TensorBoard?

[desaibhargav]

I just wanted to get people's thoughts on whether TensorBoard integration would be a good thing to have (or even make sense). TensorBoard is basically a real-time dashboard for (pretty much) all things related to model training and I've personally found it useful for debugging.

An example of what this looks like: https://huggingface.co/bigscience/bloom/tensorboard

A central dashboard for monitoring and logging experiments feels like a good to have for the number of experiments we run. It IS a little finicky to set up, but I can take initiative to get it working for us if everyone is onboard with the idea.

Some use-cases (other than the obvious ones) that come to mind are:

• Adding short videos of agents while the training is in progress. It may be useful debugging such inspecting the videos with the reward signals side-by-side.
• Since it's all real-time, it may be useful to pull plugs on experiments if necessary and save time. Plus, a lot of things can be logged and monitored (such as @Zach-Attach's very cool density map of the area covered by the agents). It can log whatever we want, audio, video, text or scalars.

What do we think?

[desaibhargav]

The idea of a TV monitor with live images / videos / high-level conceptual metrics from various experiments is growing on me 😵‍💫