OpenSAP Sustai1 Week 2 All Slides

Week 2: Enterprise AI in Action – Defect Detection
Unit 1: SAP AI Overview and Introduction

to SAP AI Core
SAP AI overview and introduction to SAP AI Core
Agenda
SAP AI Strategy and Overview

▪ Introduction to AI
▪ Market overview
▪ AI at SAP
– SAP AI Core and SAP AI Launchpad
– SAP AI Business Services
– SAP Conversational AI
– SAP HANA ML
– SAP Data Intelligence
© 2022 SAP SE or an SAP affiliate company. All rights reserved. ǀ PUBLIC 2

What is artificial intelligence (AI)?
Advanced analysis
Support decision making
Any means to make machines solve
Artificial Intelligence tasks like humans do Take action
Many applications
Learn through training
Machine Learning Computers can learn from examples
to solve tasks without being explicitly Statistical model
programmed to do so
Manual feature extraction
Artificial Neural Brain-inspired machine learning algorithms
Networks
Multiple layers
Handle complex data
Deep Subset of ML that uses deep artificial
neural networks Enable large datasets
Learning
Advanced applications
https://www.gartner.com/en/topics/artificial-intelligence
https://www.britannica.com/technology/artificial-intelligence
Market overview: enterprise applications with embedded AI
The addressable market of enterprise applications with 60%

embedded AI is projected to increase to $136B (2024). of all apps
16% (227B)
CAGR
36%
of all apps
(98B)
197 +15.7%
AI Non-Centric 169
95 106 121 143
AI
Apps
AI Centric 12
14 17 20 25 30 +19.4%
Other Apps 191 178 168 162 157 153 CAGR

(No AI)
$Billion USD 2019 2020 2021 2022 2023 2024
Source: IDC Worldwide Semiannual Artificial Intelligence Tracker, 2H19 – Draft

*Gartner MQ for Data Management for Analytics Jan 2019
Artificial intelligence at SAP
SAP is committed to building intelligent and sustainable enterprises

by infusing AI technologies into applications and business scenarios
▪ Embedding AI natively into SAP applications
▪ Exposing AI via SAP Business Technology Platform
▪ Providing AI services to extend SAP systems

SAP AI solution portfolio
AI embedded Custom Applications AI embedded SAP Applications
manage & operate AI content centrally

AI API standardized integration of AI
SAP AI Launchpad & tooling

SAP AI Business Services
SAP
Conversational
Partner SAP Data SAP HANA Document Business Service Data Attribute …
AI Information Entity Ticket Recommen-
Offerings* Intelligence* (APL, PAL)* Extraction Recognition Intelligence dation
SAP AI Core
Complementary AI runtimes training & serving AI scenarios
SAP Business Technology Platform

SAP AI solution portfolio: SAP AI Core Available since
October 12th, 2021


SAP
Conversational
SAP AI Core
Complementary AI runtimes training
training&&serving
servingAI scenarios©
AI scenarios

SAP AI solution portfolio: SAP AI Launchpad Available since
October 12th, 2021


SAP
Conversational
SAP AI Core

SAP AI solution portfolio: SAP AI Business Services


SAP
Conversational
SAP AI Core
© 2022 SAP SE or an SAP affiliate company. All rights reserved. ǀ PUBLIC

https://api.sap.com/package/SAPLeonardoMLBusinessServices/overview
9
SAP AI solution portfolio: SAP Conversational AI


SAP
Conversational
SAP AI Core
© 2022 SAP SE or an SAP affiliate company. All rights reserved. ǀ PUBLIC

https://api.sap.com/package/SAPConversationalAI/overview
10
SAP AI solution portfolio: SAP HANA ML


SAP
Conversational
SAP AI Core

SAP AI solution portfolio: SAP Data Intelligence


SAP
Conversational
SAP AI Core

Thank you.
Contact information:
open@sap.com
Follow all of SAP
www.sap.com/contactsap
© 2022 SAP SE or an SAP affiliate company. All rights reserved.

No part of this publication may be reproduced or transmitted in any form or for any purpose without the express permission of
SAP SE or an SAP affiliate company.
The information contained herein may be changed without prior notice. Some software products marketed by SAP SE and its
distributors contain proprietary software components of other software vendors. National product specifications may vary.
These materials are provided by SAP SE or an SAP affiliate company for informational purposes only, without representation or
warranty of any kind, and SAP or its affiliated companies shall not be liable for errors or omissions with respect to the materials.
The only warranties for SAP or SAP affiliate company products and services are those that are set forth in the express warranty
statements accompanying such products and services, if any. Nothing herein should be construed as constituting an additional
warranty.
In particular, SAP SE or its affiliated companies have no obligation to pursue any course of business outlined in this document or
any related presentation, or to develop or release any functionality mentioned therein. This document, or any related presentation,
and SAP SE’s or its affiliated companies’ strategy and possible future developments, products, and/or platforms, directions, and
functionality are all subject to change and may be changed by SAP SE or its affiliated companies at any time for any reason
without notice. The information in this document is not a commitment, promise, or legal obligation to deliver any material, code, or
functionality. All forward-looking statements are subject to various risks and uncertainties that could cause actual results to differ
materially from expectations. Readers are cautioned not to place undue reliance on these forward-looking statements, and they
should not be relied upon in making purchasing decisions.
SAP and other SAP products and services mentioned herein as well as their respective logos are trademarks or registered
trademarks of SAP SE (or an SAP affiliate company) in Germany and other countries. All other product and service names
mentioned are the trademarks of their respective companies.
See www.sap.com/trademark for additional trademark information and notices.
Unit 2: SAP AI Core Deep-Dive Part 1:
Onboarding, Terminology, AI API
SAP AI Core deep-dive part 1: onboarding, terminology, AI API
Agenda
▪ SAP AI Core key components

▪ Terminology
▪ SAP AI Core and AI Launchpad onboarding
– Demo 1:
▫ How to connect AI Core to AI Launchpad
▪ The standard AI API
– The standard AI API and his clients
– The standard AI API specification

SAP AI Core key components and architecture
▪ SAP AI Core provides an engine for running any

AI workflows and model serving workloads.
▪ AI API provides a standard way of managing the
AI scenario lifecycle on different runtimes.
▪ SAP AI Launchpad manages the SAP AI Core
runtime by means of the standardized AI API.
▪ SAP AI Core is integrated with third party
tools, widely used for developing cloud-native
applications and AI solutions.

SAP AI Core onboarding: service instance
▪ SAP AI Core is provided as a reuse service in

SAP BTP Marketplace
▪ Service instance of AI Core = AI Core main
tenant + default resource group
▪ One AI Core main tenant per BTP subaccount
– Multiple AI Core instances from the same
subaccount will point to the same AI Core
main tenant

SAP AI Core onboarding: service key
▪ After configuring the AI Core service, we get the

service key needed for authentication purposes
and therefore for managing AI Core through the
AI API.
▪ Keep the service key private. Anyone with these
credentials can access all your assets in AI Core.
▪ Instructions for SAP AI Core provisioning
▪ Instructions for SAP AI Launchpad provisioning

SAP AI Launchpad onboarding: service subscription

Demo 1: How to connect SAP AI Core to SAP AI Launchpad

Terminology
Templates
▪ Templates are used to define the training pipelines and
how a model has to be deployed and served at the level
of the main tenant
Executable
▪ An executable is a template that is instantiated for a
purpose, such as training a model or creating a
deployment
Scenario
▪ A scenario is a group of related executables for a use
case within the user’s tenant
Full AI Core glossary

Terminology
Execution
▪ The operation of executing an executable with a given
configuration
Artifact
▪ A reference to data or a file that is produced or
consumed by an execution or deployment which
typically resides on an object store
Configuration
▪ A collection of parameters, artifact references, and
executables that is used to run an execution or a
deployment
Full AI Core glossary

AI operations
AI API
▪ AI API provides a standard way of managing your artifacts and workflows (such as training scripts, data, models, and model servers).
▪ Open API format, HTTP-based RESTful API definition.
Clients
SDK AI Launchpad operations manager Other clients

▪ SAP AI Core SDK provides a ▪ A multitenant software as a service (SaaS) application on ▪ cURL
programmatic interface to SAP Business Technology Platform to manage AI use cases
across multiple instances of AI runtimes (such as SAP AI ▪ Postman
communicate with SAP AI Core.
Core).

AI API specification for ML lifecycle management
▪ Accessing metadata about scenarios and

executables
▪ Starting/updating/stopping executions (training
pipelines) and deployments (model servers)
▪ Managing artifacts (datasets, models)
▪ Getting metrics (by products of executions/models)
▪ AI API Specification

Lessons learned
▪ SAP AI Core integrates with all the most popular

tools for developing cloud-native AI solutions.
▪ SAP AI Core and SAP AI Launchpad can be
provisioned very easily in SAP BTP.
▪ A service key is fundamental to manage AI Core, for
example through the AI Launchpad interface.
▪ A proper terminology is used to name entities in AI
Core.
▪ The standard AI API allows to manage every aspect
of the AI scenario lifecycle.
▪ Several clients for the AI API are available to manage
AI Core.

Thank you.
open@sap.com
Follow all of SAP

warranty.
Unit 3: SAP AI Core Deep-Dive Part 2: MLOps
SAP AI Core deep-dive part 2: MLOps
Agenda
▪ Code development, containerization and execution in SAP AI Core

▪ Templates and CI/CD in SAP AI Core
– Demo 1:
▫ How to create AI API client in Python
▫ How to connect a GitHub repository to SAP AI Core
▫ How to connect a Docker repository to SAP AI Core
▪ The concept of “application” in SAP AI Core
– Demo 2:
▫ How to create an “application”

Machine learning workflow and code development
Identify the business challenge

Automatize the inspection of final products
Identify the most suitable AI approach

Defect detection with deep learning-enabled computer vision
Identify the phases of your ML workflow

Collect the data, train the DL model, and make it available online through an endpoint
Training part
Investigate the model architecture and design the training Serving part
pipeline
Design the web application that will use the trained model and
Write the code that executes all the needed steps that will serve the inference requests
Write the code for the web-serving application

Code containerization and execution
▪ A Docker repository needs to be ▪ SAP AI Core is an abstraction for the
Kubernetes
Docker
connected to SAP AI Core for developing, Kubernetes layer that is used to

shipping, and running applications in orchestrate the Docker container platform.
containers. ▪ Kubernetes assures scalability,
▪ Docker registry secret authorizes SAP AI monitoring, and cross-platform
Core to pull your Docker images from your deployments.
Docker repository.
Container Lifecycle

Code containerization and execution
Training and serving

codes converted into
Docker images
Templates
Templates are used to define the training pipelines and how a model has to be deployed and served
▪ Workflow templates are used to: ▪ Serving templates are used to deploy one or
– train models more trained models on a model server and rely
– ingest data, pre-process and post-process on KServe
data ▪ KServe provides a standard model inference
– execute batch inference pipelines platform on Kubernetes, regardless of the
framework used to develop the model
▪ Workflow templates are based on the Argo
Workflows workflow engine ▪ KServe supports modern serverless inference
workload with autoscaling including Scale to
▪ Workflow templates instruct AI Core about how
Zero on GPU
to execute the Docker images, what inputs to
provide, what outputs to generate, etc. ▪ Serving templates instruct AI Core about the
container to execute, the model to use, etc.

Example of workflow template
Name of the scenario
Name of the training executable
Docker secret
Resource plan
Input artifact: dataset name and location
Output artifact: model name and location
Container name and location
Execution command of the code within the

Docker image
Example of serving template
Name of the scenario
Name of the serving executable
Name of the input artifact

(i.e. the model produced by the training)
Resource plan
Docker secret and replicas
Container name and location

GitHub and templates
Workflow and
serving templates

CI/CD in SAP AI Core
▪ A GitHub repository contains the declarative descriptions (templates) of the K8s infrastructure currently
desired in SAP AI Core.
▪ An automated process, ArgoCD, makes the underlying K8s infrastructure match the described state in the
repository.
▪ To create a new scenario/executable or update the existing ones, you only need to update the files in the
repository, then ArgoCD handles everything else.

Demo 1

More about Continuous Delivery in SAP AI Core
Git repository SAP AI Core Tenant
Defect
Detection defect-det-app
Training template
Serving template
defect-det-scenario Training executable Serving executable
Docker Registry Data Storage
Defect Detection
Image Dataset
Docker image

Demo 2

Lessons learned
▪ In SAP AI Core, workloads and services are

executed in the form of Docker containers thanks to
the underlying Kubernetes infrastructure.
▪ Access to a Docker registry has to be provided to
SAP AI Core.
▪ Declarative configuration files (templates) are used
to instruct AI SAP Core about how to execute the
Docker images, what inputs to provide, what outputs
to generate, etc.
▪ A GitHub repository is needed to host the
configuration files and access has to be granted.
▪ ArgoCD makes the Kubernetes environment match
the described state in the GitHub repository.

Thank you.
open@sap.com
Follow all of SAP

warranty.
Unit 4: SAP AI Core Deep-Dive Part 3:
Storage Access and Resource Groups
SAP AI Core deep-dive part 3: storage access and resource groups
Agenda
▪ Storage options in SAP AI Core

▪ Resource groups and multitenancy
– Demo:
▫ How to create a resource group
▫ How to connect an AWS S3 bucket
▫ Check all the configurations in SAP AI Launchpad
▪ End-to-end ML workflow in SAP AI Core and SAP AI Launchpad

SAP AI Core key components and architecture
▪ SAP AI Core provides an engine for running

any AI workflows and model serving
workloads.
▪ AI API provides a standard way of managing
the AI scenario lifecycle on different runtimes.
▪ SAP AI Launchpad manages the SAP AI
Core runtime by means of the standardized AI
API.
▪ SAP AI Core is integrated with third party Trained model
tools, widely used for developing cloud-native

applications and AI solutions.
Datasets

Storage options in SAP AI Core
The main storage option is a Hyperscaler object store

to connect to SAP AI Core.
Purpose
▪ Hyperscale storage is a storage mechanism with the ability
to massively expand under a controlled and efficient
managed, software-defined framework.
PROs
▪ Cost-effective, robust, and durable, rich tooling.
▪ Data access is handled via the Argo artifacts feature which
copies data in/out of pipeline containers.
Current status
▪ AWS S3 and Azure Blob storage.
▪ Support for SAP HANA Cloud data lake files.
Future roadmap
▪ Adding other hyperscalers (like GCP).
SAP AI Core deep-dive part 3: storage access and resource groups ups
Resource groups and multitenancy
▪ A resource group represents a virtual collection of related resources within the scope of one SAP AI
Core tenant.
▪ SAP AI Core tenants use resource groups to isolate related ML resources and workloads.
▪ Scenarios, executables, and Docker registry secrets are shared across all resource groups.
▪ Runtime entities such as executions, deployments, configurations, and artifacts belong to specific resource
groups and cannot be shared across resource groups.

Examples of resource group mapping
Service Provider Tenant
Service Service Service

Default
consumer 1 consumer 2 consumer 3

Default Dev Test Prod

Default Team 1 Team 2 Team 3

Business Business Business

Default
service 1 service 2 service 3

Demo

SAP AI Core
end-to-end ML workflow

End-to-end ML workflow in SAP AI Core and SAP AI Launchpad
2.1 Docker: Create own Docker image
0.1 SAP BTP: Create AI Core and AI 2.2 Git: Create a training workflow 4.1 Docker: Create own Docker image
Launchpad instances 2.3 AI Core: Choose the scenario 4.2 Git: Create a serving template
0.2 AI Launchpad: Connect the AI Core 2.4 AI Core: Register input dataset as an 4.3 AI Core: Choose the scenario
runtime to AI Launchpad artifact
0.3 4.4 AI Core: Create a configuration
AI Core: create an AI API client 2.5 AI Core: Create a configuration
4.5 AI Core: Trigger the serving execution
0.4 AI Core: Connect a GitHub repo 2.6 AI Core: Trigger the workflow
execution 4.6 AI Launchpad: Check the status of
0.5 AI Core: Connect a Docker repo deployment, check the logs
2.7 AI Launchpad: Check the status of
execution, check the logs
0 1 2 3 4 5 6
Upload the
AI Core and Use
Data in S3 Get the Get
AI Submit for Submit for deployment
and create a trained deployment
Launchpad training deployment URL for
resource model URL
onboarding inferencing
group
AI Core one-time initial configuration ML model training ML model Inferencing

Lessons learned
▪ In SAP AI Core artifacts (datasets, models) are

stored in hyperscaler object stores.
▪ A hyperscaler object store has to be registered in
SAP AI Core.
▪ Recently also the use of SAP HANA Cloud Data
Lake Files has become possible.
▪ Resource groups allow to segregate AI assets within
the scope of an SAP AI Core tenant.
▪ The main steps of a typical end-to-end ML workflow
in SAP AI Core.

Thank you.
open@sap.com
Follow all of SAP

warranty.
Unit 5: Computer Vision and Defect Detection in
TensorFlow
Computer vision and defect detection in TensorFlow
Agenda
▪ Business challenge: LGP automated defect detection

▪ Computer vision and image segmentation
▪ BYOM: image semantic segmentation model for LGP
automated defect detection
– Introduction to convolutional neural networks
– The U-Net architecture
– Designing the training pipeline
– Monitor your model quality and accuracy
– Coding your training pipeline for SAP AI Core
▪ The computer vision package

Business challenge: automated defect detection
Light Guide Plate (LGP)

▪ Optical product which can
transform a linear light source
into a surface light source.
▪ High luminous efficiency, luminance
Sustainability
Goals & Challenges Solution uniformity, long service life, energy
benefits conservation.
▪ Applied in liquid crystal display
screens such as computer monitors,
car navigations, and PADs.
▪ The LGP production lines Deep learning ▪ Reduced waste
must improve quality and techniques have found
▪ Reduced price of end
speed wide application in
product surface defect product
▪ The distribution of the detection and are ▪ Safer work place
LGP points is uneven, suitable for the
and the size, shape, and detection of complex
brightness of the same defects of the LGP
kind of defects are also
different

Trained AI Models Sales Records Maintenance Orders etc.
▪ Defect Detection Model
Procurement Equipment Inventory
▪ Sound-Based Predictive SAP S4/HANA Cloud
Maintenance Model
Monitoring day-to-day ops
Data
Recording energy
▪ Annotated Defect AI Model Serving APIs consumption Maintenance cost
Images AI models scoring on live planning
▪ Annotated Sound sensors data
Anomalies ▪ Defect detection Production and
AI Model Training
▪ Sound-based predictive Sustainability and
Cloud Storage maintenance KPIs monitoring
Placing proactive
SAP AI Core maintenance orders in S/4
& SAP AI Launchpad Smart Factory App SAP Analytics Cloud

Microphones Data Batches
Camera Data Batches
IoT Gateway
x6

Computer vision: purpose and tasks
Computer vision
Computer vision is an interdisciplinary scientific field that deals with how computers can gain high-level
understanding from digital images or videos.
Classification Object Detection Segmentation Generation Restoration
Single object Multiple object

https://www.v7labs.com/blog/what-is-computer-vision
Image segmentation and some of its applications
Industrial inspection
Image segmentation
Self-driving cars
Medical scans
Satellite imagery

What is a CNN?
▪ A convolutional neural network (CNN) is a deep learning algorithm which can take in an input image,
assign importance to various aspects/objects in the image and be able to differentiate one from the other.
▪ The architecture of a CNN is analogous to that of the connectivity pattern of neurons in the human brain and
was inspired by the organization of the visual cortex.

A model architecture for image segmentation: U-Net
▪ U-Net is an image segmentation model Contracting path Expanding path

architecture that has become very
popular (Ronneberger et al. 2015).
▪ The basic idea of the U-Net is to:
– first obtain a lower-dimensional
representation of the image through a
traditional convolutional neural network;
– then upsample that low-dimensional
representation to produce the final
output segmentation map: a binary
mask of 1s and 0s for each pixel, where
1 indicates a defect and 0 indicates
background.
MobileNetV2 pix2pix

Designing the training pipeline
Load surface images (OpenCV) Pre-processing (OpenCV) Build U-Net (TensorFlow) Train the model
Modified U-Net
Normal surface
MobileNet
V2
Defected surface
pix2pix
upsample
block
Mask/ground truth

Focus on the loss function and metrics
▪ The Intersection-Over-Union (IoU) is one of the most ▪ Overlap = number of pixels

commonly used metrics in semantic segmentation. belonging to the defect class shared
▪ This metric ranges from 0 to 1 with 0 signifying no overlap among prediction and ground truth
and 1 signifying perfectly overlapping segmentation. ▪ Union = sum of the number of
▪ We defined a specific and more suitable loss function for pixels belonging to the defect class
our use case: 1- IoU. in the prediction and ground truth
minus the overlap

Prepare the code to run in SAP AI Core
The key point is to prepare some structured code

(a Python class, for example) which implements all
the required steps in the pipeline:
▪ read_dataset: load the images and apply some
pre-processing;
▪ split_dataset: split the dataset;
▪ build_model: build the chosen model;
▪ init_model: initialize and compile the model;
▪ train_model: configure the training and execute it;
▪ save_model: save the model;
▪ get_model: load the model;
▪ model_metrics: define and register the desired
metrics in SAP AI Core.

The computer vision package
▪ The computer vision package helps with the implementation

of computer vision use-cases on top of SAP AI Core.
▪ It extends detectron2, a state-of-the-art library for object
detection and image segmentation.
▪ Our package adds image classification and feature
extraction capabilities.
▪ The package offers training and evaluation methods and
other helpful components, such as a large set of
augmentation functions.
▪ The package can also be used standalone without AI Core,
and AI Core integration can be added to the project later.
– https://pypi.org/project/sap-computer-vision-package/
– https://developers.sap.com/tutorials/cv-package-aicore-train-
object-detection.html

Lessons learned
▪ Advanced computer vision is the best approach to

automate the product inspection in Bagnoli & CO.
▪ CNN is the most popular and effective choice for
implementing an accurate visual automatic
inspection.
▪ SAP AI Core allows to train and exploit any custom
model built by profiting from the tools and knowledge
available in the Open Source community.
▪ Designing the training pipeline and translating it into
code is fundamental to run it in AI Core.
▪ In SAP AI Core it is possible to easily monitor the
performance of your model.

Thank you.
open@sap.com
Follow all of SAP

warranty.
Unit 6: Defect Detection Training with
SAP AI Core SDK
Defect detection training with SAP AI Core SDK
Agenda
▪ Model training in SAP AI Core

▪ Demos:
– How to convert a training Python code to a Docker image
– How to create a training configuration
– How to run and check the training
– How to retrieve the metrics from SAP AI Core




Training part
pipeline
Write the code for the web serving application

0 1 2 3 4 5 6
Upload the Submit for Get the trained
Use
AI Core and data in S3 training (using model Get
Submit for deployment
AI Launchpad and create a customer’s (registered as deployment
deployment URL for
onboarding resource training artifact URL
inferencing
group scripts) automatically)

Convert the training Python code to a Docker image
docker login docker.io -u YOUR_DOCKER_REPO
docker push docker.io/YOUR_DOCKER_REPO/image-seg-train:0.0.1
docker buildx build -o type=docker --platform=linux/amd64 -t YOUR_DOCKER_REPO/image-seg-train:0.0.1 .
Training Dockerfile requirements.txt train.py

Focus on the training Dockerfile
Load an official Docker image with

Python and all the TensorFlow
dependencies for GPUs
Add path pointing to the NVIDIA libraries

for using GPUs
Install all the packages specified in the

requirements.txt file
For retrieving the pix2pix upsample block
Create all the directories needed to host:

the code to run, the output model, the
dataset copied from the AWS S3 bucket
It copies the train.py from the local folder

to the previously created path

Demo: How to convert a training Python code to a Docker image

0 1 2 3 4 5 6
Use
deployment URL for
inferencing

Demo: How to create a training configuration and how to run and check the training

Focus on the loss function and metrics
▪ The Intersection-Over-Union (IoU) is one of the most ▪ Overlap = number of pixel

commonly used metric in semantic segmentation. belonging to the defect class shared
▪ This metric ranges from 0 to 1 with 0 signifying no overlap among prediction and ground truth
and 1 signifying perfectly overlapping segmentation. ▪ Union = sum of the number of pixel
▪ We defined a specific and more suitable loss function for belonging to the defect class in the
our use case: 1- IoU. prediction and ground truth minus
the overlap

Demo: How to retrieve the metrics from AI Core

Lessons learned
▪ The code prepared for implementing the training pipeline has to be

converted into one or more Docker images that are then uploaded
to the remote Docker registry.
▪ A workflow template has to be prepared and uploaded to the GitHub
repository.
▪ The workflow template is written by following the Argo Workflow
engine specifications and reflects the pipeline structure.
– Metaflow plugin for generating multi-step Argo templates:
▫ Metaflow library for SAP AI Core
▫ Train your model in SAP AI Core using the Metaflow-Argo plugin
▪ The training pipeline can be easily run and monitored by means of
the available endpoints.
▪ The model accuracy and quality can be monitored by defining/
retrieving the metrics in/from AI Core by means of specific
endpoints.

Thank you.
open@sap.com
Follow all of SAP

warranty.
Unit 7: Defect Detection Serving with
SAP AI Core SDK
Defect detection serving with SAP AI Core SDK
Agenda
▪ Model deploying in SAP AI Core

▪ Demos:
– How to convert a serving Python code to a Docker image
– How to create a serving configuration
– How to deploy a model in SAP AI Core
– How to make an inference with the deployed model




Training part
pipeline
Write the code for the web serving application

Serving application workflow
Custom consuming SAP AI Core

service
Serving application
API call Prepare the data Load the DL model

▪ Image is provided in the ▪ Decode binary string ▪ Download the model from the
form of a binary string in ▪ Apply pre-processing, resize object store
the body of the call and normalize the image
Prepare the response Make an inference

▪ Convert with the mask to a ▪ Produce a segmentation map
Response binary string from the prediction
▪ Return the defected area ▪ Calculate the defected area
together with the binary string
© 2022 SAP SE or an SAP affiliate company. All rights reserved. ǀ PUBLIC https://flask.palletsprojects.com/en/2.1.x/ 4
0 1 2 3 4 5 6
Use
deployment URL for
inferencing

Convert the serving Python code to a Docker image
docker login docker.io -u YOUR_DOCKER_REPO
docker push docker.io/YOUR_DOCKER_REPO/image-seg-train:0.0.1
docker buildx build -o type=docker --platform=linux/amd64 -t YOUR_DOCKER_REPO/image-seg-train:0.0.1 .
Serving Dockerfile requirements.txt train.py

Focus on the serving Dockerfile
Load an official Docker image with

Python and all the TensorFlow
dependencies for GPUs
Add path pointing to the NVIDIA libraries

for using GPUs
Install all the packages specified in the

requirements.txt file
Create the directory needed to host the

code to run
It copies the infer.py from the local folder

to the previously created path
Start the WSGI server to run our Python

web application
https://gunicorn.org/
Demo: How to convert a serving Python code to a Docker image

0 1 2 3 4 5 6
Use
deployment URL for
inferencing

Demo: How to create a serving configuration and how to deploy a model in AI Core

0 1 2 3 4 5 6
Use
deployment URL for
inferencing

Demo: How to make an inference with the deployed model in Postman

Deployment of pre-trained models
https://developers.sap.com/tutorials/ai-core-tensorflow-byod.html
Docker: Create own Docker image

0.1 SAP BTP: Create AI Core and AI 4.1 Docker: Create own docker Image
Launchpad instances Git: Create a serving template
4.2 Git: Create a Serving template
0.2 AI Launchpad: Connect the AI Core
4.3 AI
AI Core:
Core:Choose
Choosethe scenario
the Scenario
runtime to AI Launchpad
0.3 AI Core: create an AI API client
4.4 AI Core:
AI Core:Create
Createa aconfiguration
Configuration
4.5 AI Core:
AI Core:Trigger
Triggerthe
the Serving
serving Execution
execution
0.4 AI Core: Connect a GitHub repo
AI Launchpad: Check the status of
0.5 AI Core: Connect a Docker repo deployment,check
deployment, checkthe
the logs
logs
0 2 3 4 5 6
Get the pre-
Use
AI Core and Create a trained model Get
AI Launchpad resource from open deployment
deployment URL for
onboarding group source URL
inferencing
framework
and register it
as artifact
AI Core one-time initial configuration ML model Inferencing

Lessons learned
▪ A web application for serving the training model has

to be designed and coded.
▪ The serving application code has to be converted
into a Docker image that is then uploaded to the
remote Docker registry.
▪ A serving template has to be prepared and uploaded
into the GitHub repository in order to instruct AI Core
about the Docker image to execute and the needed
resources.
▪ The web application can be easily started and
managed by means of the available endpoints.
▪ Pre-trained models can be easily registered in AI
Core and made available online by exposing an
endpoint with the very same deployment procedure.

Useful links
SAP-specific
▪ AI Core help page
▪ AI Launchpad help page
▪ SAP BTP help page
▪ SAP AI Portfolio blog post
Technical Links
▪ VS code plugin for AI Core
▪ SAP Computer Vision Package
▪ Metaflow library for SAP AI Core

Additional material
Specific to this course

▪ “Building Intelligent Scenarios and Sustainability on SAP BTP” blog
post series
▪ “Building Intelligent Scenarios and Sustainability on SAP BTP”
GitHub repository
Tutorials
▪ SAP HANA Academy video series about the AI Core and AI
Launchpad onboarding
▪ SAP HANA Academy video series about Defect Detection with
TensorFlow in AI Core
▪ Create Your First Machine Learning Project using SAP AI Core
▪ Use Pre-Trained TensorFlow Models with a GPU in SAP AI Core
▪ Create an Object Detection Project using Computer Vision Package
and SAP AI Core
Useful links
Data science-specific
▪ https://www.britannica.com/technology/artificial-intelligence
▪ https://ml-ops.org/content/end-to-end-ml-workflow
▪ https://www.v7labs.com/blog/image-segmentation-guide
▪ https://towardsdatascience.com/a-comprehensive-guide-to-
convolutional-neural-networks-the-eli5-way-3bd2b1164a53
▪ https://www.v7labs.com/blog/data-preprocessing-guide
▪ https://resources.github.com/ci-cd/
▪ https://ubuntu.com/blog/kubernetes-versus-docker
▪ https://www.v7labs.com/blog/what-is-computer-vision
▪ https://neptune.ai/blog/performance-metrics-in-machine-learning-
complete-guide
▪ https://www.analyticsvidhya.com/blog/2017/09/common-machine-
learning-algorithms/
Thank you.
open@sap.com
Follow all of SAP

warranty.

OpenSAP Sustai1 Week 2 All Slides

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

OpenSAP Sustai1 Week 2 All Slides

Uploaded by

Copyright:

Available Formats

Week 2: Enterprise AI in Action – Defect Detection

Unit 1: SAP AI Overview and Introduction

SAP AI Strategy and Overview

© 2022 SAP SE or an SAP affiliate company. All rights reserved. ǀ PUBLIC 2

The addressable market of enterprise applications with 60%

Other Apps 191 178 168 162 157 153 CAGR

$Billion USD 2019 2020 2021 2022 2023 2024

Source: IDC Worldwide Semiannual Artificial Intelligence Tracker, 2H19 – Draft

SAP is committed to building intelligent and sustainable enterprises

© 2022 SAP SE or an SAP affiliate company. All rights reserved. ǀ PUBLIC 5

AI embedded Custom Applications AI embedded SAP Applications

manage & operate AI content centrally

SAP AI Launchpad & tooling

SAP Business Technology Platform

© 2022 SAP SE or an SAP affiliate company. All rights reserved. ǀ PUBLIC 6

AI embedded Custom Applications AI embedded SAP Applications

manage & operate AI content centrally

SAP AI Launchpad & tooling

SAP Business Technology Platform

© 2022 SAP SE or an SAP affiliate company. All rights reserved. ǀ PUBLIC 7

AI embedded Custom Applications AI embedded SAP Applications

manage & operate AI content centrally

SAP AI Launchpad & tooling

SAP Business Technology Platform

© 2022 SAP SE or an SAP affiliate company. All rights reserved. ǀ PUBLIC 8

AI embedded Custom Applications AI embedded SAP Applications

manage & operate AI content centrally

SAP AI Launchpad & tooling

SAP Business Technology Platform

© 2022 SAP SE or an SAP affiliate company. All rights reserved. ǀ PUBLIC

AI embedded Custom Applications AI embedded SAP Applications

manage & operate AI content centrally

SAP AI Launchpad & tooling

SAP Business Technology Platform

© 2022 SAP SE or an SAP affiliate company. All rights reserved. ǀ PUBLIC

AI embedded Custom Applications AI embedded SAP Applications

manage & operate AI content centrally

SAP AI Launchpad & tooling

SAP Business Technology Platform

© 2022 SAP SE or an SAP affiliate company. All rights reserved. ǀ PUBLIC 11

AI embedded Custom Applications AI embedded SAP Applications

manage & operate AI content centrally

SAP AI Launchpad & tooling

SAP Business Technology Platform

© 2022 SAP SE or an SAP affiliate company. All rights reserved. ǀ PUBLIC 12

© 2022 SAP SE or an SAP affiliate company. All rights reserved.

▪ SAP AI Core key components

© 2022 SAP SE or an SAP affiliate company. All rights reserved. ǀ PUBLIC 2

▪ SAP AI Core provides an engine for running any

© 2022 SAP SE or an SAP affiliate company. All rights reserved. ǀ PUBLIC 3

▪ SAP AI Core is provided as a reuse service in

© 2022 SAP SE or an SAP affiliate company. All rights reserved. ǀ PUBLIC 4

▪ After configuring the AI Core service, we get the

© 2022 SAP SE or an SAP affiliate company. All rights reserved. ǀ PUBLIC 5

© 2022 SAP SE or an SAP affiliate company. All rights reserved. ǀ PUBLIC 6

© 2022 SAP SE or an SAP affiliate company. All rights reserved. ǀ PUBLIC 7

Full AI Core glossary

Full AI Core glossary

SDK AI Launchpad operations manager Other clients

© 2022 SAP SE or an SAP affiliate company. All rights reserved. ǀ PUBLIC 10

▪ Accessing metadata about scenarios and

© 2022 SAP SE or an SAP affiliate company. All rights reserved. ǀ PUBLIC 11

▪ SAP AI Core integrates with all the most popular

© 2022 SAP SE or an SAP affiliate company. All rights reserved. ǀ PUBLIC 12