Designing Streamlined Model Management Workflows at Canva Using Weights & Biases
Canva is one of the Most Innovative Companies in enterprise software, with the popular design and publishing tool boasting more than 150 million monthly active users. The company has also invested heavily in AI, with new suites of AI tools to automatically translate text in Canva designs, add new image features, remove unwanted details and much more
The focus on AI also extended behind the scenes, where Canva pioneers innovative approaches to machine learning operations to help them tackle a whole host of challenges. From generative models and recommendation systems to personalization models and search improvements, Canva is deeply invested in ML in the name of creating more delightful user experiences.
Driving a big part of this ML-driven culture and investment is Thibault Main de Boissiere, the ML Platform Team Lead at Canva. Thibault is responsible for helping Canva’s team of over 100 ML engineers train models in a reproducible way, enabling those models to be productionized smoothly, and maintaining the team’s pipelines and infrastructure.
The world-class ML team at Canva relies on Weights & Biases (W&B) for experiment tracking, as well as leaning on the Model Registry to serve as the centralized hub and single pane of glass in facilitating experimentation into deployment.
Model Management and Deployment via Model Registry
Thibault and his team wanted a clean separation between experimental models and production-ready models. Before using the W&B Model Registry, there was a lot more noise in the entire deployment workflow. It was hard to separate production training runs from experimental training runs, and the deployment logic relied on a complex combination of tags. Using the W&B Model Registry provided a clear separation between production and experimentation, while aliases made it much clearer for users to understand which model would be deployed to production or to carry out A/B testing.
With many on the team already using Weights & Biases for experiment tracking, integrating the Model Registry proved seamless and hugely impactful.
“The W&B Model Registry simplifies our lives in so many ways,” said Thibault. “It brings less noise to the user experience, as we are now only seeing models that are production-ready. It stores all the production-level information we need.”
The W&B Model Registry currently sits at the middle of the ML workflow and tech stack at Canva, that also includes leveraging Anyscale to support on-demand notebooks and distributed training, using Nix to manage dependencies, and with all production workflows deployed on Amazon Elastic Container Service (ECS). Weights & Biases now provides that single pane of glass Thibault and the Canva ML team needs to facilitate experimentation, and then ingestion for the deployment process.
And the team has even greater ambitions for how they can use Weights & Biases to simplify their ML workflow and pipelines. As Thibault describes, “W&B Model Registry can be a key way to change how we’re doing ML deployment.”
“All our MLEs and some product managers have access to Weights & Biases,” said Thibault. “We love the W&B UI, everything out of the box is really useful with very helpful system metrics, and it’s easy to manage access and security on the admin side.”