def test_tune_metrics_availability(self):
log_info = LogInfo()
for metric in _AVAILABLE_TUNE_METRICS:
self.assertTrue(metric in log_info.data)
def evaluate(
experiment_dir,
seed,
agent_ids,
policy_classes,
checkpoint_dirs,
scenario_info,
num_episodes,
max_episode_steps,
headless,
timestep_sec,
log_dir,
eval_mode=True,
):
torch.set_num_threads(1)
# Create the agent specifications matched with their associated ID.
agent_specs = {
agent_id: make(
locator=policy_classes[agent_id],
checkpoint_dir=checkpoint_dirs[agent_id],
experiment_dir=experiment_dir,
max_episode_steps=max_episode_steps,
agent_id=agent_id,
)
for agent_id in agent_ids
}
# Create the environment with the specified agents.
env = gym.make(
"ultra.env:ultra-v0",
agent_specs=agent_specs,
scenario_info=scenario_info,
headless=headless,
timestep_sec=timestep_sec,
seed=seed,
eval_mode=eval_mode,
)
# Build each agent from its specification.
agents = {
agent_id: agent_spec.build_agent()
for agent_id, agent_spec in agent_specs.items()
}
# A dictionary to hold the evaluation data for each agent.
summary_log = {agent_id: LogInfo() for agent_id in agent_ids}
# Define an 'etag' for this experiment's data directory based off policy_classes.
# E.g. From a ["ultra.baselines.dqn:dqn-v0", "ultra.baselines.ppo:ppo-v0"]
# policy_classes list, transform it to an etag of "dqn-v0:ppo-v0".
etag = ":".join(
[policy_class.split(":")[-1] for policy_class in policy_classes])
for episode in episodes(num_episodes, etag=etag, log_dir=log_dir):
# Reset the environment and retrieve the initial observations.
observations = env.reset()
dones = {"__all__": False}
infos = None
episode.reset(mode="Evaluation")
while not dones["__all__"]:
# Get and perform the available agents' actions.
actions = {
agent_id: agents[agent_id].act(observation, explore=False)
for agent_id, observation in observations.items()
}
observations, rewards, dones, infos = env.step(actions)
# Record the data from this episode.
episode.record_step(agent_ids_to_record=infos.keys(),
infos=infos,
rewards=rewards)
episode.record_episode()
for agent_id, agent_data in episode.info[episode.active_tag].items():
for key, value in agent_data.data.items():
if not isinstance(value, (list, tuple, np.ndarray)):
summary_log[agent_id].data[key] += value
# Normalize by the number of evaluation episodes.
for agent_id, agent_data in summary_log.items():
for key, value in agent_data.data.items():
if not isinstance(value, (list, tuple, np.ndarray)):
summary_log[agent_id].data[key] /= num_episodes
env.close()
return summary_log
def evaluate(
experiment_dir,
seed,
agent_id,
policy_class,
itr_count,
checkpoint_dir,
scenario_info,
num_episodes,
headless,
timestep_sec,
):
torch.set_num_threads(1)
spec = make(
locator=policy_class,
checkpoint_dir=checkpoint_dir,
experiment_dir=experiment_dir,
)
env = gym.make(
"ultra.env:ultra-v0",
agent_specs={agent_id: spec},
scenario_info=scenario_info,
headless=headless,
timestep_sec=timestep_sec,
seed=seed,
eval_mode=True,
)
agent = spec.build_agent()
summary_log = LogInfo()
logs = []
for episode in episodes(num_episodes):
observations = env.reset()
state = observations[agent_id]
dones, infos = {"__all__": False}, None
episode.reset(mode="Evaluation")
while not dones["__all__"]:
action = agent.act(state, explore=False)
observations, rewards, dones, infos = env.step({agent_id: action})
next_state = observations[agent_id]
state = next_state
episode.record_step(agent_id=agent_id, infos=infos, rewards=rewards)
episode.record_episode()
logs.append(episode.info[episode.active_tag].data)
for key, value in episode.info[episode.active_tag].data.items():
if not isinstance(value, (list, tuple, np.ndarray)):
summary_log.data[key] += value
for key, val in summary_log.data.items():
if not isinstance(val, (list, tuple, np.ndarray)):
summary_log.data[key] /= num_episodes
env.close()
return summary_log