def _setup_actor_critic_agent(self, config: Config, load_from_ckpt: bool,
ckpt_path: str) -> None:
r"""Sets up actor critic and agent.
Args:
config: MODEL config
Returns:
None
"""
config.defrost()
config.TORCH_GPU_ID = self.config.TORCH_GPU_ID
config.freeze()
if config.CMA.use:
self.actor_critic = CMAPolicy(
observation_space=self.envs.observation_spaces[0],
action_space=self.envs.action_spaces[0],
model_config=config,
)
else:
self.actor_critic = Seq2SeqPolicy(
observation_space=self.envs.observation_spaces[0],
action_space=self.envs.action_spaces[0],
model_config=config,
)
self.actor_critic.to(self.device)
self.optimizer = torch.optim.Adam(self.actor_critic.parameters(),
lr=self.config.DAGGER.LR)
if load_from_ckpt:
ckpt_dict = self.load_checkpoint(ckpt_path, map_location="cpu")
self.actor_critic.load_state_dict(ckpt_dict["state_dict"])
logger.info(f"Loaded weights from checkpoint: {ckpt_path}")
logger.info("Finished setting up actor critic model.")
def get_defaut_config():
c = Config()
c.INPUT_TYPE = "blind"
c.MODEL_PATH = "data/checkpoints/blind.pth"
c.RL.PPO.hidden_size = 512
c.RANDOM_SEED = 7
c.TORCH_GPU_ID = 0
return c