did you know that 87 of data science projects never make it into production that's just one out of every 10 ai projects so how do you make sure your project does not become one of those 87 percent i am priyanka vergara and in this video i will walk you through steps to operationalize your machine learning workflow using vertex pipelines so your project becomes one of those 10 successful ai projects we will look at mlab's framework and vertex ai pipelines which helps you break down the complex multi-step ml workflow into a pipeline and ensure that each
step runs in a reproducible auditable cost effective and a scalable way in contrast with typical software systems which are traditionally code centric ml systems involve an intricate relationship between data code and the models now data and models are unique artifacts with their own dependencies and pitfalls and the joint management of these three artifacts is a challenge in delivering and maintaining production machine learning systems that is what ml ops is ml ops is a culture in practice that aims to unify ml system development and operations and guide through the challenges of taking ml projects from experimentation
to production it is just like devops but for ml hence mlobs which handles managing the life cycle of data models and code let's start with a high level ml ops framework so the first step is ml development which is inherently experimental where data scientists and model builders explore and transform the data sets explore different algorithms train many models and then compare them then the second step is continuous training now production training should be automatable and repeatable so it can take new data or other triggers and generate new and better models on demand and then the
third step is model deployment which is really all about continuous integration and continuous delivery this involves running a b tests evaluation of model behavior in production and then approving them for releasing following a rigorous auditable and even reversible ci cd process and then lastly it's continuous monitoring you need to continuously monitor your models when they are running in production because you want to get a sense of how they're performing this is important to ensure the quality and business continuity but also crucial for getting high quality signals into how to improve a model for your next
iteration model management and governance is basically the step that applies to all of these steps it's the it's the entire framework we need traceability verifiability and auditability at each stage of this workflow ml pipeline models this ml ops workflow a pipeline is a way of modeling a workflow as a set of connected steps each step takes as input the output of the previous step and performs some additional computations and then produces outputs that can be utilized by the future components now here in this ml ops workflow to keep the spirit of experimentation in the ml
development phase we recommend using the reusable training pipelines and the reason for usable pipelines is because the data scientists can share components amongst themselves so they don't have to start with a blinking cursor and to rerun and iterate quickly to create those optimal models the output of this step would be the trained pipeline source code which you feed into the continuous training phase the pipeline generates generated earlier can be now treated like a software application and is version controlled and deployed through ci cd process as a training pipeline that can now be invoked with new
data or parameters in production as needed now the output of this step would be your trained model and then the output of the model deployment is your live predictions and these predictions can not not just be predictions but also logs and other records of production inferences now the output of continuous monitoring because you are monitoring um the train the model continuously the output of this continuous monitoring is those alerts that go out to the team so that they can continuously uh measure and monitor what the model's performance is looking like is it degrading or triggering
rollbacks and retaining a b testing of any candidate models the model management and governance capabilities obviously as i said earlier span across the entire framework including the features such as model registry model approvals and model provenance now what are vertex pipelines vertex pipelines help you automate and monitor your ml systems by orchestrating this ml ops workflow in a serverless manner they are based on containers which help you make the ml ops process uh really portable and scalable and the vertex pipeline store the workflow artifacts using vertex aiml metadata which makes it easier to analyze the
lineage of our workflow and the items that are being created for example an ml model's lineage may include training data the hyper parameters and the code that we use to create the models which are critical pieces of information to understand the changes in performance or accuracy of our ml systems now as you can imagine managing this metadata by yourself in an ad hoc manner can be difficult and time consuming and vertex pipelines makes it easy to do that now vertex pipelines support both kubeflow pipelines and also tensorflow extended now if you are already using tensorflow
tensorflow extended is a good choice for you it provides a rich set of components that helps you take the tensorflow code and make it into an ml pipeline i did a video on tfx not too long ago which i will include in the description below now if you're not using tensorflow then use a kfp which is an open source machine learning pipeline it offers a great deal of flexibility it's easy to plug in code from any ml framework including the ml frameworks that aren't python based such as apache mxnet now you might ask why do
i need vertex pipelines if i use kubeflow pipelines already today now the biggest reason is that it is managed and because it is managed you don't have to maintain or create servers by yourself now when you're using kfp you need to build scale and maintain a kubernetes cluster all by yourself but with managed pipelines we don't have to do any of that work it's all serverless okay so in this video we got a quick overview of ml ops workflow and vertex pipelines what's next let's take a look at vertex pipeline's demo in our next video
in the meantime try out the codelab linked below and subscribe to get notified about the next video you
