Skip to content

Usage Guide

Using it as a standalone wrapper for ML-Agents

In this mode, users should interact with wandb-mlagents-learn, which acts as a wrapper passthrough for mlagents-learn.

To use WandB with ML-Agents, simply add an extra "wandb" field in a config file that is otherwise completely compatible with ML-Agents Python package. To run, enter wandb-mlagents-learn in the terminal. This wrapper accepts the same cli arguments as mlagents-learn.

Note: In this use case, wandb-mlagents-learn does not support cli arguments. In other words, to use this wrapper, WandB's configurations must be present in a mlagents-learn compatible config file. The syntax for the various configs can be found below.

Using it with realm-tune

In this mode, users interact solely through realm-tune, as they have been tightly integrated.

To use WandB with realm-tune, similar to previously, simply add an extra "wandb" field to the realm-tune config file. An example of which can be found here.

Comparing to using wandb-mlagents-learn as a standalone, using it within realm-tune allows cli arguments to be passed (more information can be found in realm-tune's config docs).

wandb-mlagents-learn configs

In the realm-tune or mlagents-learn config file (it should be a *.yaml file), add the following: 

wandb:
    project: <default: realm_ai>
    entity: <default: None>
    offline: <default: False>
    group: <default: None>
    jobtype: <default: None> # "job_type" is accepted too!

Notes

  • Resuming a run on wandb does not work (yet). Since we use the run_id as the name of the wandb run, we will see multiple runs with the same name when resuming a run
  • If we intend to use mlagents default parameters, it is okay to pass in a config file that solely contains wandb config (as shown above)
  • If "wandb" is not defined in the config file, wandb will not be used, and functionality will then be identical to mlagents!
  • When resuming runs, multiple tensorboard files may be present. Currently, this wrapper uses the latest tensorboard file, based on time of creation (Windows)/last modification (Unix)