Hosted Tensorboard

Posted on
December 20, 2018
by
Gabe Smedresman, Software Engineer

We love using Tensorboard to analyze our runs and visualize our model structure. But when it's running locally, it can be hard to preserve that information or share it across the team. To make this easier, we're now hosting Tensorboard for all runs instrumented with W&B.

To use hosted Tensorboard, just add tensorboard=True when you call wandb.init. For example:

wandb.init(project="groovy", tensorboard=True)

As you run your training script, your events and graph will be streamed to W&B, and a new "Tensorboard" tab will appear in your run page:

And that's it! You now have access to Tensorboard for all your runs, in perpetuity, without any additional setup or configuration.

My personal favorite aspect of this project was the loading page: thanks Jamie Wong for the fluid simulation code that makes waiting for the instance to start much more enjoyable.

You can see my example run at this link. Let us know how you use this feature and if you have any ideas for how to make it better!

Try our tools →