Search Unity

Resolved Train traceback of mlagents

Discussion in 'ML-Agents' started by PxZzYYDS, Apr 4, 2021.

  1. PxZzYYDS

    PxZzYYDS

    Joined:
    Mar 26, 2021
    Posts:
    13
    QQ图片20210404191234.png i change the parameters in GridWorld in trainer_config.yaml
    When I start the training from the Anaconda prompt, everything looks fine at first but then, I get a BrokenPipeError and the training stops. I searched and couldn't find any solution.
     
    PxZzYYDS, Apr 4, 2021
    #1

  2. ruoping_unity

    ruoping_unity

    Unity Technologies

    Joined:
    Jul 10, 2020
    Posts:
    134
    Hi, since I saw this post labeled as "resolved", have you already solved this problem?
     
    ruoping_unity, Apr 5, 2021
    #2

  3. PxZzYYDS

    PxZzYYDS

    Joined:
    Mar 26, 2021
    Posts:
    13
    sorry
    this problem haven't been solved.
    i can not find the measure to deal with,also i'm waiting
     
    PxZzYYDS, Apr 5, 2021
    #3

  4. ruoping_unity

    ruoping_unity

    Unity Technologies

    Joined:
    Jul 10, 2020
    Posts:
    134
    The error message indicates that the connection between python and Unity is interrupted.
    Does it happen consistently? Are you able to reproduce the error with any of our example environments?
    Could you also share the exact mlagents-learn command you used to run this training?

    Also can you check the Player log (Player-X.log) in the results folder and see if there's any errors?

    The "GetOverlappedResult" error looks like the connection is probably connected to one or more of the instances crashing. Did you try rebooting your machine?
     
    ruoping_unity, Apr 5, 2021
    #4

  5. PxZzYYDS

    PxZzYYDS

    Joined:
    Mar 26, 2021
    Posts:
    13
    i use '"mlagentslearn config/trainer_config.yaml --run-id=GridWorld --train"
    i reset the file "trainer_config.yaml",now that is ok
    i think that maybe the changes of parameters of GridWorld in traienr_config.yaml lead to this error
     
    PxZzYYDS, Apr 5, 2021
    #5

  6. ruoping_unity

    ruoping_unity

    Unity Technologies

    Joined:
    Jul 10, 2020
    Posts:
    134
    Could you share what exactly did you change in the config file?
     
    ruoping_unity, Apr 5, 2021
    #6

  7. PxZzYYDS

    PxZzYYDS

    Joined:
    Mar 26, 2021
    Posts:
    13
    of course
    GridWorld:
    batch_size: 32
    normalize: false
    num_layers: 3
    hidden_units: 512
    beta: 5.0e-3
    buffer_size: 256
    max_steps: 5000000
    summary_freq: 20000
    time_horizon: 5
    reward_signals:
    extrinsic:
    strength: 1.0
    gamma: 0.9

    but I trained two million times ，and i set if eat goal ,agent will get 3points,if ear pit it will get -2 points.Also each step will get -0.001.The results are as follows：
     

    Attached Files:

    PxZzYYDS, Apr 5, 2021
    #7

  8. PxZzYYDS

    PxZzYYDS

    Joined:
    Mar 26, 2021
    Posts:
    13
    and if i use :

    GridWorld:
    batch_size: 256
    normalize: false
    num_layers: 5
    hidden_units: 512
    beta: 5.0e-3
    buffer_size: 256000
    max_steps: 5000000
    summary_freq: 10000
    time_horizon: 5
    reward_signals:
    extrinsic:
    strength: 1.0
    gamma: 0.9
    it will cause error
     
    PxZzYYDS, Apr 5, 2021
    #8