Ml-agents: Accessing activations from the agent's brain?

[MarkTension]

Is your feature request related to a problem? Please describe.
For my project I'd like to visualize some of the raw activations of the agent's brain.
For my own case, I'm mostly interested in obtaining the distributions of the action probabilities during runtime for artistic purposes.
However, in other networks that I've trained I've used this functionality a lot for debugging as well. e.g. checking the activation pattern of the encoder's first layer to see if anything sensical is learned

Describe the solution you'd like
Would love to see a c# interface for accessing the brain's specific activations, so they can be experimented with.

Describe alternatives you've considered
I was told that this can theoretically already be done via barracuda but that this formal request might mark it for evaluation.

Getting the raw activations to build this kind of insight:

[surfnerd]

Hi @MarkTension,
Thanks for making a formal request. This feature request has been brought up internally before. Hopefully, we can get to it sooner rather than later. We have logged this request internally as MLA-812. All updates on this request will be posted on this thread.

[MarkTension]

Just checking: any updates in the meantime on this? Or work in progress to try?

[slantBoy]

@surfnerd I was following this thread for a while now and was wondering what the status on the topic is. It would be really helpful to get access to the agent's brain to use it for artistic interpretation.