connection to internal online services (Ubiservices)

​

• Integration in SDK (in game)

​

• Standard call

Details

Technologies used/available

32

SERVING

33

ID of the user

ID of the targeted application

IN-GAME PERSONALIZATION

OUT-GAME PERSONALIZATION

​

ML ENHANCED

APPLICATIONS …

34

… POWERED IN

35

MAP

36

TALKS

37

Women Techmakers MTL

MONITORING

FEATURES

ML TRACKING

PORTAL

SDK

38

PORTAL

• Main entry point to the Merlin platform

​

• Interact with experimentation workspace and serving part

​

• Embedded guidebook for onboarding and quick links to documentation

​

​

​

Details

39

PORTAL

40

PORTAL

41

ML TRACKING

• Based on mlflow

​

• Support offline experimentation and online pipeline

​

• Hosting our own instance on AWS

​

• Using mostly the tracking API

​

​

Details

Technologies used

42

ML TRACKING

43

MONITORING

• Pipeline in parallel of production pipeline

​

• Check the performance of the system

​

• Measure drifting of the metrics (alert if issue)

​

• Build report with Datapane

​

​

Details

Technologies used

44

MONITORING

45

MERLIN SDK

• Wrapper of useful functions for our users

​

• Currently focus on alerting, interaction with Ubisoft online services

​

• Planning to make it collaborative (inner source) for Merlin’s user

​

​

​

Details

46

TAKEAWAYS AND NEXT STEPS

47

LIVE EXPERIMENT

TAKEAWAYS OF ML OPERATIONS (DS POV)

START SIMPLE

MODULARITY

ML REEVALUATION

48

MODULARITY

• All the pipeline is not mandatory

​

• Some teams are only using the serving …

​

• … other only the training …

​

• … but most are using the full pipeline

​

​

​

Details

49

START SIMPLE

• Build simple first pipeline (with no ML in it)

​

• Iterate with available tools

​

• Don’t be blind by papers and tech blog of tech companies

​

​

​

Details

50

LIVE EXPERIMENT

• Key to make innovations and test in real time situation

​

• Mitigate the risks but need some methodology

​

• Find the right balance (we have not this kind of numbers)

​

​

​

Details

51

ML REEVALUATION

• Starting to have pipeline running since 3 years

​

• A game is going through multiple phases (launch, live phase and back catalog)

​

• Still have to continue to monitor and reevaluate the ML pipeline regularly

​

​

​

Details

52

FEATURE STORE

PRODUCTION 2.0

NEXT STEPS

LIVE PREDICTION

53

FEATURE STORE

• Currently not a real feature store

​

• Give more control on the data used for ML (point query in time, feature monitoring)

​

• Build PoC with AWS feature store

​

​

​

​

​

Details

Technologies tested

54

FEATURE STORE

55

LIVE PREDICTION

• Batch mode has his limitations

​

​

• Live prediction is opening new horizons (with some constraints #latency)

​

• PoC up and running but still work to do

​

​

​

​

​

Details

Technologies tested

56

LIVE PREDICTION

57

Huyenchip.com

PRODUCTION SPACE 2.0

• Handle infra to do CPU, GPU and cluster jobs (the same way)

​

• Easier environment deployment

​

• Rework Gitlab/project management

​

• ML observability

​

​

​

​

Details

Technologies tested

58

Part III

A JOURNEY TO MLOPS

59

MLOPS LANDSCAPE IS A MESS

60

mihaileric.com

MLOPS LANDSCAPE IS A MESS

61

REVIEWED TOOLS

62

METAFLOW

• Born at Netflix

​

• Toolbox to operate data science projects locally or remotely without effort

​

• Leverage decorator to define execution and environment (step by step)

​

• Ability to log and track artifacts

​

​

​

​

​

Details

63

Metaflow.org

The-odd-dataguy.com

AWS SAGEMAKER

• ML tooling built on top of AWS

​

• Handle the ML pipeline end to end (from experimentation to production) easily

​

• Everything is working great but be ready to pay the sagemaker tax

​

• Equivalent at Google (Vertex) and Microsoft (AzureML)

​

​

​

​

​

Details

64

AWS

The-odd-dataguy.com

DVC

• Built by Iterative 😉😏😜

​

• Define as a data versioning but it’s more than that

​

• Ability to build DAG to orchestrate ML task (using and producing data)

​

• Total integration with evolving iterative ecosystem

​

​

​

​

​

Details

65

DVC

The-odd-dataguy.com

SOME CONTENDERS

66

TESTING PROCESS

67

Stanford CS 329S

ljvmiranda921.github.io

SOME INSPIRATION

68

Arxiv

GitHub

CONCLUSION

69

ML IN PRODUCTION AT UBISOFT

THIS IS A REALITY

STILL WORK TO DO

Plenty of Ubisoft applications are running with this tech stack and are bringing value in our experinece

To support the complete ML ecosystem at Ubisoft, need to collect the needs and adapt the components

70

ML IN PRODUCTION (QUESTIONS)

71

BE AWARE OF YOUR ENVIRONMENT

START (AND FAIL !?) FAST

FOLLOW YOUR COST

Reuse a maximum of your data stack to start and iterate on that

Using vendor lock is not bad to begin

Help to take decision

ML IN PRODUCTION (MY PERSPECTIVE)

​

72

UBISOFT NEEDS YOU

73

Ubisoft careers

UDO-ML-Recruitment@ubisoft.com

THANK YOU!

DaignanJM

​

​

jeanmicheldaignan/

​

​

the-odd-dataguy.com/

​

​

​

MAP

75

COMPANY

76

Source : NEW ZOO, CMK ESTIMATES, NPD, GFK, GSD, APP ANNIE

• Founded in 1986
• 20,000 employees
• 45+ studios around the world
• Based in 30 countries
• More than 85% of our teams dedicated to production
• An international network with more than 90 nationalities and 65 spoken languages 
• Leading publisher in terms of unit sales on all platforms combined in 2020
• Leading third-party publisher on Switch
• Ubisoft ranked second-leading publisher on new generation consoles with Assassin’s Creed Valhalla, Watch Dogs: Legion, and Immortals Fenyx Rising among the best sellers.
• Assassin’s Creed Valhalla: Biggest Assassin’s Creed game launch in history
• Tom Clancy’s Rainbow Six Siege: 80M players
• Just Dance 2020: Best-selling Just Dance title of the past 6 years

​

TESTING PROCESS

77

TESTING PROCESS

78

COMPANY

79

Source : NEW ZOO, CMK ESTIMATES, NPD, GFK, GSD, APP ANNIE

Founded in 1986

20 000 employees

45+ studios around the world

Based in 30 countries

More than 85% of our teams dedicated to productions

Leading publisher in terms of unit sales on all platforms combined in 2020

Tom Clancy’s Rainbow Six Siege: 80M players

MLOPS LANDSCAPE IS A MESS

80

THE VISIBLE TIP OF THE MACHINE LEARNING PROJECTS

81

Procedural generation of content

Reinforcement learning for game development

Zoobuilder

VIDEO GAMES IS A BIG INDUSTRY

82

Source : NEW ZOO, CMK ESTIMATES, NPD, GFK, GSD, APP ANNIE

SVOD**

$56 BILLION

THEATER*

$6,8 BILLION

LIVE MUSIC*

$8 BILLION

RECORDED

MUSIC

$23 BILLION

PHYSICAL HOME VIDEO

$30 BILLION

56%

*UNPRECEDENTED RESULTS DUE TO COVID-19