Announcing Registry from Weights & Biases
We’ve made some big improvements to our Registry feature. Here’s what you need to know.
Created on July 25|Last edited on July 26
Comment
Registry is now live and it’s a game-changer for how organizations publish and share their machine learning models and datasets. Registry lets your team:
- Create multiple registries to store any type of machine learning artifact
- Share models and datasets outside of the team where they originate
- Control access with new, assignable roles per registry for users throughout the enterprise
Let’s dive into these improvements to better understand how your team can benefit. You can also see it action in the demo video below or check out the docs for more information.
Multiple registries, multiple ML artifact types
Users want to publish and share more than just models. They also want to publish and share the datasets used to train and evaluate their models. Collecting and preparing the right training data for any model building exercise is resource-intensive, whether it's internal data—such as sales data, web logs, or multimedia files—or third-party data, like as geolocation information. Datasets have enormous value as a key ingredient in the model development process and it's critically important that approved datasets are available to every individual and team.
Registry offers users the ability to publish and share models, datasets, and other artifacts necessary for model reproducibility and integration with a CI/CD pipeline. Registry also provides users with the flexibility to store artifacts in core registries available to the entire organization or to create custom registries and set up user or team-based access permissions. Core registries store only one type of artifact (models in model registry and datasets in dataset registry), but customer registries can store models, datasets, and any other artifact type.
Registry at the organization level
When training or fine-tuning models, teammates working together on an ML project need to share models, datasets, and other artifacts. But limiting artifact sharing to a single team risks building ML development silos and sacrificing wider collaboration and artifact discoverability throughout the organization. Multiple teams working on different projects benefit from access to the same models and datasets. For example, a modified churn detection model may also be relevant for predicting other future events and a website log dataset may be used for behavioral clustering or fraud analysis.
Registry is a central repository that stores models and datasets for the entire organization. ML practitioners from different teams with the right permissions can search, investigate, and use models and datasets published by other teams in their own experiments. More sharing means more collaboration, less duplication of work, and greater productivity.
Better access control
Many teams want to provision access to important artifacts across their teams. Lack of effective governance and access control can result in unexpected changes to artifacts or a model promotion process and inadequate access control can also lead to unauthorized and unintended disclosure of sensitive information.
Registry provides governance and role-based access control to ensure that the right artifacts are only available with the right level of access to the right users. Registry permits the administrator to assign each user a viewer, member, or administrator role for each custom registry. The shareability of models and datasets is balanced with the ability to protect artifacts intended only for a specific audience.
Getting started
Registry is available now and the time is right for your organization to level up your MLOps workflow. To learn more about Registry and how to get started, just head on over to our product docs.
Add a comment
Tags: Articles, W&B Features
Iterate on AI agents and models faster. Try Weights & Biases today.