def evaluate_performance(self, itr, all_samples, loss_before, loss_after,
kl_before, kl, policy_entropy):
"""Evaluate performance of this batch.
Args:
itr (int): Iteration number.
all_samples (list[list[MAMLTrajectoryBatch]]): Two
dimensional list of MAMLTrajectoryBatch of size
[meta_batch_size * (num_grad_updates + 1)]
loss_before (float): Loss before optimization step.
loss_after (float): Loss after optimization step.
kl_before (float): KL divergence before optimization step.
kl (float): KL divergence after optimization step.
policy_entropy (float): Policy entropy.
Returns:
float: The average return in last epoch cycle.
"""
tabular.record('Iteration', itr)
for i in range(self._num_grad_updates + 1):
all_rewards = [
path_rewards for task_samples in all_samples
for path_rewards in task_samples[i].rewards.numpy()
]
discounted_returns = [
tensor_utils.discount_cumsum(path_rewards,
self._inner_algo.discount)[0]
for path_rewards in all_rewards
]
undiscounted_returns = np.sum(all_rewards, axis=-1)
average_return = np.mean(undiscounted_returns)
with tabular.prefix('Update_{0}/'.format(i)):
tabular.record('AverageDiscountedReturn',
np.mean(discounted_returns))
tabular.record('AverageReturn', average_return)
tabular.record('StdReturn', np.std(undiscounted_returns))
tabular.record('MaxReturn', np.max(undiscounted_returns))
tabular.record('MinReturn', np.min(undiscounted_returns))
tabular.record('NumTrajs', len(all_rewards))
with tabular.prefix(self._policy.name + '/'):
tabular.record('LossBefore', loss_before)
tabular.record('LossAfter', loss_after)
tabular.record('dLoss', loss_before - loss_after)
tabular.record('KLBefore', kl_before)
tabular.record('KLAfter', kl)
tabular.record('Entropy', policy_entropy)
return average_return
def _train_once(self):
"""Perform one iteration of training."""
policy_loss_list = []
qf_loss_list = []
contrastive_loss_list = []
alpha_loss_list = []
alpha_list = []
for _ in range(self._num_steps_per_epoch):
indices = np.random.choice(range(self._num_train_tasks),
self._meta_batch_size)
policy_loss, qf_loss, contrastive_loss, alpha_loss, alpha = self._optimize_policy(
indices)
policy_loss_list.append(policy_loss)
qf_loss_list.append(qf_loss)
contrastive_loss_list.append(contrastive_loss)
alpha_loss_list.append(alpha_loss)
alpha_list.append(alpha)
with tabular.prefix('MetaTrain/Average/'):
tabular.record('PolicyLoss',
np.average(np.array(policy_loss_list)))
tabular.record('QfLoss', np.average(np.array(qf_loss_list)))
tabular.record('ContrastiveLoss',
np.average(np.array(contrastive_loss_list)))
tabular.record('AlphaLoss', np.average(np.array(alpha_loss_list)))
tabular.record('AlphaLoss', np.average(np.array(alpha_loss_list)))
tabular.record('Alpha', np.average(np.array(alpha_list)))
def evaluate(self, algo, test_rollouts_per_task=None):
"""Evaluate the Meta-RL algorithm on the test tasks.
Args:
algo (metarl.np.algos.MetaRLAlgorithm): The algorithm to evaluate.
test_rollouts_per_task (int or None): Number of rollouts per task.
"""
if test_rollouts_per_task is None:
test_rollouts_per_task = self._n_test_rollouts
adapted_trajectories = []
logger.log('Sampling for adapation and meta-testing...')
for env_up in self._test_task_sampler.sample(self._n_test_tasks):
policy = algo.get_exploration_policy()
traj = TrajectoryBatch.concatenate(*[
self._test_sampler.obtain_samples(self._eval_itr, 1, policy,
env_up)
for _ in range(self._n_exploration_traj)
])
adapted_policy = algo.adapt_policy(policy, traj)
adapted_traj = self._test_sampler.obtain_samples(
self._eval_itr, test_rollouts_per_task * self._max_path_length,
adapted_policy)
adapted_trajectories.append(adapted_traj)
logger.log('Finished meta-testing...')
with tabular.prefix(self._prefix + '/' if self._prefix else ''):
log_multitask_performance(
self._eval_itr,
TrajectoryBatch.concatenate(*adapted_trajectories),
getattr(algo, 'discount', 1.0),
task_names=self._test_task_names)
self._eval_itr += 1
def evaluate(self, algo, test_rollouts_per_task=None):
"""Evaluate the Meta-RL algorithm on the test tasks.
Args:
algo (garage.np.algos.MetaRLAlgorithm): The algorithm to evaluate.
test_rollouts_per_task (int or None): Number of rollouts per task.
"""
if test_rollouts_per_task is None:
test_rollouts_per_task = self._n_test_rollouts
adapted_trajectories = []
logger.log('Sampling for adapation and meta-testing...')
if self._test_sampler is None:
self._test_sampler = self._sampler_class.from_worker_factory(
WorkerFactory(seed=get_seed(),
max_path_length=self._max_path_length,
n_workers=1,
worker_class=self._worker_class,
worker_args=self._worker_args),
agents=algo.get_exploration_policy(),
envs=self._test_task_sampler.sample(1))
for env_up in self._test_task_sampler.sample(self._n_test_tasks):
policy = algo.get_exploration_policy()
traj = self._trajectory_batch_class.concatenate(*[
self._test_sampler.obtain_samples(self._eval_itr, 1, policy,
env_up)
for _ in range(self._n_exploration_traj)
])
adapted_policy = algo.adapt_policy(policy, traj)
adapted_traj = self._test_sampler.obtain_samples(
self._eval_itr, test_rollouts_per_task * self._max_path_length,
adapted_policy)
adapted_trajectories.append(adapted_traj)
logger.log('Finished meta-testing...')
if self._test_task_names is not None:
name_map = dict(enumerate(self._test_task_names))
else:
name_map = None
with tabular.prefix(self._prefix + '/' if self._prefix else ''):
log_multitask_performance(
self._eval_itr,
self._trajectory_batch_class.concatenate(
*adapted_trajectories),
getattr(algo, 'discount', 1.0),
trajectory_class=self._trajectory_batch_class,
name_map=name_map)
self._eval_itr += 1
if self._trajectory_batch_class == TrajectoryBatch:
rewards = self._trajectory_batch_class.concatenate(
*adapted_trajectories).rewards
else:
rewards = self._trajectory_batch_class.concatenate(
*adapted_trajectories).env_rewards
return sum(rewards) / len(rewards)
def log_multitask_performance(itr, batch, discount, name_map=None):
r"""Log performance of trajectories from multiple tasks.
Args:
itr (int): Iteration number to be logged.
batch (garage.TrajectoryBatch): Batch of trajectories. The trajectories
should have either the "task_name" or "task_id" `env_infos`. If the
"task_name" is not present, then `name_map` is required, and should
map from task id's to task names.
discount (float): Discount used in computing returns.
name_map (dict[int, str] or None): Mapping from task id's to task
names. Optional if the "task_name" environment info is present.
Note that if provided, all tasks listed in this map will be logged,
even if there are no trajectories present for them.
Returns:
numpy.ndarray: Undiscounted returns averaged across all tasks. Has
shape :math:`(N \bullet [T])`.
"""
traj_by_name = defaultdict(list)
for trajectory in batch.split():
try:
task_name = trajectory.env_infos['task_name'][0]
except KeyError:
try:
task_id = trajectory.env_infos['task_id'][0]
task_name = name_map[task_id]
except KeyError:
task_name = 'Task #{}'.format(task_id)
traj_by_name[task_name].append(trajectory)
if name_map is None:
task_names = traj_by_name.keys()
else:
task_names = name_map.values()
for task_name in task_names:
if task_name in traj_by_name:
trajectories = traj_by_name[task_name]
log_performance(itr,
garage.TrajectoryBatch.concatenate(*trajectories),
discount,
prefix=task_name)
else:
with tabular.prefix(task_name + '/'):
tabular.record('Iteration', itr)
tabular.record('NumTrajs', 0)
tabular.record('AverageDiscountedReturn', np.nan)
tabular.record('AverageReturn', np.nan)
tabular.record('StdReturn', np.nan)
tabular.record('MaxReturn', np.nan)
tabular.record('MinReturn', np.nan)
tabular.record('CompletionRate', np.nan)
tabular.record('SuccessRate', np.nan)
return log_performance(itr, batch, discount=discount, prefix='Average')
def log_multitask_performance(itr, batch, discount, name_map=None):
r"""Log performance of episodes from multiple tasks.
Args:
itr (int): Iteration number to be logged.
batch (EpisodeBatch): Batch of episodes. The episodes should have
either the "task_name" or "task_id" `env_infos`. If the "task_name"
is not present, then `name_map` is required, and should map from
task id's to task names.
discount (float): Discount used in computing returns.
name_map (dict[int, str] or None): Mapping from task id's to task
names. Optional if the "task_name" environment info is present.
Note that if provided, all tasks listed in this map will be logged,
even if there are no episodes present for them.
Returns:
numpy.ndarray: Undiscounted returns averaged across all tasks. Has
shape :math:`(N \bullet [T])`.
"""
eps_by_name = defaultdict(list)
for eps in batch.split():
task_name = '__unnamed_task__'
if 'task_name' in eps.env_infos:
task_name = eps.env_infos['task_name'][0]
elif 'task_id' in eps.env_infos:
name_map = {} if name_map is None else name_map
task_id = eps.env_infos['task_id'][0]
task_name = name_map.get(task_id, 'Task #{}'.format(task_id))
eps_by_name[task_name].append(eps)
if name_map is None:
task_names = eps_by_name.keys()
else:
task_names = name_map.values()
for task_name in task_names:
if task_name in eps_by_name:
episodes = eps_by_name[task_name]
log_performance(itr,
EpisodeBatch.concatenate(*episodes),
discount,
prefix=task_name)
else:
with tabular.prefix(task_name + '/'):
tabular.record('Iteration', itr)
tabular.record('NumEpisodes', 0)
tabular.record('AverageDiscountedReturn', np.nan)
tabular.record('AverageReturn', np.nan)
tabular.record('StdReturn', np.nan)
tabular.record('MaxReturn', np.nan)
tabular.record('MinReturn', np.nan)
tabular.record('TerminationRate', np.nan)
tabular.record('SuccessRate', np.nan)
return log_performance(itr, batch, discount=discount, prefix='Average')
def evaluate(self, algo, test_episodes_per_task=None):
"""Evaluate the Meta-RL algorithm on the test tasks.
Args:
algo (MetaRLAlgorithm): The algorithm to evaluate.
test_episodes_per_task (int or None): Number of episodes per task.
"""
if test_episodes_per_task is None:
test_episodes_per_task = self._n_test_episodes
adapted_episodes = []
logger.log('Sampling for adapation and meta-testing...')
env_updates = self._test_task_sampler.sample(self._n_test_tasks)
if self._test_sampler is None:
env = env_updates[0]()
self._max_episode_length = env.spec.max_episode_length
self._test_sampler = LocalSampler.from_worker_factory(
WorkerFactory(seed=get_seed(),
max_episode_length=self._max_episode_length,
n_workers=1,
worker_class=self._worker_class,
worker_args=self._worker_args),
agents=algo.get_exploration_policy(),
envs=env)
for env_up in env_updates:
policy = algo.get_exploration_policy()
eps = EpisodeBatch.concatenate(*[
self._test_sampler.obtain_samples(self._eval_itr, 1, policy,
env_up)
for _ in range(self._n_exploration_eps)
])
adapted_policy = algo.adapt_policy(policy, eps)
adapted_eps = self._test_sampler.obtain_samples(
self._eval_itr,
test_episodes_per_task * self._max_episode_length,
adapted_policy)
adapted_episodes.append(adapted_eps)
logger.log('Finished meta-testing...')
if self._test_task_names is not None:
name_map = dict(enumerate(self._test_task_names))
else:
name_map = None
with tabular.prefix(self._prefix + '/' if self._prefix else ''):
log_multitask_performance(
self._eval_itr,
EpisodeBatch.concatenate(*adapted_episodes),
getattr(algo, 'discount', 1.0),
name_map=name_map)
self._eval_itr += 1
def evaluate(self, algo, test_episodes_per_task=None):
"""Evaluate the Meta-RL algorithm on the test tasks.
Args:
algo (MetaRLAlgorithm): The algorithm to evaluate.
test_episodes_per_task (int or None): Number of episodes per task.
"""
if test_episodes_per_task is None:
test_episodes_per_task = self._n_test_episodes
adapted_episodes = []
logger.log('Sampling for adapation and meta-testing...')
env_updates = self._test_task_sampler.sample(self._n_test_tasks)
for env_up in env_updates:
policy = algo.get_exploration_policy()
eps = EpisodeBatch.concatenate(*[
algo._sampler.obtain_samples(self._eval_itr, 1,
policy,
env_up)
for _ in range(self._n_exploration_eps)
])
adapted_policy = algo.get_adapted_test_policy(policy, eps)
adapted_eps = algo._sampler.obtain_samples(
self._eval_itr,
test_episodes_per_task * env_up().spec.max_episode_length,
adapted_policy)
adapted_episodes.append(adapted_eps)
if self._verbose:
for ep in adapted_episodes:
print(ep.env_infos['task'][0])
print(f'last observations: {ep.last_observations}')
print('------------------------------------')
logger.log('Finished meta-testing...')
if self._test_task_names is not None:
name_map = dict(enumerate(self._test_task_names))
else:
name_map = None
with tabular.prefix(self._prefix + '/' if self._prefix else ''):
log_multitask_performance(
self._eval_itr,
EpisodeBatch.concatenate(*adapted_episodes),
getattr(algo, 'discount', 1.0),
name_map=name_map)
self._eval_itr += 1
return adapted_episodes
def _log_performance(self, itr, batch, discount, prefix='Evaluation'):
self_returns = []
env_returns = []
undiscounted_self_returns = []
undiscounted_env_returns = []
completion = []
success = []
for trajectory in batch.split():
self_returns.append(
discount_cumsum(trajectory.self_rewards, discount))
env_returns.append(
discount_cumsum(trajectory.env_rewards, discount))
undiscounted_self_returns.append(sum(trajectory.self_rewards))
undiscounted_env_returns.append(sum(trajectory.env_rewards))
completion.append(float(trajectory.terminals.any()))
if 'success' in trajectory.env_infos:
success.append(float(trajectory.env_infos['success'].any()))
average_discounted_self_return = np.mean(
[rtn[0] for rtn in self_returns])
average_discounted_env_return = np.mean(
[rtn[0] for rtn in env_returns])
with tabular.prefix(prefix + '/'):
tabular.record('Iteration', itr)
tabular.record('NumTrajs', len(self_returns))
# pseudo reward
tabular.record('AverageDiscountedSelfReturn',
average_discounted_self_return)
tabular.record('AverageSelfReturn',
np.mean(undiscounted_self_returns))
tabular.record('StdSelfReturn', np.std(undiscounted_self_returns))
tabular.record('MaxSelfReturn', np.max(undiscounted_self_returns))
tabular.record('MinSelfReturn', np.min(undiscounted_self_returns))
# env reward
tabular.record('AverageDiscountedEnvReturn',
average_discounted_env_return)
tabular.record('AverageEnvReturn',
np.mean(undiscounted_env_returns))
tabular.record('StdEnvReturn', np.std(undiscounted_env_returns))
tabular.record('MaxEnvReturn', np.max(undiscounted_env_returns))
tabular.record('MinEnvReturn', np.min(undiscounted_env_returns))
tabular.record('CompletionRate', np.mean(completion))
if success:
tabular.record('SuccessRate', np.mean(success))
return undiscounted_self_returns, undiscounted_env_returns
def log_performance(itr,
batch,
discount,
trajectory_class=TrajectoryBatch,
prefix='Evaluation'):
"""Evaluate the performance of an algorithm on a batch of trajectories.
Args:
itr (int): Iteration number.
batch (TrajectoryBatch): The trajectories to evaluate with.
discount (float): Discount value, from algorithm's property.
prefix (str): Prefix to add to all logged keys.
Returns:
numpy.ndarray: Undiscounted returns.
"""
returns = []
undiscounted_returns = []
completion = []
success = []
for trajectory in batch.split():
if trajectory_class == TrajectoryBatch:
returns.append(discount_cumsum(trajectory.rewards, discount))
undiscounted_returns.append(sum(trajectory.rewards))
else:
returns.append(discount_cumsum(trajectory.env_rewards, discount))
undiscounted_returns.append(sum(trajectory.env_rewards))
completion.append(float(trajectory.terminals.any()))
if 'success' in trajectory.env_infos:
success.append(float(trajectory.env_infos['success'].any()))
average_discounted_return = np.mean([rtn[0] for rtn in returns])
with tabular.prefix(prefix + '/'):
tabular.record('Iteration', itr)
tabular.record('NumTrajs', len(returns))
tabular.record('AverageDiscountedReturn', average_discounted_return)
tabular.record('AverageReturn', np.mean(undiscounted_returns))
tabular.record('StdReturn', np.std(undiscounted_returns))
tabular.record('MaxReturn', np.max(undiscounted_returns))
tabular.record('MinReturn', np.min(undiscounted_returns))
tabular.record('CompletionRate', np.mean(completion))
if success:
tabular.record('SuccessRate', np.mean(success))
return undiscounted_returns
def train_once(self, itr, paths):
"""Train the algorithm once.
Args:
itr (int): Iteration number.
paths (list[dict]): A list of collected paths
Returns:
dict: Processed sample data, with key
* average_return: (float)
"""
obs, actions, rewards, valids, baselines = self.process_samples(
itr, paths)
loss = self._compute_loss(itr, obs, actions, rewards, valids,
baselines)
self._old_policy.load_state_dict(self.policy.state_dict())
self._optimizer.zero_grad()
loss.backward()
kl_before = self._compute_kl_constraint(obs).detach()
self._optimize(itr, obs, actions, rewards, valids, baselines)
with torch.no_grad():
loss_after = self._compute_loss(itr, obs, actions, rewards, valids,
baselines)
kl = self._compute_kl_constraint(obs)
policy_entropy = self._compute_policy_entropy(obs)
average_returns = log_performance(itr,
TrajectoryBatch.from_trajectory_list(
self.env_spec, paths),
discount=self.discount)
with tabular.prefix(self.policy.name):
tabular.record('LossBefore', loss.item())
tabular.record('LossAfter', loss_after.item())
tabular.record('dLoss', loss.item() - loss_after.item())
tabular.record('KLBefore', kl_before.item())
tabular.record('KL', kl.item())
tabular.record('Entropy', policy_entropy.mean().item())
self.baseline.fit(paths)
return np.mean(average_returns)
def log_multitask_performance(itr,
batch,
discount,
name_map={},
task_names=None):
traj_by_name = defaultdict(list)
for trajectory in batch.split():
try:
task_name = trajectory.env_infos['task_name'][0]
except KeyError:
try:
task_id = trajectory.env_infos['task_id'][0]
task_name = name_map[task_id]
except KeyError:
task_name = 'Task #{}'.format(task_id)
traj_by_name[task_name].append(trajectory)
if task_names is None:
for (task_name, trajectories) in traj_by_name.items():
log_performance(itr,
metarl.TrajectoryBatch.concatenate(*trajectories),
discount,
prefix=task_name)
else:
for task_name in sorted(task_names):
if task_name in traj_by_name:
trajectories = traj_by_name[task_name]
log_performance(
itr,
metarl.TrajectoryBatch.concatenate(*trajectories),
discount,
prefix=task_name)
else:
with tabular.prefix(task_name + '/'):
tabular.record('Iteration', -1)
tabular.record('NumTrajs', -1)
tabular.record('AverageDiscountedReturn', -1.)
tabular.record('AverageReturn', -1)
tabular.record('StdReturn', -1)
tabular.record('MaxReturn', -1)
tabular.record('MinReturn', -1)
tabular.record('CompletionRate', -1)
tabular.record('SuccessRate', -1)
return log_performance(itr, batch, discount=discount, prefix="Average")
def log_performance(itr, batch, discount, prefix='Evaluation'):
"""Evaluate the performance of an algorithm on a batch of episodes.
Args:
itr (int): Iteration number.
batch (EpisodeBatch): The episodes to evaluate with.
discount (float): Discount value, from algorithm's property.
prefix (str): Prefix to add to all logged keys.
Returns:
numpy.ndarray: Undiscounted returns.
"""
returns = []
undiscounted_returns = []
termination = []
success = []
for eps in batch.split():
returns.append(discount_cumsum(eps.rewards, discount))
undiscounted_returns.append(sum(eps.rewards))
termination.append(
float(
any(step_type == StepType.TERMINAL
for step_type in eps.step_types)))
if 'success' in eps.env_infos:
success.append(float(eps.env_infos['success'].any()))
average_discounted_return = np.mean([rtn[0] for rtn in returns])
with tabular.prefix(prefix + '/'):
tabular.record('Iteration', itr)
tabular.record('NumEpisodes', len(returns))
tabular.record('AverageDiscountedReturn', average_discounted_return)
tabular.record('AverageReturn', np.mean(undiscounted_returns))
tabular.record('StdReturn', np.std(undiscounted_returns))
tabular.record('MaxReturn', np.max(undiscounted_returns))
tabular.record('MinReturn', np.min(undiscounted_returns))
tabular.record('TerminationRate', np.mean(termination))
if success:
tabular.record('SuccessRate', np.mean(success))
return undiscounted_returns
def _log(self, itr, paths, loss_before, loss_after, kl_before, kl,
policy_entropy):
"""Log information per iteration based on the collected paths.
Args:
itr (int): Iteration number.
paths (list[dict]): A list of collected paths
loss_before (float): Loss before optimization step.
loss_after (float): Loss after optimization step.
kl_before (float): KL divergence before optimization step.
kl (float): KL divergence after optimization step.
policy_entropy (float): Policy entropy.
Returns:
float: The average return in last epoch cycle.
"""
average_discounted_return = (np.mean(
[path['returns'][0] for path in paths]))
undiscounted_returns = [sum(path['rewards']) for path in paths]
average_return = np.mean(undiscounted_returns)
self._episode_reward_mean.extend(undiscounted_returns)
tabular.record('Iteration', itr)
tabular.record('AverageDiscountedReturn', average_discounted_return)
tabular.record('AverageReturn', average_return)
tabular.record('Extras/EpisodeRewardMean',
np.mean(self._episode_reward_mean))
tabular.record('NumTrajs', len(paths))
tabular.record('StdReturn', np.std(undiscounted_returns))
tabular.record('MaxReturn', np.max(undiscounted_returns))
tabular.record('MinReturn', np.min(undiscounted_returns))
with tabular.prefix(self.policy.name):
tabular.record('LossBefore', loss_before)
tabular.record('LossAfter', loss_after)
tabular.record('dLoss', loss_before - loss_after)
tabular.record('KLBefore', kl_before)
tabular.record('KL', kl)
tabular.record('Entropy', policy_entropy)
return average_return
def train(self, runner):
"""Obtain samplers and start actual training for each epoch.
Args:
runner (LocalRunner): LocalRunner is passed to give algorithm
the access to runner.step_epochs(), which provides services
such as snapshotting and sampler control.
"""
if not self._eval_env:
self._eval_env = runner.get_env_copy()
for epoch in runner.step_epochs():
if self._eval_env is not None:
log_performance(epoch,
obtain_evaluation_samples(
self.learner, self._eval_env),
discount=1.0)
losses = self._train_once(runner, epoch)
with tabular.prefix(self._name + '/'):
tabular.record('MeanLoss', np.mean(losses))
tabular.record('StdLoss', np.std(losses))
def _train_once(self, itr, eps):
"""Train the algorithm once.
Args:
itr (int): Iteration number.
eps (EpisodeBatch): A batch of collected paths.
Returns:
numpy.float64: Calculated mean value of undiscounted returns.
"""
obs = torch.Tensor(eps.padded_observations)
rewards = torch.Tensor(eps.padded_rewards)
returns = torch.Tensor(
np.stack([
discount_cumsum(reward, self.discount)
for reward in eps.padded_rewards
]))
valids = eps.lengths
with torch.no_grad():
baselines = self._value_function(obs)
if self._maximum_entropy:
policy_entropies = self._compute_policy_entropy(obs)
rewards += self._policy_ent_coeff * policy_entropies
obs_flat = torch.Tensor(eps.observations)
actions_flat = torch.Tensor(eps.actions)
rewards_flat = torch.Tensor(eps.rewards)
returns_flat = torch.cat(filter_valids(returns, valids))
advs_flat = self._compute_advantage(rewards, valids, baselines)
with torch.no_grad():
policy_loss_before = self._compute_loss_with_adv(
obs_flat, actions_flat, rewards_flat, advs_flat)
vf_loss_before = self._value_function.compute_loss(
obs_flat, returns_flat)
kl_before = self._compute_kl_constraint(obs)
self._train(obs_flat, actions_flat, rewards_flat, returns_flat,
advs_flat)
with torch.no_grad():
policy_loss_after = self._compute_loss_with_adv(
obs_flat, actions_flat, rewards_flat, advs_flat)
vf_loss_after = self._value_function.compute_loss(
obs_flat, returns_flat)
kl_after = self._compute_kl_constraint(obs)
policy_entropy = self._compute_policy_entropy(obs)
with tabular.prefix(self.policy.name):
tabular.record('/LossBefore', policy_loss_before.item())
tabular.record('/LossAfter', policy_loss_after.item())
tabular.record('/dLoss',
(policy_loss_before - policy_loss_after).item())
tabular.record('/KLBefore', kl_before.item())
tabular.record('/KL', kl_after.item())
tabular.record('/Entropy', policy_entropy.mean().item())
with tabular.prefix(self._value_function.name):
tabular.record('/LossBefore', vf_loss_before.item())
tabular.record('/LossAfter', vf_loss_after.item())
tabular.record('/dLoss',
vf_loss_before.item() - vf_loss_after.item())
self._old_policy.load_state_dict(self.policy.state_dict())
undiscounted_returns = log_performance(itr,
eps,
discount=self._discount)
return np.mean(undiscounted_returns)
def log_performance(self, indices, test, epoch):
"""Get average returns for specific tasks.
Args:
indices (list): List of tasks.
"""
discounted_returns = []
undiscounted_returns = []
completion = []
success = []
traj = []
for idx in indices:
eval_paths = []
for _ in range(self._num_evals):
paths = self.collect_paths(idx, test)
paths[-1]['terminals'] = paths[-1]['terminals'].squeeze()
paths[-1]['dones'] = paths[-1]['terminals']
# HalfCheetahVel env
if 'task' in paths[-1]['env_infos'].keys():
paths[-1]['env_infos']['task'] = paths[-1]['env_infos'][
'task']['velocity']
eval_paths.append(paths[-1])
discounted_returns.append(
discount_cumsum(paths[-1]['rewards'], self._discount))
undiscounted_returns.append(sum(paths[-1]['rewards']))
completion.append(float(paths[-1]['terminals'].any()))
# calculate success rate for metaworld tasks
if 'success' in paths[-1]['env_infos']:
success.append(paths[-1]['env_infos']['success'].any())
if test:
env = self.test_env[idx]()
temp_traj = TrajectoryBatch.from_trajectory_list(
env, eval_paths)
else:
env = self.env[idx]()
temp_traj = TrajectoryBatch.from_trajectory_list(
env, eval_paths)
traj.append(temp_traj)
if test:
with tabular.prefix('Test/'):
if self._test_task_names:
log_multitask_performance(
epoch,
TrajectoryBatch.concatenate(*traj),
self._discount,
task_names=self._test_task_names)
log_performance(epoch,
TrajectoryBatch.concatenate(*traj),
self._discount,
prefix='Average')
else:
with tabular.prefix('Train/'):
if self._train_task_names:
log_multitask_performance(
epoch,
TrajectoryBatch.concatenate(*traj),
self._discount,
task_names=self._train_task_names)
log_performance(epoch,
TrajectoryBatch.concatenate(*traj),
self._discount,
prefix='Average')
