There is generally a gap between data science practitioners who create machine learning models and the engineering/devops teams and processes that put those models into production. This often leads to data scientists not having efficient development cycles to properly test and extend the models before they make it to production, as well as limiting the ability to extend to production ML capabilities introducing advanced monitoring such as outlier detectors, drift detectors, explainers, etc.
In this talk we introduce key Python based tools and techniques that will enable machine learning practitioners to have efficient end to end workflows to productionise their machine learning models in scalable Kubernetes clusters without requiring any devops or infrastructure knowledge, whilst still providing the key functionality to allow for structured handover to devops teams once all advanced ML components are developed and tested.
The key library we will be using is the Tempo SDK, an open source Python framework for data scientists to easily prepare and test their machine learning models for production and provides capabilities to deploy them either directly or as part of CI/CD and Gitops processes. It allows data scientists to be involved in the inference logic needed to call those models correctly and allows them to easily add other required inference components for their models such as outlier detectors and explainers as part of their work.
The library is designed to be agnostic to the inference runtime allowing local Docker runtimes where production container images can be validated before passing the model artifacts on to production runtimes such as Seldon or KFServing for running the models on production Kubernetes clusters.
Tempo allows any custom python logic to be packaged along with dependencies by using tools such as conda-pack and cloudpickle to provide a simple APIs to package and push model artifacts ready for local and remote deployment.
This talk will discuss Tempo’s design decisions and illustrate its utility for data scientists deploying simple models to more complex outlier detection, explainer and multi-model ensembles with custom business logic.
Speaker: Mr. Alejandro Saucedo / London / The Institute for Ethical AI & Machine Learning - Website, GitHub, Twitter, LinkedIn Language: English Date and Time : October 8, 2021 / 16:30-17:00 (UTC+8)
Speaker Introduction
Alejandro is the Chief Scientist at the Institute for Ethical AI & Machine Learning, where he leads the development of industry standards on machine learning explainability, adversarial robustness and differential privacy. Alejandro is also the Director of Machine Learning Engineering at Seldon Technologies, where he leads large scale projects implementing open source and enterprise infrastructure for Machine Learning Orchestration and Explainability. With over 10 years of software development experience, Alejandro has held technical leadership positions across hyper-growth scale-ups and has a strong track record building cross-functional teams of software engineers.