Load agent without starting an env or creating multiple workers

How severe does this issue affect your experience of using Ray?

  • High: It blocks me to complete my task.

I want to load a trained agent to use its compute_action() method. To do this, I am loading the trainer and calling trainer = tune.registry.get_trainable_cls(class_name)(config=config) with config loaded from params.pkl saved in the agent directory. However, this does several unnecessary things that I do not want to happen, such as starting local and remote workers, each one starting an environment (which is what happens during training).

How can I use the compute_action() method without all this additional overhead? Should I load the agent differently?

Hi @fedetask,

as you want to load a trained agent, you surely want to load a checkpoint (see the documentation):

agent = ppo.PPOTrainer(config=config, env=env_class)

Then you have the agent and can use it to compute actions:

action = agent.compute_action(obs)
1 Like