Introduction
In this blog, we shall be looking into how we can deploy an image classification model that you have stored on your local computer with SAP AI core. Together with this blog and our SAP AppGyver blog, you will be able to host your very own image classification model on AppGyver and make live predictions of images.
Business Context
This blog is directly linked to SAP’s mobile application called Ariba which uses the “Aribot” model to make classifications for images. SAP's Ariba mobile application helps field technicians to identify the type of a mechanical part using image recognition. It retrieves specifics in detail, enabling users to create a purchase request on the spot, automating this usually tedious process. We shall be tackling the image modelling part that goes into SAP Ariba.
What you will learn
At the end of this blog, you will learn how to deploy and make inferences on images using your own keras model on SAP AI Core and retrieve a deployment link that exposes prediction API endpoints for mobile applications to use.
Pre-requisites
- SAP AI Core access keys via SAP BTP Account
- AWS S3 Service keys
- Python (We shall be using the SAP AI Core SDK)
- Jupyter Notebook installed with Anaconda
- Docker desktop account synced with SAP AI Core. If you have not synced your Docker with SAP AI Core, find out how here by following steps 1,2 and 7.
- Github account synced with SAP AI Core. If you have not synced your Github with SAP AI Core, find out how here by following steps 1 to 3.
- Pre-built keras model.
It is highly recommended that you watch the SAP AI Core tutorials to familiarize yourself with SAP AI Core.
Let’s begin!
Step 1
Launch a Jupyter notebook and connect to SAP AI Core by copying the code below. Please enter your SAP AI Core service keys.
from ai_core_sdk.models import Artifact
from ai_core_sdk.ai_core_v2_client import AICoreV2Client
ai_core_client = AICoreV2Client(
# `AI_API_URL`
base_url = "<base url>" + "/v2", # The present SAP AI Core API version is 2
# `URL`
auth_url= "<auth url>" + "/oauth/token",
# `clientid`
client_id = "<client id>",
# `clientsecret`
client_secret = "<client secret>"
)
Step 2
Create a resource group with ID “aribot”. We have named it aribot as the model aribot has been created with SAP AI Core previously. You can choose any name you want.
response = ai_core_client.resource_groups.create("aribot")Step 3
Upload your keras model to AWS S3 Bucket. Run the following function. Ensure that your model is located within your present working directory.
import logging
import boto3
from botocore.exceptions import ClientError
import os
def upload_file_to_s3(file_name, bucket, object_name=None):
"""Upload a file to an S3 bucket
:param file_name: File to upload
:param bucket: Bucket to upload to
:param object_name: S3 object name. If not specified then file_name is used
:return: True if file was uploaded, else False
"""
# If S3 object_name was not specified, use file_name
if object_name is None:
object_name = os.path.basename(file_name)
# Upload the file
s3_client = boto3.client('s3')
try:
response = s3_client.upload_file(file_name, bucket, object_name)
except ClientError as e:
# logging.error(e)
return False
return True
model_filename = 'keras_model.h5'
res = upload_file_to_s3(img_filename, '<S3 BUCKET NAME>', f'aribot/model/{model_filename}')
To see your model in S3, run the following command in your terminal:
aws s3 ls s3://<BUCKET NAME>/aribot/model/Step 4
Create an object store secret to connect your AWS S3 bucket to SAP AI Core. Note that the parameter path_prefix should point to the parent folder of the folder where your model is in AWS S3. Use your AWS S3 Access keys to fill up the keys.
response = ai_core_client.object_store_secrets.create(
resource_group = 'aribot',
type = "S3",
name = "aribot-secret",
path_prefix = "aribot",
endpoint = "<fill up>",
bucket = "<fill up>",
region = "<fill up>",
data = {
"AWS_ACCESS_KEY_ID": "<fill up>",
"AWS_SECRET_ACCESS_KEY": "<fill up>"
}
)
print(response.__dict__)
You should see the response: ‘message secret has been created’.
Step 5
Register the model as an artifact and save the artifact id in a variable. Make a note of the ID. This ID uniquely identifies your artifact.
response = ai_core_client.artifact.create(
resource_group = 'aribot',
name = "aribot-model",
kind = Artifact.Kind.MODEL,
url = "ai://aribot-secret/model",
scenario_id = "aribot-run-1",
description = "Aribot model run 1"
)
print(response.__dict__)
artifact_id = response.__dict__['id']
Step 6
Set up the serving code to be containerized in Docker. Note, the current serving code is meant for the Aribot model.
Download and save each of the following files to your local computer. These are the code files that you will upload as a docker image and use later for deployment in SAP AI Core. All the files should be located in the same folder.
Dockerfile
requirements.txt
tf_template.py
serve.py
Next, log into your docker account using terminal. This command can only be run if you have the docker desktop set up in your computer.
docker login <YOUR_DOCKER_REGISTRY> -u <YOUR_DOCKER_USERNAME>Navigate to your downloaded files location and build docker the image using the code files that you downloaded.
docker build -t docker.io/<YOUR_DOCKER_USERNAME>/aribot:0.0.1 .
docker push docker.io/< YOUR_DOCKER_USERNAME>/aribot:0.0.1
Step 7
Create workflow to serve your model. Copy and paste the following .yaml file within your Github repository folder which is connected to SAP AI Core. Your folder directory should look something like this.
apiVersion: ai.sap.com/v1alpha1
kind: ServingTemplate
metadata:
name: aribot-classifier-1 #Enter a unique name here
annotations:
scenarios.ai.sap.com/description: "Aribot image classification"
scenarios.ai.sap.com/name: "aribot-image-clf"
executables.ai.sap.com/description: "Aribot Tensorlfow GPU Serving executable"
executables.ai.sap.com/name: "aribot-serve-exectuable"
labels:
scenarios.ai.sap.com/id: "aribot-clf"
ai.sap.com/version: "1.0.0"
spec:
inputs:
artifacts:
- name: modelArtifact
template:
apiVersion: "serving.kubeflow.org/v1beta1"
metadata:
labels: |
ai.sap.com/resourcePlan: starter
spec: |
predictor:
imagePullSecrets:
- name: docker-registry-secret
containers:
- name: kfserving-container
image: "docker.io/<YOUR_DOCKER_USERNAME>/aribot:0.0.1"
ports:
- containerPort: 9001
protocol: TCP
env:
- name: STORAGE_URI
value: "{{inputs.artifacts.modelArtifact}}"
- Under ai.sap.com/resourcePlan, you can choose to use 'Infer-S' instead. This will enable the GPU node in deployment. Find all the available resource plans here.
- Replace docker-registry-secret with the name of your docker registry secret. You can create and use multiple docker secrets in SAP AI Core.
- Under containers, set your docker image URL.
Step 8
Sync workflow with SAP AI Core. We shall now create an application using AI Core SDK. Run the following code in your jupyter notebook. Remember to replace the github repository url with your own one.
response = ai_core_client.applications.create(
application_name = "aribot-clf-app-1",
revision = "HEAD",
repository_url = "https://github.com/<GITHUB_USERNAME>/aicore-pipelines",
path = "aribot"
)
print(response.__dict__)
You should then see the following response: ‘Application has been successfully created’
We shall now verify that the workflow has synced between SAP AI Core and Github. Run the following code till you see the message
- ‘healthy_status’: ‘Healthy’,
- ‘sync_status’: ‘Synced’
response = ai_core_client.applications.get_status(application_name = 'aribot-clf-app-1')
print(response.__dict__)
print('*'*80)
print(response.sync_ressources_status[0].__dict__)
Step 9
After your workflow has been synced, launch SAP AI Lauchpad and log into your SAP AI Core account. We shall now create a configuration for deployment with our model.
Create a new configuration by heading to Configuration > Create. Ensure that you have the correct resource group chosen.
Next, select the respective variables for the configuration. Use the screenshot below as a reference.
Next, choose your model as the artifact and select review > create > create deployment.
Once completed, you should be able to see this:
Copy the deployment URL and the deployment ID.
Step 10
We shall now make a prediction on a test image. Copy the following code:
img_filename = './raw_images/IMG_5694.JPG' #images are stored in the raw_image folder in your local machine.
def get_img_byte_str(): #returns img in base 64 encoded
from io import BytesIO
import base64
with open(img_filename, "rb") as fh: #rb - read img as raw bytes
buf = BytesIO(fh.read()) #reading file and constructing bytes io obj from the file
img_data = base64.b64encode(buf.getvalue()).decode() #so can send over the network
return img_data #base 64 encoded
base64_data = get_img_byte_str()
import requests
# URL
deployment_url = "<INSERT_YOUR_DEPLOYMENT_URL>" + “/v1/predict” #need to append endpoint
# Preparing the input for inference
test_input = {'imgData':get_img_byte_str()}
endpoint = f"{deployment_url}" # endpoint implemented in serving engine
headers = {"Authorization": ai_core_client.rest_client.get_token(),
'ai-resource-group': "aribot",
"Content-Type": "application/json"}
response = requests.post(endpoint, headers=headers, json=test_input)
print('Inference result:', response.content)
You should then see your predictions being made. For example, for this blog, we get the output as so:
Inference result: b'{"Prediction": "Adapter Pk2", "Probability": "1.0"}Conclusion
You have reached the end of the tutorial, congratulations! Hope you were able to run through all the code blocks successfully. After this, you can follow the next tutorial on how to use the deployment link with SAP AppGyver to deploy this model on a mobile application. If you have any questions or comments, feel free to post one down below in the comment section! To learn more about this topic, check out the following links:
If you liked this blog, follow my profile as I'll be posting more in the future!
- SAP Managed Tags:
2 Comments
