Search Unity

Question Use pre-trained CNNs in model architecture

Discussion in 'ML-Agents' started by kpalko, Sep 18, 2020.

  1. kpalko

    kpalko

    Joined:
    May 14, 2020
    Posts:
    5
    Good afternoon everyone.

    I am wondering if anyone has any ideas about implementing a pre-trained CNN (like AlexNet or ResNet) and building that into the model architecture. As far as I know, the CNNs that are included in the brain are not pre-trained and have to learn on the visuals in the environment.
     
    kpalko, Sep 18, 2020
    #1

  2. ruoping_unity

    ruoping_unity

    Unity Technologies

    Joined:
    Jul 10, 2020
    Posts:
    134
    Yes currently the model in ML-Agents are trained from scratch. We do have ResNet as an option to the visual encoder architecture but it is randomly initialized.

    It's unclear if those pre-trained CNN trained on general image dataset would be very helpful in this case.

    If you have a specific use case and find this particularly helpful, you can submit a feature request on our Github page to let us know.
     
    ruoping_unity, Sep 24, 2020
    #2
    kpalko likes this.