Weights & Biases (W&B)

A cloud platform for logging and comparing deep-learning training runs online.

Weights & Biases is a machine-learning experiment-management platform built by the company of the same name, acquired by the cloud-compute company CoreWeave in 2025. Adding a few lines — wandb.init and wandb.log — to training code streams loss curves, hyperparameters, GPU utilization, and sample videos to a web dashboard in real time, viewable from a phone or browser, with tools for comparing dozens of runs side by side and sharing them with collaborators. Mainstream open-source embodied-AI frameworks such as LeRobot and openpi have built-in wandb logging. When Chinese network access to it is unreliable, people commonly substitute a local TensorBoard setup or the domestic tool SwanLab.

ExampleAdding --wandb.enable=true when training a diffusion policy with LeRobot makes the loss curve and evaluation-episode videos viewable on the wandb website.

Also called

wandb

Related

TensorBoard · SwanLab · Hyperparameter · Checkpoint · LeRobot

Sources

Weights & Biases 官网 (Chinese)
W&B 文档 (Chinese)

As of

2025-05