Iterative, the MLOps company today announced the release of Data Versioning Control (DVC), introducing industry-first, experiment versioning. Experiment versioning gives developers an easy way to save, compare, and reproduce ML experiments at scale in ways that neither traditional software version control nor existing experiment tracking tools can.
“current experiment tracking tools usually provide an API to log experiment information, a database to store it, and a dashboard to compare and visualize. DVC experiment versioning builds on newer version control principles which allow to address experiment tracking needs and give developers an integrated way to iterate their experiments.”
Experiment versioning in DVC builds on modern version control principles to address experiment tracking needs and give developers a way to iterate their experiments. Experiment versioning is lightweight, using an existing tech stack eliminating the need for additional services. Automated reproduction saves time and complexity while providing confidence and audit-ability, while distributed and flexible collaboration enables any size team to generate experiments individually and share them as they choose.
With experiment versioning, data science teams can:
- Restore or reproduce any experiment automatically
- Log experiments end-to-end and track changes introduced by each
- Keep experiments connected to their Git repo, with no external services needed
With open tools and formats, Iterative is cloud-agnostic, providing flexibility and removing the need and lock-in for proprietary AI Platforms.
DVC provides users with a Git-like interface for versioning data, models, and pipelines, bringing version control to machine learning and solving the challenges of reproducibility. Experiment versioning extends DVC’s capabilities beyond simple experiment tracking.