I’m trying to train an agent using MAML on a meta-environment made up of five different tasks.
I’m wondering if, once the training is complete, there is a way to retrive, from the trained agent, the specialized weights obtained after the last adaptation step on each single task at a given saved iteration/checkpoint.
I know that the non-specialized weights at a given checkpoint can be easily obtained with something similar to:
agent = ray.rllib.agents.maml.MAMLTrainer(config=config, env=config["env"]) agent.restore(checkpoint) weights = agent.get_weights()
but I don’t know if there is a built-in function which instead returns the specialized weights at the N adaptation step.
Alternatively, is there a simple way to re-compute the specialized weights on the single task starting from the non-specialized ones?
I’m using RLlib library, version 1.3.0, with python 3.8.10 on a conda virtual environment.
Thank you in advance