Sample for setting up different kinds of SageMaker endpoints with the YOLOv9 object detection model running in a custom Docker container.
Github Actions is used to build and push a custom Docker container to AWS ECR. A YOLOv9 model is baked into that image, which is used to deploy the YOLOv9 object detection SageMaker endpoints.
- OIDC set up between Github and your AWS account.
- An ECR repository to hold the custom image
- Update placeholders
Update placeholders with your own values in both Github Actions workflows (OIDC_ROLE, ECR_REPOSITORY, AWS_REGION and S3_ASYNC_OUTPUT)
- Check in and commit the code to your repo
Check in the code to your Github repository to trigger the Docker image build and SageMaker endpoint Cloudformation deployment. Note that the first Cloudformation deployment may fail, since the the image build may not have finished in time. If that happens, simply trigger the deployment again.
- Run test script with your endpoint
In test_endpoint.ipynb there's sample code that invokes the endpoint and draws bounding boxes on the provided image. You can find the SageMaker endpoint name in the Cloudformation Stack outputs in the AWS console or using the AWS cli.
Note
There are 2 base images in the Dockerfile (one is for GPU and one is for CPU). You can use whichever you want, but serverless endpoints do not support GPU, and the GPU image is too large for even using on a serverless endpoint.
Warning
Make sure to remove the SageMaker endpoint types from the cloudformation you don't want to use to avoid unnecessary cost.