Reusing Tensorboard's Frontend in Custom Plugins

[Markus28]

Hey,

I am interested in creating a plugin to processes several runs and visualize some metrics as a margin plot. I would like the user interface to be consistent with the rest of Tensorboard. After taking a look at the custom scalars plugin, I believe that the existing plotting frontend in Tensorboard would be suited for this job.

What is the recommended practice to reuse Tensorboard's frontend code? Should I build Tensorboard from source with my plugin? I was not able to find an answer to this question on the internet. This might be a basic question, but unfortunately, I am not completely familiar with front-end development.

I think it would be immensely helpful for plugin developers, if a minimal plotting example (or some guidelines for creating plots) could be provided, as this task is central to most plugins. I would be grateful for some pointers!

Best,
Markus

[stephanwlee]

Hi @Markus28, thank you for the question.

Unfortunately, we have no plans to provide widgets and foundational components for plugin developers. We do not want plugin authors to be locked by the library/framework choice of ours (we want you to use React, Vue.js and others if you can!) and we did not want to increase the API surface more than the our small team can provide. We will continue to provide support for our plugin APIs (e.g., https://github.com/tensorflow/tensorboard/blob/master/tensorboard/components/experimental/plugin_lib/core.ts, https://github.com/tensorflow/tensorboard/blob/master/tensorboard/components/experimental/plugin_lib/runs.ts) instead though.