Building an MLOps stack for rapid delivery of reproducible computer vision projects

At Plant & Food Research there are many opportunities to use computer vision solutions to help ask questions across a range of science disciplines. The diversity and volume of potential applications has strongly motivated our work for standardisation and automation of workflows. In addition, productionising machine learning is a key step to produce useful downstream applications.

These goals are fundamental to the emerging concept of MLOps (machine learning operations), which is the set of practices that brings DevOps (development operations) from the field of software engineering to machine learning. Over the past two years, we have been developing a computer vision platform that incorporates MLOps practices. This technological stack allows us to deliver computer vision applications in a repeatable, reproducible and efficient manner.

This talk will provide a framework of modules comprising an MLOps stack, and discuss how each is implemented in our platform. It will cover the motivations, methods, learnings and future plans of our platform, as well as some examples of successful PFR computer vision projects.

ABOUT THE AUTHOR

Daniel Bentall trained as a mechatronics engineer, working for 7 years in electronics and embedded software design for mobile robotics at a startup company before retraining as a data scientist. During his Master of Applied Data Science degree at the University of Canterbury he interned at Plant & Food Research, developing a deep learning model for detecting kiwifruit on the vine. After completing his degree in mid-2019 he began working full time at PFR on a wide variety of computer vision problems.