def evaluate(self, env_fn, hparams, stochastic):
if stochastic:
policy_to_actions_lambda = lambda policy: policy.sample()
else:
policy_to_actions_lambda = lambda policy: policy.mode()
hparams.add_hparam("policy_to_actions_lambda",
policy_to_actions_lambda)
hparams.add_hparam("force_beginning_resets", False)
hparams.add_hparam("env_fn", env_fn)
hparams.add_hparam("frame_stack_size", self.frame_stack_size)
rl_trainer_lib.evaluate(hparams, self.agent_model_dir)
def evaluate_single_config(hparams, agent_model_dir):
"""Evaluate the PPO agent in the real environment."""
eval_hparams = trainer_lib.create_hparams(hparams.ppo_params)
eval_hparams.num_agents = hparams.num_agents
env = setup_env(hparams, batch_size=hparams.num_agents)
environment_spec = rl.standard_atari_env_spec(env)
eval_hparams.add_hparam("environment_spec", environment_spec)
eval_hparams.add_hparam("policy_to_actions_lambda",
hparams.policy_to_actions_lambda)
env.start_new_epoch(0)
rl_trainer_lib.evaluate(eval_hparams, agent_model_dir)
rollouts = env.current_epoch_rollouts()[:hparams.num_agents]
env.close()
assert len(rollouts) == hparams.num_agents
return tuple(
compute_mean_reward(rollouts, clipped) for clipped in (True, False))