def log_diagnostics(self, paths, **kwargs):
list_of_rewards, terminals, obs, actions, next_obs = split_paths(paths)
returns = []
for rewards in list_of_rewards:
returns.append(np.sum(rewards))
statistics = OrderedDict()
statistics.update(
create_stats_ordered_dict(
'Undiscounted Returns',
returns,
))
statistics.update(
create_stats_ordered_dict(
'Rewards',
list_of_rewards,
))
statistics.update(create_stats_ordered_dict(
'Actions',
actions,
))
fraction_of_time_on_platform = [o[1] for o in obs]
statistics['Fraction of time on platform'] = np.mean(
fraction_of_time_on_platform)
for key, value in statistics.items():
logger.record_tabular(key, value)
return returns
def log_diagnostics(self, paths, **kwargs):
list_of_rewards, terminals, obs, actions, next_obs = split_paths(paths)
returns = []
for rewards in list_of_rewards:
returns.append(np.sum(rewards))
last_statistics = OrderedDict()
last_statistics.update(
create_stats_ordered_dict(
'UndiscountedReturns',
returns,
))
last_statistics.update(
create_stats_ordered_dict(
'Rewards',
list_of_rewards,
))
last_statistics.update(create_stats_ordered_dict(
'Actions',
actions,
))
for key, value in last_statistics.items():
logger.record_tabular(key, value)
return returns
def _statistics_from_paths(self, paths, stat_prefix):
rewards, terminals, obs, actions, next_obs = split_paths(paths)
np_batch = dict(
rewards=rewards,
terminals=terminals,
observations=obs,
actions=actions,
next_observations=next_obs,
)
batch = np_to_pytorch_batch(np_batch)
statistics = self._statistics_from_batch(batch, stat_prefix)
statistics.update(create_stats_ordered_dict(
'Num Paths', len(paths), stat_prefix=stat_prefix
))
return statistics
def test_split_paths(self):
paths = [create_path(0), create_path(1)]
rewards, terminals, obs, actions, next_obs = split_paths(paths)
expected_rewards = np.array([-1, 0, 1, 0, 1, 2]).reshape(-1, 1)
expected_terminals = np.array([0, 0, 1, 0, 0, 1]).reshape(-1, 1)
expected_obs = np.array([
[2], [4], [8],
[3], [5], [9],
])
expected_actions = np.array([
[5], [7], [9],
[6], [8], [10],
])
expected_next_obs = np.array([
[4], [5], [9],
[5], [6], [10],
])
self.assertNpEqual(rewards, expected_rewards)
self.assertNpEqual(terminals, expected_terminals)
self.assertNpEqual(obs, expected_obs)
self.assertNpEqual(actions, expected_actions)
self.assertNpEqual(next_obs, expected_next_obs)