{"id":1073,"date":"2021-10-23T08:41:54","date_gmt":"2021-10-23T08:41:54","guid":{"rendered":"https:\/\/salarydistribution.com\/machine-learning\/2021\/10\/23\/improve-your-data-science-workflow-with-a-multi-branch-training-mlops-pipeline-using-aws\/"},"modified":"2021-10-23T08:41:54","modified_gmt":"2021-10-23T08:41:54","slug":"improve-your-data-science-workflow-with-a-multi-branch-training-mlops-pipeline-using-aws","status":"publish","type":"post","link":"https:\/\/salarydistribution.com\/machine-learning\/2021\/10\/23\/improve-your-data-science-workflow-with-a-multi-branch-training-mlops-pipeline-using-aws\/","title":{"rendered":"Improve your data science workflow with a multi-branch training MLOps pipeline using AWS"},"content":{"rendered":"
\n

In this post, you will learn how to create a multi-branch training MLOps continuous integration and continuous delivery (CI\/CD) pipeline using AWS CodePipeline<\/a> and AWS CodeCommit<\/a>, in addition to Jenkins and GitHub. I discuss the concept of experiment branches, where data scientists can work in parallel and eventually merge their experiment back into the main branch. I also show you how to create an Amazon SageMaker<\/a> projects template that you can use from within Amazon SageMaker Studio<\/a>.<\/p>\n

SageMaker projects give organizations the ability to set up and standardize developer environments for data scientists and CI\/CD systems for MLOps engineers. With SageMaker projects, MLOps engineers or organization admins can define templates that bootstrap the machine learning (ML) workflow with source version control, automated ML pipelines, and a set of code to start iterating over ML use cases. With SageMaker projects, organizations can set up dependency management, code repository management, build reproducibility, and artifact sharing and management. SageMaker projects are provisioned using AWS Service Catalog products.<\/p>\n

SageMaker projects already provides a few MLOps CI\/CD pipeline templates, which are the recommended way to get started with CI\/CD in SageMaker. These training templates let you modify a single predefined branch called main, and all changes to this branch launch a training job. In some cases, though, you may want to use a multi-branch (trunk-based) training pipeline. This type of pipeline lets you get more flexibility and better performance with a thorough review of the code and experiment results before approving models and merging changes into main. This solution enables you to create several experimental branches that each launch their own training job and create their own model artifact. We can then use pull requests to approve some of these models, using a modified SageMaker template available in GitHub<\/a>.<\/p>\n

When data scientists are working on a new model, that work is often experimental, meaning that unsuccessful experiments may be discarded and those with good results can go into production. Each data scientist may be working on a particular attempt at improving the current objective function. While one may be trying out a different architecture, another may be trying a new set of hyperparameters, for instance.<\/p>\n

Whatever may be the case, these experiments should be version controlled, tracked, and automated<\/a>. When the winning experiment is found and becomes eligible to be merged into the main branch, it can be assessed by a lead data scientist. They can view the exact code that was run, identify the key metrics output by the experiment, and have the results of any automated tests before model approval and release. When each experiment has its own branch, you can automate training whenever a Git push is run. This way, we can make sure that all metrics and tests for the model are stored in Studio.<\/p>\n

The following image illustrates a potential Git history of two data scientists experimenting on the same project.
\"\"<\/a><\/p>\n

This post includes the following sections:<\/p>\n