In an period of fixed innovation, there may be an growing want for steady iteration, which results in a extra advanced mannequin growth lifecycle. Retaining observe of all of the inputs and outputs together with options, metrics and artifacts for every mannequin model could be troublesome and, at occasions, tedious.
In pursuit of simplifying this course of, Capital One created rubicon-ml, an open-source machine studying (ML) answer that may observe, visualize and share experiments with collaborators and reviewers. These capabilities might help knowledge scientists and technologists experiment, practice and govern fashions designed to unravel advanced enterprise issues.
“Earlier than a mannequin is definitely pushed to manufacturing, ML specialists conduct hundreds of experiments with completely different enter parameters that end in numerous outputs,” mentioned Sri Ranganathan, director of ML engineering at Capital One and proprietor of rubicon-ml. “Rubicon tracks these experiments throughout the mannequin growth lifecycle and might present the standing of code for any given parameter.”
Ranganathan went on to clarify that rubicon-ml simplifies mannequin governance, auditability and reproducibility by explaining how numerous parameters affect the general output of a mannequin. ”This may be significantly helpful for an inner Mannequin Threat Workplace as they search to approve, validate and govern fashions throughout a company,” she mentioned.
rubicon-ml is straightforward to make use of and integrates immediately right into a consumer’s Python mannequin pipeline. It leverages present open supply tooling together with Scikit-learn for mannequin coaching; Sprint and Plotly for visualizations; and Consumption for sharing experimental outcomes. In keeping with Ranganathan, what differentiates rubicon-ml from different related instruments at the moment in the marketplace is the ability that it offers to the consumer to decide on the platform or file format.
“The open supply nature of this answer additionally units it other than competing instruments,” mentioned Nureen D’Souza, director of the Open Supply Program Workplace at Capital One. With contributions from consultants throughout the ecosystem, open supply software program creates a high-quality product that grows even stronger over time.
In keeping with D’Souza, it’s essential for Capital One to offer again to the open supply group and work collectively to enhance the software program that everybody wants. By open sourcing our options, we will make a a lot larger affect than would have ever been doable in any other case. “Plus, we all know that open supply software program growth creates higher high quality and safer code.”
As rubicon-ml continues to steadily develop, Ranganathan mentioned that Capital One is at all times trying to make enhancements to the answer. ”We’re planning to make new integrations with the most recent Python ML libraries. And we’re at all times in search of new contributions to make the answer even stronger.”
D’Souza and Ranganathan will likely be talking about rubicon-ml and different new open supply options from Capital One on the All Issues Open convention later this yr. Within the meantime, go to rubicon-ml on GitHub to learn the way the answer can standardize the mannequin growth lifecycle at your group.
