Regarding inferencing the learnt policy about marllib HOT 4 CLOSED

It is doable. However, MARLlib decides not to incorporate the loading and rendering functions as we find it hard to unify all ten environments to render in a similar pattern.

We would like to provide you with instructions on how to implement this.
You can find a example for rendering here.
Also, the way to load the checkpoint is adding a restore path in tune.run(restore=YourCheckPointPath)
The complete configuration can be found in Trainer.

There is a thorough solution provided by Sven: multiagent-load-only-one-policy-from-checkpoint.

Any further question is welcome. We are happy to help you out.

from marllib.

Hi Thanks for the previous answer. But we have been unable to use the learned policy to compute actions for our agents. Our objective is to compute agent's actions based on the learnt policy. But when we use the function agent.compute_single_action(obs), where obs = env.reset()
We get the error that seq_lens in None type. We are not able to find where to find the sequence lengths to resolve this error. We added some print statements in the training and could observe the sequence lengths being printed there but compute_single_action does not seem to be workking with that.

It would be really helpful if you could provide some insights for the same. Also kindly let us know if we should be using some function other than agent.compute_single_action for the same purpose.

from marllib.

Hi SiddSS,

Sorry for the late reply. Check out mamujoco example and mpe example for loading the checkpoint and rendering the environment in MARLlib.
You are welcomed for any further question.

Siyi

from marllib.

New APIs are here for guiding how to render the pretrained model.

from marllib.