Run tune.Tuner with a given policy

RLlib
1

Hi,

I would like to set up a tune.Tuner object and run tuner.fit() for an experiment using a Policy object restored from a checkpoint. Could you please guide me on how to achieve this?

To clarify, here’s what I’m currently doing:

from ray.rllib.policy import Policy
policy = Policy.from_checkpoint(<path>)['default_policy']
weights = policy.get_weights()
weights = {'default_policy': weights}

from ray.tune.registry import get_trainable_cls
config = (
            get_trainable_cls("PPO")
            .get_default_config()
            .environment(..., env_config=env_config)
            .training(..., model={custom_model_config=custom_model_config})
            )
algo = config.build()
algo.set_weights(weights)

At this point, I can call algo.train(), but how can I achieve something similar using a tuner?

Specifically, I’m interested in configuring a new environment (a new env_config) and passing on a custom model config (a new custom_model_config) in this setup.

Thank you in advance for your help!