Tutorial

Deploy deep learning models on Red Hat OpenShift

Deploy model-serving microservices from the Model Asset Exchange on Red Hat OpenShift

Save Like

By Patrick Titzler
Published August 16, 2019

Learning objective

By completing this introductory tutorial, you learn how to deploy a deep learning microservice from the Model Asset Exchange on Red Hat OpenShift.

Upon completion of the tutorial, you know how to use the OpenShift web console or OpenShift Container Platform command-line interface (CLI) to:

  • Create a new project
  • Deploy a model-serving microservice from a public container image on Docker Hub
  • Create a route that exposes the microservice to the public

Prerequisites

To follow this tutorial, you must have:

The tutorial was created using Red Hat OpenShift v3.11 on IBM Cloud. You will notice a few GUI differences if you complete the tutorial using Red Hat OpenShift Online, which at the time of writing was version 3.

Estimated time

It should take you approximately 30 minutes to complete this tutorial. The tutorial modules are:

Tutorial setup

To complete this tutorial you must have access to a small 1-node OpenShift cluster.

Provision a Red Hat OpenShift cluster

If you don’t have access to an existing cluster, provision one on the IBM cloud or on openshift.com.

You can complete this tutorial using the free starter tier on openshift.com. However, note that RAM is limited to 2 GiB, which might not always be sufficient for resource-intensive models.

Throughout this tutorial, the instructions and screen shots refer to a cluster named os-max-test-cluster. Whenever necessary, substitute this name with the name of your cluster.

You are ready to start the tutorial.

Deploy a deep learning model-serving microservice on Red Hat OpenShift

In this tutorial, you will deploy a model-serving microservice from the Model Asset Exchange on Red Hat OpenShift using the OpenShift web console or the OpenShift Container Platform CLI. You can complete both parts or only one part.

Deploy using the OpenShift web console

In this first part of the tutorial, you deploy the MAX-Object-Detector model-serving microservice, which can be used to identify objects in pictures, as shown in this demo.

  1. Open the OpenShift web console.

    1. In a web browser, open the Kubernetes cluster page on IBM Cloud.
    2. Locate your OpenShift cluster entry os-max-test-cluster.

    3. From the … drop-down menu, select OpenShift web console. OpenShift web console button

      The catalog browser is displayed.

  2. From the drop-down menu, select the Application console. A list of your projects is displayed. A project is used in OpenShift to organize related resources.

    Switch to application console

  3. Create a new project and assign it a name (such as max-deployments), a display name, and an optional description.
  4. Open the project.

Deploy the microservice Docker image

You can deploy Docker images that are hosted in public or private registries. The MAX-Object-Detector Docker image codait/max-object-detector is hosted on Docker Hub.

  1. On the project overview page, choose Deploy Image.

    Project screen

    If your Overview page looks different (for example, because you are using an existing project that already contains an application), use the Add to Project drop-down menu in the menu bar and select Deploy Image.

  2. Select the model-serving Docker image that you want to deploy.

    1. Choose the Image Name radio button to select a public or private repository as the source.

    2. Enter codait/max-object-detector as Image name or pull spec (or enter the name of another model-serving Docker image from our repository on Docker Hub).

      Select container image

    3. Click on the magnifying glass next to the image name (or press Enter) to load the Docker image’s metadata.

  3. Review the deployment configuration for the selected Docker image.

    Review deployment configuration

    The Name field is pre-populated using the name of the Docker image. OpenShift uses this name to identify the resources being created when the application is deployed.

    If you append, modify, or delete a few characters, you’ll notice how the change is impacting the generated image stream name, the deployment configuration name, the service name, and the host name. (However, leave the name as it is in this tutorial!)

    Review environment variable settings

    You can customize the behavior of most model-serving microservices by setting environment variables. For example, you can enable Cross-Origin Resource Sharing support by setting the CORS_ENABLE variable in the deployment configuration to true. Refer to the model documentation for model-specific environment settings.

  4. Deploy the Docker image.

  5. Close the window and open the Overview tab if it is not selected by default. The deployment configuration for the Object Detector model-serving microservice is displayed.

    View deployment

    By default, only one instance of the microservice is started. You can manually scale out the deployment (“run more instances of the microservice to handle more requests concurrently”) by increasing the target number of running pods. You can also enable autoscaling (“run up to X instances if needed”) by clicking the deployment configuration name, and selecting Configuration and Add Autoscaler.

After the microservice is deployed, it is only visible internally (at port 5000) to the cluster. To expose it to the public, you must create a route.

Create a route

When you create a route in OpenShift, you have the option to expose an unsecured or a secured route. Because the model-serving miroservice communicates over HTTP, you can configure the router to expose an unsecured HTTP connection (which is what you’ll do in this tutorial) or expose a secured HTTP connection, which the OpenShift router automatically terminates.

  1. Select Create Route under the NETWORKING section.

  2. Review the configuration settings. The defaults expose an unsecured HTTP connection at a generated URL if you leave the hostname empty. Review route configuration

    To configure a secured HTTP route, you’d select Secure route, set TLS Termination to Edge, and set Insecure Traffic to None. You can specify your own TLS certificate or use the router’s default certificate.

  3. Create the route. Upon completion of the operation, the public URL for the deployed model-serving microservice is displayed. Route created

    If an error is displayed indicating that the router rejected the request, shorten the route name. The maximum length of the generated URL (derived by concatenating the route name, the project name, the cluster name, and the router’s domain) likely exceeds the 63 character maximum.

  4. Open the displayed URL to verify that the microservice’s OpenAPI specification endpoint can be accessed.

    Review OpenAPI specification

  5. If you have been following this tutorial using the Object Detector, open the embedded sample application by appending /app to the URL displayed in your browser, and test the deployed service by submitting an image.

    Test embedded sample application

Skip to the tutorial summary or review the instructions in the next section to learn how to deploy a model-serving microservice using the command-line interface.

Deploy using the CLI

You can deploy and manage applications using the OpenShift Container Platform CLI (oc). Before you can use the CLI to deploy the microservice in your cluster, you must log in.

  1. Open the OpenShift web console and copy the login command.

    1. In a web browser, open the Kubernetes cluster page on IBM Cloud.
    2. Locate your OpenShift cluster entry os-max-test-cluster.

    3. From the … drop-down menu, select OpenShift web console.

      OpenShift web console button

      The catalog browser is displayed.

    4. From the avatar drop down, select Copy Login Command.

      Copy login command

    Do not share the login command. It contains a token that grants access to the resources in your account.

  2. Log in to your cluster using the OpenShift CLI.

    1. Open a terminal window.

    2. Paste the copied login command.

      oc login https://... --token=...

    Tip: To learn more about a command, run oc <command> --help, for example, oc login --help.

  3. Create a new project using a name such as max-deployments-cli. OpenShift uses a project to organize related resources.

    oc new-project max-deployments-cli

    You can list existing projects using oc projects or switch between projects using oc project <project-name>.

Deploy the Docker image

You can deploy Docker images that are hosted in public or private registries. In this part of the tutorial, you deploy the MAX-Image-Caption-Generator model-serving microservice, which can be used to describe the content of a picture in a sentence, as shown in this demo.

  1. Verify that you can access the codait/max-image-caption-generator Docker image on Docker Hub.

    oc new-app --search codait/max-image-caption-generator

    The output should look as follows:

     Docker images (oc new-app --docker-image=<docker-image> [--code=<source>])
 -----
 codait/max-image-caption-generator
 Registry: Docker Hub
 Tags:     latest

  2. Deploy the Docker image.

    oc new-app codait/max-image-caption-generator

    Review the output and note that the image name max-image-caption-generator is used by default to name the generated resources, such as the image stream, the deployment configuration, the service, and the host.

    --> Found Docker image ... (... days old) from Docker Hub for "codait/max-image-caption-generator"

 * An image stream tag will be created as "max-image-caption-generator:latest" that will track this image
 * This image will be deployed in deployment config "max-image-caption-generator"
 * Port 5000/tcp will be load balanced by service "max-image-caption-generator"
   * Other containers can access this service through the hostname "max-image-caption-generator"
 * WARNING: Image "codait/max-image-caption-generator" runs as the 'root' user which may not be permitted by your cluster administrator

--> Creating resources ...
 imagestream.image.openshift.io "max-image-caption-generator" created
 deploymentconfig.apps.openshift.io "max-image-caption-generator" created
 service "max-image-caption-generator" created
--> Success
...
Run 'oc status' to view your app.

    You can change the default by supplying the --name <my-custom-name> parameter when you deploy the image.

  3. You can customize the behavior of most model-serving microservices by setting environment variables. For example, you can enable Cross-Origin Resource Sharing support by setting the CORS_ENABLE variable in the deployment configuration to true:

    oc set env deploymentconfig max-image-caption-generator CORS_ENABLE=true

    Refer to the model documentation for model-specific environment settings.

  4. Query the deployment status.

    oc status

    Note that after deployment one microservice instance is running and it is only visible internally (at port 5000) to the cluster:

    In project max-deployments-cli on server https://...

svc/max-image-caption-generator - ...:5000
  dc/max-image-caption-generator deploys istag/max-image-caption-generator:latest
    deployment #1 deployed 15 minutes ago - 1 pod

    You can manually scale out the deployment by increasing the target number of running pods (oc scale --replicas=2 deploymentconfig max-image-caption-generator) or configure autoscaling.

    To expose the service to the public, you must create a route.

Create a route

When you create a route in OpenShift, you have the option to expose an unsecured or a secured route. Because the model-serving miroservice communicates over HTTP, you can configure the router to expose an unsecured HTTP connection (which is what you’ll do in this tutorial) or expose a secured HTTP connection, which the OpenShift router automatically terminates.

  1. Create a route for the service.

    oc expose service max-image-caption-generator

  2. Retrieve the microservice’s public URL.

    oc get route max-image-caption-generator

    Under the HOST/PORT column, the generated host name is displayed.

    NAME                          HOST/PORT                                                          ...     
max-image-caption-generator   max-image-caption-generator-max-deployments-cli...appdomain.cloud  ...

    If the displayed HOST/PORT reads InvalidHost, the generated host name is invalid. This commonly happens when the length exceeds the 63-character maximum. To resolve the issue, shorten the route name. For example, to shorten the route name for the max-image-caption-generator service, run oc expose service max-image-caption-generator --name <shorter-route-name> followed by oc get route <shorter-route-name> to retrieve the public URL.

  3. Open the displayed URL in your web browser to verify that the microservice is accessible.

    Review OpenAPI specification

  4. Validate that the /model/predict endpoint returns the expected result for a picture of your choice.

    Note: For some compute-intensive models, calls to the /model/predict endpoint might result in HTTP error 504 (Gateway timeout) if you use the router’s default configuration. To resolve the issue, increase the router’s timeout value by running oc annotate route <router-name> --overwrite haproxy.router.openshift.io/timeout=<number-of-seconds>s.

Summary

In this tutorial, you learned how to use the OpenShift web console or OpenShift Container Platform CLI to:

  • Create a new project
  • Deploy a model-serving microservice from a public container image on Docker Hub
  • Create a route that exposes the microservice to the public

To learn more about the Model Asset Exchange or how to consume the deployed model-serving microservice in Node-RED or JavaScript, take a look at Learning Path: An introduction to the Model Asset Exchange and these pens.

You might find the following resources useful if you’d like to learn more about Red Hat OpenShift:

Legend

Related

Event | Conference

Codemotion Online Tech Conference - Italian Edition

November 24, 2020 2:00 pm CET
Event | Conference

Open Source 101 at Home Conference

May 12, 2020 10:00 am EST
Event | Hackathon

Hack for Wuhan against COVID-19

March 6, 2020 9:00 am CST