Search Unity

  1. Welcome to the Unity Forums! Please take the time to read our Code of Conduct to familiarize yourself with the forum rules and how to post constructively.
  2. Unite 2023 Registrations are now LIVE! We are thrilled to announce that Unite 2023 Registration is now live and open for all!
    Dismiss Notice
New Forum User Notice Unite 2023 Registrations are now LIVE!

Tensorboard when using the inference engine

Discussion in 'ML-Agents' started by tobblatzius, Apr 27, 2021.

  1. tobblatzius

    tobblatzius

    Joined:
    Oct 22, 2020
    Posts:
    2
    I have trained a model and now only want to run it with inference. When selecting inference only, the mlagents-learn methods won't work. This is a problem if I want to use the handy tensorboard log for monitoring environment parameters in the inference-only simulation. These environment parameters are added to a StatsRecorder and it worked very well during the training, when the mlagents-learn command was used. Is there any easy way to modify my code when running inference to store the StatsRecorder data?
     
    tobblatzius, Apr 27, 2021
    #1

  2. vincentpierre

    vincentpierre

    Unity Technologies

    Joined:
    May 5, 2017
    Posts:
    160
    I do not see why you would want to connect to TensorBoard when running inference within the Unity Engine. One hack I can see is to use the resume training functionality but set the learning rate to 0. This way the model will no longer change and the TensorBoardStats should still be written.
    What is your use case for monitoring the TensorBoard stats when doing inference ?
     
    vincentpierre, Apr 29, 2021
    #2

  3. tobblatzius

    tobblatzius

    Joined:
    Oct 22, 2020
    Posts:
    2
    Thanks for your reply. My reason is that I monitor environment parameters, such as population sizes, as a function of step and I use pretrained agents that I don't necessarily want to keep training. Your solution, with setting learning rate to 0 is what I did and it worked fine but the simulations might have been faster if it was possible to run on inference only mode.
     
    tobblatzius, May 17, 2021
    #3

  4. ervteng_unity

    ervteng_unity

    Unity Technologies

    Joined:
    Dec 6, 2018
    Posts:
    150
    The best way to do this is to use --resume and --inference, where the trainer won't train at all but should still record to tensorboard.

    You could further speed it up by setting your batch and buffer size to something super small. Taking it a step further you could also make a "fake" agent that has very small obs/action space and set the real agents to heuristic.

    Otherwise, it's probably easier to write monitoring code inside Unity than to pass the data to Python for write out into Tensorboard.
     
    Last edited: May 18, 2021
    ervteng_unity, May 18, 2021
    #4