Machine Learning & IBM i - Common France
←
→
Transcription du contenu de la page
Si votre navigateur ne rend pas la page correctement, lisez s'il vous plaît le contenu de la page ci-dessous
IBM Systems Center Montpellier
Machine Learning & IBM i
• Les Solutions IBM de Machine Learning
• Démarrer sur IBM i avec des solutions
Open Source et Scikit-learn
Benoit MAROLLEAU - Cloud Architect
IBM Cognitive Systems - Client Center Montpellier, France
benoit.marolleau@fr.ibm.com
Replay & More info
https://ibm.biz/bma-wiki
https://callforcode.org
Introduction & cas d’usage
Optimisation de la logisitique, prediction de récolte, prediction de fraude, customer churn
(attrition) , maintenance predictive, etc.
Avant de démarrer sur Scikit–learn : Introduction au Machine Learning
Solutions IBM pour l’intelligence artificielle
Agenda Solutions Logicielles et Materielles, Private/Public Cloud : Watson , Linux & IC922/AC922
Solutions IBM i / AIX au plus près des données
Démonstration – Scikit-learn sur IBM i
Installation dans un IFS container dans PASE, premiers contacts sur un cas concret
Questions / Réponse , Pour aller plus loin…
Détails des Solutions, Quelques liens utiles, TP à distance disponibles, contactez moi!!
Machine Learning Introduction Cas d’usage
Le Machine learning est partout …
et influence notre vie de tous les jours!
Netflix provides personalized
movie recommendations
Waze provides a personalized
driving experience for its users
Le Machine learning est partout …
et influence notre vie de tous les jours!
‘Epidemics Progression Index (EPI)’
https://towardsdatascience.com/italian-covid-19-analysis-with-python-1bdb0e64d5ac
WW Dataset Johns Hopkins University
https://github.com/resCovid19/COVID-19
Le Machine learning est partout … et influence notre vie de tous les jours!
Machine Learning ? • Machine Learning : moyen d’automatisation en apprenant des données, sans programmer explicitement les règles. • établit des relations de causalité dans les données, alors qu’on observe un phénomène (complexe?) récurrent • L’expertise métier est nécessaire en phase d’apprentissage pour sélectionner les données, les exemples, façonner l’algorithme, le modèle. Elle est nécessaire également pour surveiller et maintenir ces modèles. • Le ML n’est pas nouveau (1950’s), mais la puissance de calcul, les librairies et Framework nécessaires sont maintenant disponibles et accessibles au plus grand nombre. • Deep Learning (DL) : classe particulière d’algorithmes Machine Learning faisant appel à des algorithmes à base de réseau de neurones – très efficaces sur des phénomènes « complexes » • ML fait partie de l’AI, domaine à la croisée de la robotique, la philosophie, la sociologie, les mathématiques…
Machine Learning?
Pourquoi s’y intéresser?
à Je peux dès maintenant améliorer une application avec du Machine Learning !!
Aide à la décision
Détection d’un évènement à base d’images, de données structurées,
Classification, segmentation des clients : Retail, Assurance, Banque,
Surveillance video, Maintenance…
User Experience (UX)
User Interface par l’image, la voix, le texte, helpdesk, chabot (NLP),
Traitement plus rapide des demandes, amélioration continue de services
et produits
Augmentation (aide) et automatisation
Taches répétitives, ex: sur les images, ou des documents
Un médecin passerait des centaines d’heures à lire les publications utiles
pour sa prise de décision: l’IA peut être utile pour l’aider.
Exemple – Optimisation du Turnover Avoir le bon produit dans le bon point de vente. Optimisation logistique, des stocks. Optimisation de la Supply Chain basée sur les habitudes et données des clients (CRM) Solution: IBM i + Machine Learning with Driverless AI
Exemple – Prediction de récolte Estimation de récolte d’une année à l’autre, d’un trimestre à l’autre etc. Résultat en tonnes (modèle de regression) par parcelle et type. Basé sur des données historisées : Données de capteurs, relevés par les ingénieurs et personnes sur le terrain. Solution: IBM i + Machine Learning avec H2O Driverless AI Projet initié suite à une session de type ‘Design Thinking’ , reflexions métier, apport de la technologie …
Exemple: Maintenance Prévisionnelle
Historical
Data
Live
Classification:
Data Class 0 : asset will fail within the next 15 cycles
Class 1 : asset won‘t fail within the next 15 cycles
Regression (prediction):
Schedule Maintenance Asset Time to failure = 18 cycles
Predictive Maintenance 11Avant de démarrer sur Scikit-learn: Machine Learning
Machine Learning ?
L’objectif de l’entrainement d’un algorithme Machine Learning
§ n = nombre d’exemples dans le dataset d’entrainement, yi est la valeur réelle de colonne à prédire (label)
§ Entrainement: on cherche à minimiser l’écart entre le résultat obtenu par l’algorithme, f(x), et la réalité y
§ Les paramètres de l’algorithme vont être ajustés durant l’apprentissage en fonction de l’erreur.
§ Le volume et la qualité des données : Training Dataset conséquent, connaissance du phénomène, du métier, afin de sélectionner
(ou créer) les bons features (colonnes)
§ Choix de l’algorithme en fonction du problème est clé, l’initialisation des hyperparamètres également.
§ A la fin de l’entrainement, le modèle est façonné, les paramètres optimaux sont figés.
§ L’exécution du modèle entrainé, validé, testé, avec un nouvel exemple jamais vu avant, s’appelle l’inférence
§ Des accélérateurs - carte graphiques GPU (ou FPGA) – sont utiles en Training, moins souvent en Inférence : Milliers de cores,
spécialisés dans la multiplication matricielle (calcul tensoriel, vectoriel) peuvent diviser le temps d’entrainement par 50
§ Des outils logiciels, des librairies , existent afin d’assister dans ce travail.Machine Learning ?
L’objectif de l’entrainement d’un algorithme Machine Learning
Error Function Regression
Minimisation à A chaque iteration, en
(presque) chaque cherchant à minimiser la
itération fonction de cout, on se
raproche d’un modele
pertinent
§ Dans la réalité: plusieurs centaines/milliers de dimensions
§ Dimensionality Reduction : Dimensions redondantes ou inutiles
Classification
§ Feature Extraction: Création de features supplémentaires qui aident à mieux
comprendre le phénomène. Un reseau de neurones fait cette extraction « tout
seul » , grande différence avec du machine learning.
§ Estimation de la qualité du modèle : précision du modèle sur des jeux de tests .
est il généralisable sur de nouvelles données ? overfitting vs. underfitting , biais
& variance. Méthodes de Régularisation…Machine Learning ?
L’objectif de l’entrainement d’un algorithme Machine Learning
Deep Learning = Training Artificial Neural Networks
Example solving a non linear classification problem
boundary formation
with 3 layers and ‘relu’
activation
Neural Nets Simulator:
http://playground.tensorflow.org
1516 Machine Learning Durée d’un projet ? De quelques semaines à (en general) quelques mois… Par exemple: Analyse métier et travail sur les données (60%) , modélisation (20%), déploiement et integration (20%)
Solutions
Hardware / Software
Machine Learning sur IBM i
AutoML accéléré (GPU)
sur Linux/AC922?
Machine Learning
Dev
Business Analist ?
Data Scientist
?
Le choix des outils Data Engineer
Operations ?
?
Data Science libraries & frameworks
• ML: scikit learn, R, Rstudio, SQL, H2O-3 , H2O Driverless AI …
• DL: PyTorch, TensorFlow, Caffe , Theano, Chainer …
• Distributed : Apache Spark, H2O Sparking Water, HPC tooling…
• Accelerated ML: Nvidia CUDA, Rapids , IBM Snap ML
à Hundreds to come every month…most of them are free open source
Desktop & Server Appliance with Accelerators (GPU)
Nvidia DGX , Intel … GPU / FPGA based etc.
IBM AC922 : Oak Ridge ‘Summit’ SuperComputer Building BlockMachine Learning Dev
Business Analist
?
?
Public Cloud ? Dans son Datacenter ?
Data Scientist
Data Engineer
?
Operations ?
?
1. La standardisation des offres AI dans le
Public Cloud répond à l’essentiel des
besoins
• Modèles Pré-entrainés prêt à l’usage
• Invocation par API
• Des Solutions Cloud permettent de créer des
modèles cousus main : Watson Studio, etc.
2. Limitations:
• Data Gravity (Transfert des données vs. volume)
• Legal, Compliance & Regulations
• Pay Per Use vs. Investissement , cout perfomance
à AI = Public + Private Cloud (Datacenter)
+ Edge (embarqué)Solutions IBM : Data Science/ML/DL
WML WML
Structured Data All use cases Vision/Acoustic All use cases
Tabular data: IoT/Logs/Database Program your models w/ Open source Retail/Inspection Data Science Cockpit
Data Engineer Acceleration & AutoML with Driverless AI & Accelerated HW . Scale w/o limit /Security, Quality… AutoML
Open Source Libraries
H2O Driverless AI WML-CE WML-A Visual Insights WML Studio
AIX / IBM i
ML on IBM i /AIX Auto Machine learning Free, Easy to install Enterprise Solution for Collaborative ML/DL
Ready to use
No GPU, runs on CPUs Simplified deployment,, High Performance DL Deep Learning – WML dev & deployment
Deep Learning
Description Traditional datasets, Relevant for automatic pipelines, LMS (Large Model) Advanced Features : DDL
with Video tools
AutoAI
Model Inference (production) model "explainability", DDL (Distributed) 4+ Nodes… Use WMLCE if installed
Train/Infer/Edge
Enrichable with Watson Services end to end automation HPC Job Scheduler Can Use WML-A plugin
Inference:
Support Avaialable from IBM H2O L 1-3 Available from IBM IBM L 1-3 Included IBM L1-3 Included Available from IBM
AIX , IBM i, Linux x86, IC/AC922, LC9xx x86/AC922/ IC922 x86/AC922/ IC922 x86/AC922/ IC922 x86/AC922/ IC922
Environments
Server Technologies
PaaS / SaaS IBM Cloud Power Virtual Server BYOL / AC922 IBM Cloud Yes No Trial only Watson Studio
MultiCloud (PAK) On Prems / Cloud Yes Yes CP4D Yes CP4D Yes Yes CP4D
Private CloudData Science : IBM i & AIX
April 2019
Les Applications AIX et IBM i hébergent les données
essentielles de l’entreprise
à Créer encore plus de la valeur facilement avec
les outils Open Source à disposition pour la Data
Science !!H2O Driverless AI Complements IBM Visual Insights
Integration with Cloud Pak for Data - powered by POWER9, OpenPower NVlink & GPU Acceleration
IBM Visual Insights H2O Driverless AI is
delivers Deep Learning for Images Automatic Machine Learning
Transactional
Time Series
Sensors
Log
NLP
22H2O Driverless AI: How it Works
Drag and drop data Automatic Visualization Automatic Machine Learning Automatic Scoring Pipelines
1 2 3 4
Use best practice model recipes Deploy ultra-low latency Python or
Ingest data from Understand the data shape, and the power of high Java Automatic Scoring Pipelines
cloud, big data and outliers, missing values, etc. performance computing to Iterate that include feature
desktop systems across thousands of possible transformations and models.
models including advanced feature
HDFS engineering and parameter tuning
SQL Automatic Machine Learning
Model Recipes: Survival of the Fittest Model
Local
Modelling • i.i.d. data Documentation
• Time-series Deploy Low-
Automatic latency
• More on the way
X Y Scoring Pipeline Scoring to
Amazon S3
Production
Advanced Machine learning
Snowflake Dataset Feature Model
+ Algorithm + Interpretability
Engineering Tuning
Google BigQuery
Powered by GPU Acceleration
Azure Blog Storage
Confidential 23H2O Driverless AI : Fraud Detection Example
• Driverless AI matched 10
years of expert feature “Driverless AI is giving amazing
engineering results in terms of feature and
model performance “
• Increased accuracy from
0.89 to 0.947 (6%) in
Venkatesh Ramanathan
detecting fraudulent activity Senior Data Scientist, PayPal
• 6X speed up when using
H2O4GPU with Driverless AI Leader in Gartner’s 2018 Data
Science Quadrant
H2O Driverless AI and IBM POWER9 GPU Systems are bringing together the best of breed AI innovation. To handle
the increasingly complex workloads of AI you need an integrated system of software and hardware:
• supports nearly 2.6x more RAIBM POWER9 M, 9.5x more I/O bandwidth than comparable systems.
• Nearly 2X the data ingest speed and over 50% faster feature engineering.
• With GPU accelerated machine learning delivering nearly 30X speedup on model building.
• Support for up to 6 V100 GPUs on a single system.AIX/IBM i – H2O Driverless AI Solution
ETL customer Data
IBM Power Systems via JDBC Connector H2O Fast Start Bundle
Enterprise Server (AC922+H2O DAI)
Feature
Engineering & AI
Model Training
LPAR AI
LINUX Scoring
(inference engine) AC922
ML Inference Engine
Java MOJO can be deployed in IBM i , AIX and Linux LPAR 25
CONFIDENTIALIllustration Rapide
Customer Churn ou “Prediction de l’attrition”
La prédiction de l’attrition (« churn » en anglais) est l’un des cas d’usages les plus fréquents
de la data-science. Elle conduit souvent à l’instauration d’un score prédictif qui sert
déclencher des actions marketings et commerciales.
La disponibilité de données historiques de clients, d’une part, et les coûts de rétention
souvent plus faibles que les coûts d’acquisition, font de la prédiction de l’attrition l’un des
premiers travaux d’une équipe de data-scientists.
Source : Converteo.com
CONFIDENTIALSmarter IBM i apps made easy with Driverless AI
CRM & Customer Churn scenario
Initial CRM
No Customer Churn risk estimate per customerSmarter IBM i apps made easy with Driverless AI
Dataset creation from Db2 for i
Prepared
Training
Dataset
Extraction
or JDBC
Data ConnectorSmarter IBM i apps made easy with Driverless AI
CRM & Customer Churn scenario
2
Import & Visualize Data
Model Export
Create & Test Models
3
1
Recommendation Engine
Extract Scoring Pipeline
CRM data Optimized for GPU-Accelerated
(or JDBC) Power9 Servers 4
Augment IBM i Applications
Monitor Models Model Inference
5
& Iterate
Business Database Business Application 29Export Scoring Pipeline 3
Automatic Scoring Pipelines
Export ultra-low latency Python or
Java Automatic Scoring Pipelines
that include feature
transformations and models.
Here, the Java/ Python Scoring
Pipeline “scorer.zip” to be deployed
on the inference system
Inference on the Edge , datacenter,
Cloud … on any device.
Scoring Pipeline exportSmarter IBM i apps made easy with Driverless AI
Behind the scenes:
CRM application prototype
integrated with a real time
scoring Pipeline (Java MOJO)
Fig. Node-RED Prototyping - CRM dashboard
Java MOJO
Java MOJO Scoring Pipeline Db2 CRM
Invoked from the Node.js app DatabaseManual Modeling vs. Driverless AI Accelerated Auto ML
§ Customer Churn Dataset WA_Fn-UseC_-Telco-Customer-Churn.csv
Customer churn is when an existing customer, user, player, subscriber or any kind of return client stops doing
business or ends the relationship with a company.
§ Model Accuracy: 0.79 vs. 0.84
§ Time to market :
Days vs. Minutes with DAI
AutoML w/ Feature Engineering, etc.
§ Accelerated ML able to play with larger datasets
§ Accelerated ML saves time & effort
& licensing costs (PowerVM cores offload to GPU)
à Accelerated Auto ML can create high quality models, that only a team of
experienced data scientists with years of business knowledge can build.
(Real feedback from customers)Machine Learning
sur IBM i
Stratégie Open Source sur IBM iStratégie Open Source sur IBM i à Integration & Web • API Management , Web Services, • Web Development Open Source & IBM i : Unix (AIX) Runtime aka PASE • Accès Database à Cloud / Automatisation • Openstack, Cloudinit, Terraform , Ansible à Data Science , Data Analysis
Data Science : IBM i & AIX
Librairies et langages disponibles sur AIX & IBM i
§ Packages RPM “Data Analytics & Data Science”
§ Complètent les fonctions Db2
§ Liste qui s’enrichit régulièrement J Yum install
Numpy, Pandas : Data Processing
Scipy, Scikit-Learn
IPython : interactive Python
NLTK : Natural Language Processing
Matplotlib, Jupyter : Data Visualization
R Language (Interpreter, Runtime) ILE / PASE Applications
Machine Learning DB2 for i
§ Utile dans chaque phase d’un projet ML : Libraries (PASE) Business Data
à Préparation & Visualisation, Modeling, Validation…
§ Particulièrement intéressant:
à L’inférence sur IBM i intégrée à une application IBM i (RPG, Db2, Node.js)
§ L’accélération matérielle (GPU) parfois nécessaire:
è ‘large datasets’ , fort parallèlisme d’éxécution
è Utilisation de librairies ‘Deep Learning’ (neural nets)
è Économie des ressources CPU / Mémoire sur IBM i (price / performance)
è Entrainement des modèles sur Power IC922 / AC922 + WML-CE (Cloud public ou privé)Data Science sur IBM i
à Scikit-learn https://scikit-learn.org
• Outil simple et efficace pour du data mining / data analysis
• Accessible à tous, réutilisable dans divers contextes
• Repose sur les librairies NumPy, SciPy, matplotlib
• Open source depuis 2010 – INRIA
• Utilisable commercialement - BSD license
• Plus souvent utilisé avec panda (manipulation de données - dataframe)
• Cours en Français ? http://cedric.cnam.fr/vertigo/Cours/ml/tpIntroductionScikitLearn.htmlMachine learning sur IBM i
• Packages nécessaires – 1.2 GB
Package (RPM) Description
Tcl TCL language support
Tk TK package for GUI support of TCL
Python3 Python 3 support
Python3-devel Python3 development package
Yum Install python3-ibm_db DBAPI support backage for IBM i
(*ALLOBJ) Pyton3-numpy Numpy package
Python3-scipy Scipy package
Python3-scikit-learn Scikit-learn package
Libzmq Zero-MQ library
Package Python (PIP) Description
Package Python juypterlab Includes jupyter (Julia/Python/R IDE)
(utilisateur)
joblib Model and large objects persistence on FilesystemMachine learning
sur IBM i
• Chroot jails (Isolation ds PASE/IFS ) J
• n Data Scientists = n bacs à sable (chroot)
• Environnements Dev + Prod
• 1.2GB par environnement Scikit-learn
mop_chroots
• Le reste du système ( / ) isolé
yum install
--installroot /QOpenSys/mop_chroots/dev1
tcl
tk
python3 dev1 dev2 . . . devN prod
python3-devel
python3-ibm_db
SSH
python3-numpy
python3-scipy
python3-scikit-learn
ibm_dbi
python3-pyzmq odbc
libzmq
python3-pip ibm_dbi
python3-pandas
opensssl odbc
bash
https://10.7.19.71:8881 Prod
ssh dev1@BENOIT.ICC.LOCAL
bash-4.4$ pip3 install jupyterlab App
bash-4.4$ pip3 install joblib
bash-4.4$ jupyter notebook --port 8881 --ip 10.7.19.71Machine learning sur IBM i
1. Entrainement et validation modèle (IBM i ou pas)
2. Chargement du modèle depuis l’IFS
3. Scoring (Inference) sur IBM i (PASE)
4. Integration de modèle dans l’application
PASE
Program
Call or
REST PASE or ILE
Prod
App
Fig. Scoring 20 clients
dans PASE (Python3)IBM Data Science Technologies & IBM i
à Demonstration
• Customer Churn
• Supervised model – classification with Scikit-Learn
• Install yum packages and git clone
• https://github.com/bmarolleau/firstdemo-scikitlearn-ibmi/Data Science sur IBM i
https://scikit-learn.orgMachine learning sur IBM i
Machine learning sur IBM i
Pour aller plus loin…
CONFIDENTIALQuelques cours et documentations en ligne…
https://cognitiveclass.ai/courses/data-analysis-python
https://www.coursera.org/specializations/data-science-python
CONFIDENTIALQuelques documentations utiles… How to Start With Machine Learning on IBM i (by Gan Zhang, IBM - Nov 2019 ) Accessing the rest of your IBM i from Python ibm_dbi driver presentation Accessing the rest of your IBM i from IBM i / Windows/Linux : ODBC Driver github examples YiPS http://yips.idevcloud.com/wiki/ Bitbucket IBM i Open Source https://bitbucket.org/ibmi/opensource/ My blog J (K8s , App modernization , ML , H2O, Watson, Node-RED, IBM i) https://ibm.biz/bma-wiki Videos 2020 (AIX/IBM i in IBM Cloud , Terraform) My Youtube Channel
Workshops à Distance
Présentation / Démo / Travaux Pratiques Power IBM i à distance
è En Planifié avec Expert IBM ou Travaux Pratiques Autonome.
è Nous fournissons l’accès distant à nos plateformes, l’expertise et
les supports de présentation et de travaux pratiques!! Contacts Systems Center Montpellier :
Benoit Marolleau benoit.marolleau@fr.ibm.com
è Durée: 1 ou 2 heures (Webinar) à 1 semaine (en autonome) Marc Lapierre (Briefing Client) mlapierre@fr.ibm.com
Christophe Cavelier (BP) chris.desclavelles@fr.ibm.com
è Acces via OpenVPN et Web Conférence
è IBM i 7.3/7.4 pré-chargés, AIX 7.x, AC922, Linux, OpenShift…
Sujets déja disponibles:
POW4- Machine Learning, Jupyter & Scikit-learn sur IBM i
POW1 - H2O.ai Driverless AI : Machine Learning Accéléré avec integration IBM i
POW5- Open Source et IBM i : Node-red, node.js, web services.
POW2- Démarrer sur IBM Cloud et Power Virtual Server IBM i
POW7- Open Source & IBM i : Cloud et Automatisation de l'infrastructure avec Terraform
+ Tous les sujets du moment : Cloud, PowerVC, AI Accéléré (Vision, WML-CE) , Db2 Mirror / Haute Dispo
Si besoin: agenda à la carte, contactez nous!
Red Hat
OpenShift
CONFIDENTIALIntroduction & cas d’usage
Optimisation logisitique, prediction de récolte, prediction de fraude, customer churn
(attrition) , maintenance predictive, etc.
Avant de démarrer sur Scikit–learn : Introduction au Machine Learning
Solutions IBM pour l’intelligence artificielle
Agenda Solutions Logicielles et Materielles, Private/Public Cloud : Watson , Linux & IC922/AC922
Solutions IBM i / AIX au plus près des données
Démonstration – Scikit-learn sur IBM i
Installation dans un IFS container dans PASE, premiers contacts sur un cas concret
Questions / Réponse , Pour aller plus loin…
Détails des Solutions, Quelques liens utiles, TP à distance disponibles, contactez moi!!
Merci !Montpellier Cognitive Systems Lab
Manager : Philippe Chonavel
Coordinator : Alain Roy Montpellier Client Center (France) - EMEA
Team PowerAI/WML ATS Europe Our Offerings
Jean-Armand Broyelle 1. Technical Consultancy & Assistance
Regis Cely 2. Co-Creation Lab Workshops
Sebastien Chabrolles 3. Hands-On Technical Enablement
Sylvain Delabarre 4. Demonstration/PoT
Maxime Deloche 5. PoC/Benchmark
Thierry Huche
Benoit Marolleau
Team CAPI/SNAP (FPGA)
Bruno Mesnet
Alexandre Castelane
Fabrice Moyen
PowerHA Analytics , ML/DL
Team HPC Modernization Virtualization
Ludovic Enault Cloud / Hybrid Cloud Performance Benchmarks
Pascal VezolleA vous de jouer !
POW1 - H2O.ai Default Payment / customer churn scenario *** Remote Hands-on Labs option!!
Data Scientist in a Box on Power AC922 / IC922
H2O Driverless AI is an artificial intelligence platform for automatic machine learning (ML).
POW2- Get your hands dirty with IBM Cloud and Power Virtual Server ***webinar
Get started with IBM Cloud & discover the new Power Virtual Server offering with this guided hands-on
lab / demonstration session. Get AIX/IBM i/Linux VM on PowerVM is a few seconds , in a Pay as you Go
model.
POW3 - Business Continuity : Db2 Mirror, Logical Replication, and PowerHA
***webinar/Labs upon request
Positioning and technical aspects, External Storage Solutions, Q&A
CONFIDENTIALA vous de jouer !
POW4- Machine Learning, Jupyter & Scikit-learn on IBM i /AIX ***Remote Hand-on Labs Option
Get started with free open source frameworks & libraries on your favorite platform (AIX/IBM i)
POW5- Open Source on IBM i : node-red, node.js, web services. ***Remote Hand-on Labs Option
Embrace open source technologies on your IBM i platform , and build , prototype, deploy your first Web
Service.
POW6- Open Source on IBM i : Cloud & automation with Ansible ***soon
Use core Ansible modules , IBM i modules & playbooks for : Configuration of servers, Application deployment,
POW7- Open Source on IBM i : Cloud & automation with Terraform ***webinar
Terraform providers for PowerVC private & CAM (Cloud Pak for MCM) , IBM Terraform provider for Public Cloud
+ Visual Insights + AC922 / WML-CE ….
Contact us for Any customized Agenda of course!!
CONFIDENTIALGet Started Today : Contact us!
PowerAI Developer Portal
https://developer.ibm.com/linuxonpower/deep-learning-powerai/technology-previews/powerai-vision/
AI Vision Object Detection Demo
https://www.youtube.com/watch?v=19vaot75JCY & Jupyter notebook
AI Vision / Public Cloud – Get Started demo
https://github.com/IBM/powerai-vision-object-detection
PowerAI FAQ
https://developer.ibm.com/linuxonpower/deep-learning-powerai/faq/
PowerAI Vision 1.1.1 Free trial
Register for a free 3-day trial of PowerAI Vision
H2o.ai Driverless AI (Trial 21 days)
https://www.h2o.ai/products/h2o-driverless-ai/
Presentations, Demo Replays : https://ibm.biz/bma-wiki
Want to know more? Need support ?
Montpellier team & AI Environment for your PoC / Tests: Driverless/PowerAI/AI Vision Remote Access
Contact us : a2roy@fr.ibm.com / benoit.marolleau@fr.ibm.com
56Get Started Today – H20 Driverless AI
Videos and Solution Brief H2O Driverless AI Tutorials Client Reference Case Study AC922-H2O DAI bundle
Want to know more? Need support ? Contact us :
- European Cognitive Systems CoC - IBM Montpellier a2roy@fr.ibm.com / benoit.marolleau@fr.ibm.com
Power Systems & AI Environment for your PoC / Tests: Driverless/PowerAI/AI Vision Remote Access.
- IBM CSSC Workshop? AI Workshop at CSSC
WML & AI Demos on the IBM WW Demo portal : https://www.ibm.com/systems/clientcenterdemonstrations
Presentations, Demo Replays for the author : https://ibm.biz/bma-wiki
WML-CE FAQ https://developer.ibm.com/linuxonpower/deep-learning-powerai/faq/
H2o.ai Driverless AI (Trial 21 days). https://www.h2o.ai/products/h2o-driverless-ai/Gartner names H2O as Leader with the most completeness of vision • H2O.ai recognized as a technology leader with most completeness of vision • H2O.ai was recognized for the mindshare, partner network and status as a quasi-industry standard for machine learning and AI. • H2O customers gave the highest overall score among all the vendors for sales relationship and account management, customer support (onboarding, troubleshooting, etc.) and overall service and support.
Shortage of Data Scientists
“The United States alone faces a shortage of 140,000 to
190,000 people with analytical expertise and 1.5 million
managers and analysts”
–McKinsey Prediction for 2018H2O.ai: Worldwide Community Adoption * DATA FROM GOOGLE ANALYTICS EMBEDDED IN THE END USER PRODUCT
Use cases examples: AI in Financial Services
Wholesale / Commercial Banking IT Infrastructure
• Know Your Customers (KYC) • Security Cyberlake
• Anti-Money Laundering (AML) • DoS Detection and Protection
• Master Data Management
Card / Payments Business Retail Banking
• Transaction Frauds • Deposit Fraud
• Collusion Fraud • Customer Churn Prediction
• Real-Time Targeting • Auto-Loan
• Credit Risk Scoring
• In-Context PromotionH2O Driverless AI? Containerized App
IBM Cloud Private
or OpenShift CP
n Worker nodes
GPU GPU GPU … GPU
Cuda Drivers + WML-CE
IC922/ AC922
On Prems/Cloud
62Example: Predictive Maintenance 2
with Driverless AI
Automatic Visualization
Understand the data shape,
outliers, missing values, etc.
Data Visualization on Driverless AIExample: Predictive Maintenance 2
with Driverless AI
Create new model
Automatic Machine Learning
Use best practice model
recipes and the power of high
performance computing to
Iterate across thousands of
possible models including
advanced feature engineering
and parameter tuning
Target Column : Asset Failure in 15 cycles (YES/NO) for binary classification
Or estimate the Time to Live (in number of cycles) with a regressionExample: Predictive Maintenance
with Driverless AI
Create new model
Automatic Machine Learning
• POWER9 supports nearly 2.6x more RAM, 9.5x
more I/O bandwidth than comparable systems.
• Nearly 2X the data ingest speed and over 50%
faster feature engineering.
• With GPU accelerated machine learning delivering
nearly 30X speedup on model building.
• Support for up to 6 V100 GPUs on a single system
Experiment running –GPU Accelerated AutoML on POWER9Example: Predictive Maintenance with Driverless AI Turbine sensors dataset. RUL is the column to predict. Sensors data imported. Contains 21 sensor data for each turbine RUL: Remaining Useful Life 100 turbines, 200+ cycles per turbine, before failure. (or Time to Live)
Example: Predictive Maintenance
with Driverless AI
Reduced
Training on Training
19K rows time
Target: RUL
Precision
Metric
RMSE=35.7
(Root Mean
Square
Error)
Feature Importance in the model decision: 1. Cycle# 2. Sensor11, 3. Sensor4 etc.Java Mojo Scoring Pipeline on IBM i/AIX/Linux
Real time prediction on IBM i !! (AIX/Linux)
with the java MOJO Scoring Pipeline built by Driverless AI
Real time Customer Churn predictions on IBM iDAI Data Connectors: JDBC Connection to DB2 for i
How to configure a Db2 Data Connector:
1. Download a Db2 JDBC Driver first
- Ex: Db2 for i type 4 jdbc driver jt400.jar : https://sourceforge.net/projects/jt400/
- Db2 LUW : get the appropriate db2jcc4.jar
2. Get the original Driverless AI Config file aka /etc/dai/config.toml
Many ways to do that, either from the image or a running container (docker cp etc.)
- K8s example : kubectl cp -c dai-gpu-mop dai-demo-mop-6b75f6b5b9-kddd:/etc/dai .
3. Edit this config.toml as shown below (refer to the Online doc for further information)
4. Inject config.toml, and the driver jar file(s) , to be mounted in your DAI POD/Docker container.
- Inject it in your DAI image or better, use a K8s ConfigMap / persistent volume to persist it
5. Restart your DAI container to refresh the config.
/etc/dai/config.toml
# jdbc: JDBC Connector, remember to configure JDBC below. (jdbc_app_configs)
enabled_file_systems = "upload, file, hdfs, s3, jdbc "
jdbc_app_configs = '{"db2fori": {"url": "jdbc:as400://;", "jarpath":
"/etc/dai/jt400.jar", "classpath": "com.ibm.as400.access.AS400JDBCDriver"}}’
# Note: here, the jtopen driver jtopen.jar is mounted in the DAI Container, in /etc/daiDAI Data Connectors: JDBC Connection to DB2 for i
Data Connectors - JDBCDAI Data Connectors: JDBC Connection to DB2 for i
Configure your Data source:
Db2 Connection & Query with credentials
to your Db2 database.Example: Predictive Maintenance
Architecture Overview
Asset Management + IoT Solution for Predictive Maintenance
Watson IoT
Platform Asset Management
IoT MQTT Broker
On Premise or IoT Platforms on
IBM Cloud
Model API Call
Integration – Data Transformation
WebSphere IIB
Work Order
Scoring - Decision
Dashboad - Reports SPSS Modeler
Cognos BI SPSS C&DS
Accelerated ML/DL
- IBM Watson Studio
and/or IBM Watson Machine Learning (WML)
and/or H2O Driverless AI
- OpenPOWER – AC922 WML-CE & GPU Accelerators
07/05/2020
72Example: Predictive Maintenance
with Watson Studio
Model created with AutoAI on Studio. Can be deployed on WML (REST API…)Computer Vision with PowerAI Vision & IBM i
[understands natural-
language and responds in
human-like conversation]
CONVERSATION
Camera
[ Web Browser ]
Object Detection API
Node-RED Deployed
Model APIs FB Detectron
Faster RCNN
Yolo V2
DEVICE
AC922Computer Vision with PowerAI Vision & IBM i
Computer Vision with PowerAI Vision & IBM i IBM i can now detection objects using an object detection API on premise powered by PowerAI Vision On ppc64le with Nvidia GPUs for inference.
Computer Vision with PowerAI Vision & IBM i
Accelerated embedded edge devices Train on PowerAI Vision but infer on embedded devices
Made with PowerAI Vision – Drone Inspection
Acceleration in training …. days become hours
Performance… Large AI Models Train
faster training times
~4 Times Faster
Faster Training
and Inferencing
POWER9 Servers with NVLink to GPUs
vs for data scientists
x86 Servers with PCIe to GPUs
Distributed Deep Learning Traditional Model Support à Large Model Support
(Competitors) (PowerAI)
Limited memory on GPU forces Use system memory and GPU
trade-off in model size / data to support more complex models
resolution and higher resolution data
DDR4 CPU DDR4 POWER
CPU
POWER NVLink
NVLink
PCIe
Data Pipe
Graphics
Memory GPU Graphics
GPU
Memory• IBM POWER SYSTEMS
AC922
An Acceleration Superhighway
Unleash state of the art IO and accelerated
computing potential in
the post “CPU-only” era
Designed for the AI Era
Architected for the modern analytics
and AI workloads that fuel insights
Delivering Enterprise-Class AI
Flatten the time to AI value curve
by accelerating the journey to build,
train, and infer deep neural networksOpenPower Foundation – After 6 years of existence…
§ Drivers: Innovation vs. Moore law
§ Collaborative Approach
§ Domains:
– Cloud & Scale Out Architecture
– Research/ High Perf Computing
– Analytics, Big Data
– Machine Learning / Deep Learning
§ IBM Sells OpenPower servers
§ Power Systems POWER9…
https://www.ibm.com/cloud/bare-metal-servers/power
https://console.bluemix.net/docs/services/PowerAI-IBM/OpenPower Summit 2018
Source Forbes(Free) PowerAI 1.6.2 (WML-CE) Overview
5765-PAI - Linux Native install (conda) or docker image on RHEL 7.6+ / Ubuntu 18.04
https://hub.docker.com/r/ibmcom/powerai/
Works with CUDA 10 + Nvidia drivers
§ Distributed Deep Learning (DDL) 1.3.0 Not Just About Hardware Design
§ TensorFlow 1.13.1
It’s about co-optimized
§ IBM Caffe 1.0.0
§ Caffe2 1.0.1 software
§ PyTorch 1.0.1
§ Snap ML 1.2.0
+
hardware
§ Rapids cuDF /cuML 0.2.0
Available on x86-64 architecture, Optimized for
§ IBM AC922 POWER9 system with NVIDIA Tesla V100 GPUs
§ IBM S822LC POWER8 system with NVIDIA Tesla P100 GPUs
Announcement LetterWML-CE : Enterprise AI Platform
Simplicity: Integrated Ease of Use, Unique Faster Model Open AI Platform w/
Platform that Just Works Capabilities Training Time Ecosystem Partners
Power9 IBM ISV Solution
CPU
SW SW SIs
PowerAI
GPU
Curate, Test, and Support Large data & model Faster Training Times in Platform that Partners
Fast Moving Open Source support due to NVLink Single Server can build on
Provide Enterprise Acceleration of Analytics Scalability to 100s of Software Partners: H2O,
Distribution on RedHat & ML Servers (Cluster level IBM, Anaconda
Easy to deploy Enterprise AutoML: PowerAI Vision Integration)
AI Platform SIs, Solution Vendors
Elastic Training: Scale Leads to Faster Insights & Accelerator Partners
GPUs as Required and Better EconomicsCognitive Systems are built with optimized HW & SW
Not Just About Hardware Design
It’s about co-optimized
Industry Alignment
Dev Ecosystem
software
Partnerships
hardware
+
Open Frameworks
IBM Software which just work for Machine Learning,
Deep Learning and AI
POWER9 is the only processor with NVLink 2.0 from CPU to GPU
Delivering 5.6X Host-Device bandwidth vs Xeon E5-2640 v4
Open Accelerator Interfaces based systems with CUDA H2D Bandwidth Test
No code changes are required to leverage NVLink capability
Accelerator Roadmaps “We’re excited to see accelerating progress as the Oak Ridge National
Laboratory Summit supercomputer machine, which we expect will be
among the world’s fastest supercomputers. The advanced capabilities
of the IBM POWER9 CPUs coupled with the NVIDIA Volta GPUs will
significantly advance DOE’s mission critical applications,” says Buddy
Bland, Oak Ridge Leadership Computing Facility DirectorSnap ML
Distributed GPU-Accelerated Machine Learning Library
Snap Machine Learning (ML) Library
APIs for Popular ML
Frameworks
Logistic Regression Linear Regression
Support Vector
More Coming Soon
Machines (SVM)
(coming
libGLM (C++ / CUDA Distributed Hyper-Parameter soon)
Optimized Primitive Lib) Optimization
Distributed Training
pai4sk API (Included in PAI 1.5.4)
snap-ml-local API
4 APIs snap-ml-mpi API
snap-ml-spark API
87• Snap ML: Training Time Goes From Logistic Regression in Snap ML (with
An Hour to Minutes GPUs) vs TensorFlow (CPU-only)
46x faster than previous 80
record set by Google 1.1 Hours 46x Faster
Workload: Click-through rate
prediction for advertising 60
Runtime (Minutes)
Logistic Regression Classifier in
Snap ML using GPUs vs 40
TensorFlow using CPU-only
20
Dataset: Criteo Terabyte Click Logs 1.53
(http://labs.criteo.com/2013/12/download-terabyte-click-logs/) Minutes
4 billion training examples, 1 million features 0
Model: Logistic Regression: TensorFlow vs Snap ML
Test LogLoss: 0.1293 (Google using Tensorflow), 0.1292 (Snap ML)
Google Snap ML
Platform: 89 CPU-only machines in Google using Tensorflow versus CPU-only Power + GPU
4 AC922 servers (each 2 Power9 CPUs + 4 V100 GPUs) for Snap ML
Google data from this Google blog 90 x86 Servers 4 Power9 Servers
(CPU-only) With GPUs
88Deep Learning = Training Artificial Neural Networks
Based on biological neurons. Artificial neurons learn by recognizing patterns in data.
Neural net Framework
& Python Programming
Each input value is a pixel Each output contains:
RGB Value (3 Channels)
outputs = f(f(f…(x))
detection_boxes
detection_scores
Training = Adjust the parameters (w,b) detection_classes
To get the best validation results num_detections
Vs. ground truth (Training set).
Inference = apply the model to the input (w,b y = f(x)
fixed) to get the output (classification, detection, …)Demonstration
Machine Learning techniques
• Classification: predict class from observations
• - E.g. Spam Email Detection
• Clustering: group observations into “meaningful” groups
• - E.g. Amazon Recommendations
• Regression (prediction): predict value from observations
• - E.g. Energy consumption prediction
and many different technologies and libraries are available:Machine Learning Methods (examples)
K Nearest Bayesian Decision trees Support Vector Machines
Neighbours Classifiers
Reinforcement learning Cluster Analysis Neural Networks /
Deep learningDeep Learning = Training Artificial Neural Networks
Based on biological neurons. Artificial neurons learn by recognizing patterns in data.
92Deep Learning = Training Artificial Neural Networks
Example solving a non linear classification problem
boundary formation with 1 layer and non linear “relu” activation function
mlp = MLPClassifier(hidden_layer_sizes=(1), 1 layer and 2 neurons 1 layer and 3 neurons Decision plane with 30 neurons
max_iter=500000,activation=’identity’,learning_rate_init=0.01)
Rectified Linear Unit (ReLu)
Neural Nets Simulator:
http://playground.tensorflow.org
93Vous pouvez aussi lire
