Ml-agents: Assigning custom c# variables to summary in tensorboard

[MarkTension]

For my research I'd like to track specific variables in my c# scripts via tensorboard summaries.
In Tensorflow it's quite easy to do from python, but currently it seems more like a hassle to program.

Would love to see added to the ml-agents C# API a function to include any c# float / double to tensorboard summaries.

Cheers!

[unityjeffrey]

hi @MarkTension - i can see how this be very useful. its not currently in our backlog, would you be interested to contribute? we'd be happy to get your thoughts on design and code. CC: @ervteng

[MarkTension]

Great! I'd be happy to contribute. Let me know how I can help

[chriselion]

@vincentpierre possible candidate for SideChannels

[unityjeffrey]

hi @MarkTension - as Chris mentioned, we are working on a feature called SideChannels which is meant to pass data between Unity and Python outside of the training loops. Here is a few of the PRs and the latest code is in our develop branch. Let us know if that would work for this request:

https://github.com/Unity-Technologies/ml-agents/pull/2956
Docs - https://github.com/Unity-Technologies/ml-agents/pull/3011

[MarkTension]

Any updates on this in the meantime? JustChecking^^

[ervteng]

Hi @MarkTension, side channels were released with ML-Agents v0.13. There's no official support yet for moving arbitrary data from C# to Python, but you can see some example implementations of Side Channels here.

[chriselion]

I've got this logged in our backlog as MLA-675; I'm going to write up a small proposal for it today.

[chriselion]

This was implemented in https://github.com/Unity-Technologies/ml-agents/pull/3660/files - it's available on the master branch now, and will be in the next release.