Data is at the core of every ML workflow. In this talk, Chris discusses our powerful latest tool Tables and new features in W&B Artifacts to let you visualize and query datasets and model evaluations at the example level. You can use this new tool together with Ray to analyze and understand your datasets, and to measure and debug model performance.
Our goal is to give you highly scalable, flexible and configurable tools, with rich out-of-the-box visualizations available for common tasks. The system is constructed out of:
The ability to save large wandb.Table objects, optionally containing rich media (like images with bounding boxes), inside of W&B Artifacts.
Support for cross-artifact file references, and the ability to join tables together in the UI. This is used, for example, to log a set of bounding box predictions against a ground-truth dataset artifact, without duplicating the source images and labels.
An all new “typed, run-time-swappable UI-panel architecture”. This is what powers the rich visualizations and charts you see as you compare and group your data tables. Eventually we’ll open this up, so users can add completely custom visualizers that work everywhere in the W&B UI.
Use W&B Tables to log and visualize data and model predictions. Interactively explore your data by comparing changes precisely across models, epochs, or individual examples & understanding higher-level patterns in your data & capturing and communicating your insights with visual samples