TFlite example

This example builds an image recognition example in a Docker image with TFlite for Arm64 architecture.

Github Actions workflows are used to:

• build the Docker image and store it in the Github container registry
• deploy the application on a Github runner and test the application output

Local build

To build the Docker image:

./build.sh

Local run

Run the image classification example:

./run.sh  | tee app.out

The application output is saved in the app.out file. This file can be used to check the output.

Test application output locally

To run the tests, shell2junit is used. To install it:

git clone https://github.com/manolo/shell2junit

Run the tests:

. shell2junit/sh2ju.sh
./test.sh

Worflows

Build - Github-hosted

In .github/workflows/build.yml you will find instructions to build the Docker image with the application.

The worflow can be launched manually from the Actions tab and Build > Run workflow. The yml file can also be edited to trigger a build for every push or pull request.

Test - self-hosted

Add a self-hosted Github actions runner

To run the Test workflow, add a self-hosted runner from the Settings tab of the project Actions > Runners and Add self-hosted runner.

In the menu, select Linux as the runner image, and ARM64 as the achitecture. Copy and paste the instructions displayed on the target console to install and start the agent.

Note: for the agent to start automatically at launch, configure it as a service

sudo ./svc.sh install
sudo ./svc.sh start

Launch on self-hosted runner

In .github/workflows/test.yml you will find instructions to deploy the Docker image on the self-hosted runner, execute and test the application.

When a self-hosted runner is available, the worflow can be launched manually from the Actions tab and Build > Run workflow. The yml file can also be edited to trigger a build for every push or pull request.

AVH Launch Test - self-hosted

When an instance has been added as a runner and with the agent configured as a service, it can started automatically on demand with this workflow. This worflow will run the tests and stop the instance when completed.

This workflow requires:

• The ID of the instance configured as a github runner. You can get it from the URL when accessing the device in the AVH interface, e.g. https://app.avh.arm.com/devices/<ID>/connect. Save it as a github actions secret called AVH_ID.

• The AVH API token. Save it as a github actions secret called AVH_TOKEN.

It relies on a CLI bash tool to interact with the AVH API which can be found here.

References

• Original TF example code
• TFlite example guide
• TFlite Docker image example for Aarch64