Build CI/CD pipelines for YOLOv5 object detection service on Azure.

Repo Structure

• Generated from this template.
• .pipelines: contains YAML files describing the CI/CD pipeline on Azure DevOps.
• bootstrap: scripts used to bootstrap a new project from the template.
• docs: original how-to guide to setup a CI/CD on Azure, with a simple regression model.
• environment_setup: YAML files used to setup cloud enviroment, provision resources, etc.
• ml_service: scripts using Azure Python SDK to build training, registration and evaluation pipelines on Azure Machine Learning Studio.
• yolov5: main scripts to interact with Azure Virtual Machines, containing the original YOLOv5 code.

Setup CI pipeline:

• Basically, I created an Azure DevOps pipeline that was triggered whenever a commit is pushed/merged to the master branch, to run another pipeline on Azure Machine Learning Studio that would retrain my yolov5 model.
• Configure the .pipeline/yolov5-ci.yml, which trigger the build pipeline for model training. - Write training code
• Setup .env
• Register coco128 dataset on Datastores, which would be used to quickly train the model (since this repo's purpose is to demonstrate how to setup a CI/CD pipeline, I will not waste time on training the model, just use a small dataset).
• Config training params in parameters.json
• Write eval code: evaluate the newly trained mode, if its metric is better than the current model on production, register that model to Azure ML Studio, otherwise cancel the pipeline run!
• Eval metric: mean Average Precision under IoU 0.5 to 0.95 mAP_0.5_0.95, which will be saved in a tag during the training process, just read that tag to get the current model’s metric.

Setup CD pipeline

• Create a new Azure DevOps pipeline specified by yolov5_cd.yml, which would be triggered whenever the CI pipeline is done.
• Automatically deploy the newly registered model on Azure ML Realtime Endpoint, both managed compute instance and Kubernetes Services.