Building an ML Experimentation Platform for Easy Reproducibility

Machine learning workflows are not linear, where experimentation is an iterative & repetitive to and fro process between different components. What this often involves is experimentation with different data labeling techniques, data cleaning, preprocessing and feature selection methods during model training, just to arrive at an accurate model. Quality ML at scale is only possible when we can reproduce a specific iteration of the ML experiment–and this is where data is key. This means: capturing the version of training data, ML code and model artifacts at each iteration is mandatory. However, to efficiently version ML experiments without duplicating code, data and models, data versioning tools are critical. Open source tools like lakeFS make it possible to version all components of ML experiments without the need to keep multiple copies, and as an added benefit, save you storage costs as well. In this talk, you will learn how to use a data versioning engine to intuitively and easily version your ML experiments and reproduce any specific iteration of the experiment. This talk will demo through a live code example: Creating a basic ML experimentation framework with lakeFS (on Jupyter notebook) Reproducing ML components from a specific iteration of an experiment Building intuitive, zero-maintenance experiments infrastructure All with common data engineering stacks & open source tooling.