def __init__(self,
config,
policy_params,
env_creator,
noise,
monitor_path,
min_task_runtime=0.2):
self.min_task_runtime = min_task_runtime
self.config = config
self.policy_params = policy_params
self.noise = SharedNoiseTable(noise)
self.monitor_path = monitor_path
self.env = env_creator(config["env_config"])
self.memory = []
if is_atari(self.env):
self.env = wrap_deepmind(self.env, dim=84, framestack=4)
if np.random.uniform() < 0.5:
if self.monitor_path:
self.env = _monitor(self.env, self.monitor_path)
else:
self.env = env_creator(config["env_config"])
if is_atari(self.env):
self.env = wrap_deepmind(self.env, dim=84, framestack=4)
self.env2 = env_creator(config["env_config"])
if is_atari(self.env2):
self.env2 = wrap_deepmind(self.env2, dim=84, framestack=4)
from ray.rllib import models
self.preprocessor = models.ModelCatalog.get_preprocessor(
self.env, config["model"])
self.sess = utils.make_session(single_threaded=True)
self.policy = policies.GenericPolicy(self.sess, self.env, self.env2,
self.env.action_space,
self.env.observation_space,
self.preprocessor,
config["observation_filter"],
config["model"], **policy_params)
#self.sess.run(tf.global_variables_initializer())
#self.sess1 = utils.make_session(single_threaded=True)
# self.policymax = policies.GenericPolicy(
# self.sess, self.env,self.env2,self.env.action_space, self.env.observation_space,
# self.preprocessor, config["observation_filter"], config["model"],
# **policy_params)
self.sess.run(tf.global_variables_initializer())
def __init__(self,
env_creator,
policy,
policy_mapping_fn=None,
policies_to_train=None,
tf_session_creator=None,
rollout_fragment_length=100,
batch_mode="truncate_episodes",
episode_horizon=None,
preprocessor_pref="deepmind",
sample_async=False,
compress_observations=False,
num_envs=1,
observation_fn=None,
observation_filter="NoFilter",
clip_rewards=None,
clip_actions=True,
env_config=None,
model_config=None,
policy_config=None,
worker_index=0,
num_workers=0,
monitor_path=None,
log_dir=None,
log_level=None,
callbacks=None,
input_creator=lambda ioctx: ioctx.default_sampler_input(),
input_evaluation=frozenset([]),
output_creator=lambda ioctx: NoopOutput(),
remote_worker_envs=False,
remote_env_batch_wait_ms=0,
soft_horizon=False,
no_done_at_end=False,
seed=None,
extra_python_environs=None,
fake_sampler=False):
"""Initialize a rollout worker.
Arguments:
env_creator (func): Function that returns a gym.Env given an
EnvContext wrapped configuration.
policy (class|dict): Either a class implementing
Policy, or a dictionary of policy id strings to
(Policy, obs_space, action_space, config) tuples. If a
dict is specified, then we are in multi-agent mode and a
policy_mapping_fn should also be set.
policy_mapping_fn (func): A function that maps agent ids to
policy ids in multi-agent mode. This function will be called
each time a new agent appears in an episode, to bind that agent
to a policy for the duration of the episode.
policies_to_train (list): Optional whitelist of policies to train,
or None for all policies.
tf_session_creator (func): A function that returns a TF session.
This is optional and only useful with TFPolicy.
rollout_fragment_length (int): The target number of env transitions
to include in each sample batch returned from this worker.
batch_mode (str): One of the following batch modes:
"truncate_episodes": Each call to sample() will return a batch
of at most `rollout_fragment_length * num_envs` in size.
The batch will be exactly
`rollout_fragment_length * num_envs` in size if
postprocessing does not change batch sizes. Episodes may be
truncated in order to meet this size requirement.
"complete_episodes": Each call to sample() will return a batch
of at least `rollout_fragment_length * num_envs` in size.
Episodes will not be truncated, but multiple episodes may
be packed within one batch to meet the batch size. Note
that when `num_envs > 1`, episode steps will be buffered
until the episode completes, and hence batches may contain
significant amounts of off-policy data.
episode_horizon (int): Whether to stop episodes at this horizon.
preprocessor_pref (str): Whether to prefer RLlib preprocessors
("rllib") or deepmind ("deepmind") when applicable.
sample_async (bool): Whether to compute samples asynchronously in
the background, which improves throughput but can cause samples
to be slightly off-policy.
compress_observations (bool): If true, compress the observations.
They can be decompressed with rllib/utils/compression.
num_envs (int): If more than one, will create multiple envs
and vectorize the computation of actions. This has no effect if
if the env already implements VectorEnv.
observation_fn (ObservationFunction): Optional multi-agent
observation function.
observation_filter (str): Name of observation filter to use.
clip_rewards (bool): Whether to clip rewards to [-1, 1] prior to
experience postprocessing. Setting to None means clip for Atari
only.
clip_actions (bool): Whether to clip action values to the range
specified by the policy action space.
env_config (dict): Config to pass to the env creator.
model_config (dict): Config to use when creating the policy model.
policy_config (dict): Config to pass to the policy. In the
multi-agent case, this config will be merged with the
per-policy configs specified by `policy`.
worker_index (int): For remote workers, this should be set to a
non-zero and unique value. This index is passed to created envs
through EnvContext so that envs can be configured per worker.
num_workers (int): For remote workers, how many workers altogether
have been created?
monitor_path (str): Write out episode stats and videos to this
directory if specified.
log_dir (str): Directory where logs can be placed.
log_level (str): Set the root log level on creation.
callbacks (DefaultCallbacks): Custom training callbacks.
input_creator (func): Function that returns an InputReader object
for loading previous generated experiences.
input_evaluation (list): How to evaluate the policy performance.
This only makes sense to set when the input is reading offline
data. The possible values include:
- "is": the step-wise importance sampling estimator.
- "wis": the weighted step-wise is estimator.
- "simulation": run the environment in the background, but
use this data for evaluation only and never for learning.
output_creator (func): Function that returns an OutputWriter object
for saving generated experiences.
remote_worker_envs (bool): If using num_envs > 1, whether to create
those new envs in remote processes instead of in the current
process. This adds overheads, but can make sense if your envs
remote_env_batch_wait_ms (float): Timeout that remote workers
are waiting when polling environments. 0 (continue when at
least one env is ready) is a reasonable default, but optimal
value could be obtained by measuring your environment
step / reset and model inference perf.
soft_horizon (bool): Calculate rewards but don't reset the
environment when the horizon is hit.
no_done_at_end (bool): Ignore the done=True at the end of the
episode and instead record done=False.
seed (int): Set the seed of both np and tf to this value to
to ensure each remote worker has unique exploration behavior.
extra_python_environs (dict): Extra python environments need to
be set.
fake_sampler (bool): Use a fake (inf speed) sampler for testing.
"""
self._original_kwargs = locals().copy()
del self._original_kwargs["self"]
global _global_worker
_global_worker = self
# set extra environs first
if extra_python_environs:
for key, value in extra_python_environs.items():
os.environ[key] = str(value)
def gen_rollouts():
while True:
yield self.sample()
ParallelIteratorWorker.__init__(self, gen_rollouts, False)
policy_config = policy_config or {}
if (tf and policy_config.get("eager")
and not policy_config.get("no_eager_on_workers")
# This eager check is necessary for certain all-framework tests
# that use tf's eager_mode() context generator.
and not tf.executing_eagerly()):
tf.enable_eager_execution()
if log_level:
logging.getLogger("ray.rllib").setLevel(log_level)
if worker_index > 1:
disable_log_once_globally() # only need 1 worker to log
elif log_level == "DEBUG":
enable_periodic_logging()
env_context = EnvContext(env_config or {}, worker_index)
self.policy_config = policy_config
if callbacks:
self.callbacks = callbacks()
else:
from ray.rllib.agents.callbacks import DefaultCallbacks
self.callbacks = DefaultCallbacks()
self.worker_index = worker_index
self.num_workers = num_workers
model_config = model_config or {}
policy_mapping_fn = (policy_mapping_fn
or (lambda agent_id: DEFAULT_POLICY_ID))
if not callable(policy_mapping_fn):
raise ValueError("Policy mapping function not callable?")
self.env_creator = env_creator
self.rollout_fragment_length = rollout_fragment_length * num_envs
self.batch_mode = batch_mode
self.compress_observations = compress_observations
self.preprocessing_enabled = True
self.last_batch = None
self.global_vars = None
self.fake_sampler = fake_sampler
self.env = _validate_env(env_creator(env_context))
if isinstance(self.env, MultiAgentEnv) or \
isinstance(self.env, BaseEnv):
def wrap(env):
return env # we can't auto-wrap these env types
elif is_atari(self.env) and \
not model_config.get("custom_preprocessor") and \
preprocessor_pref == "deepmind":
# Deepmind wrappers already handle all preprocessing
self.preprocessing_enabled = False
if clip_rewards is None:
clip_rewards = True
def wrap(env):
env = wrap_deepmind(
env,
dim=model_config.get("dim"),
framestack=model_config.get("framestack"))
if monitor_path:
from gym import wrappers
env = wrappers.Monitor(env, monitor_path, resume=True)
return env
else:
def wrap(env):
if monitor_path:
from gym import wrappers
env = wrappers.Monitor(env, monitor_path, resume=True)
return env
self.env = wrap(self.env)
def make_env(vector_index):
return wrap(
env_creator(
env_context.copy_with_overrides(
vector_index=vector_index, remote=remote_worker_envs)))
self.tf_sess = None
policy_dict = _validate_and_canonicalize(policy, self.env)
self.policies_to_train = policies_to_train or list(policy_dict.keys())
# set numpy and python seed
if seed is not None:
np.random.seed(seed)
random.seed(seed)
if not hasattr(self.env, "seed"):
raise ValueError("Env doesn't support env.seed(): {}".format(
self.env))
self.env.seed(seed)
try:
assert torch is not None
torch.manual_seed(seed)
except AssertionError:
logger.info("Could not seed torch")
if _has_tensorflow_graph(policy_dict) and not (tf and
tf.executing_eagerly()):
if not tf:
raise ImportError("Could not import tensorflow")
with tf.Graph().as_default():
if tf_session_creator:
self.tf_sess = tf_session_creator()
else:
self.tf_sess = tf.Session(
config=tf.ConfigProto(
gpu_options=tf.GPUOptions(allow_growth=True)))
with self.tf_sess.as_default():
# set graph-level seed
if seed is not None:
tf.set_random_seed(seed)
self.policy_map, self.preprocessors = \
self._build_policy_map(policy_dict, policy_config)
if (ray.is_initialized()
and ray.worker._mode() != ray.worker.LOCAL_MODE):
if not ray.get_gpu_ids():
logger.debug(
"Creating policy evaluation worker {}".format(
worker_index) +
" on CPU (please ignore any CUDA init errors)")
elif not tf.test.is_gpu_available():
raise RuntimeError(
"GPUs were assigned to this worker by Ray, but "
"TensorFlow reports GPU acceleration is disabled. "
"This could be due to a bad CUDA or TF installation.")
else:
self.policy_map, self.preprocessors = self._build_policy_map(
policy_dict, policy_config)
self.multiagent = set(self.policy_map.keys()) != {DEFAULT_POLICY_ID}
if self.multiagent:
if not ((isinstance(self.env, MultiAgentEnv)
or isinstance(self.env, ExternalMultiAgentEnv))
or isinstance(self.env, BaseEnv)):
raise ValueError(
"Have multiple policies {}, but the env ".format(
self.policy_map) +
"{} is not a subclass of BaseEnv, MultiAgentEnv or "
"ExternalMultiAgentEnv?".format(self.env))
self.filters = {
policy_id: get_filter(observation_filter,
policy.observation_space.shape)
for (policy_id, policy) in self.policy_map.items()
}
if self.worker_index == 0:
logger.info("Built filter map: {}".format(self.filters))
# Always use vector env for consistency even if num_envs = 1
self.async_env = BaseEnv.to_base_env(
self.env,
make_env=make_env,
num_envs=num_envs,
remote_envs=remote_worker_envs,
remote_env_batch_wait_ms=remote_env_batch_wait_ms)
self.num_envs = num_envs
if self.batch_mode == "truncate_episodes":
pack_episodes = True
elif self.batch_mode == "complete_episodes":
rollout_fragment_length = float("inf") # never cut episodes
pack_episodes = False # sampler will return 1 episode per poll
else:
raise ValueError("Unsupported batch mode: {}".format(
self.batch_mode))
self.io_context = IOContext(log_dir, policy_config, worker_index, self)
self.reward_estimators = []
for method in input_evaluation:
if method == "simulation":
logger.warning(
"Requested 'simulation' input evaluation method: "
"will discard all sampler outputs and keep only metrics.")
sample_async = True
elif method == "is":
ise = ImportanceSamplingEstimator.create(self.io_context)
self.reward_estimators.append(ise)
elif method == "wis":
wise = WeightedImportanceSamplingEstimator.create(
self.io_context)
self.reward_estimators.append(wise)
else:
raise ValueError(
"Unknown evaluation method: {}".format(method))
if sample_async:
self.sampler = AsyncSampler(
self,
self.async_env,
self.policy_map,
policy_mapping_fn,
self.preprocessors,
self.filters,
clip_rewards,
rollout_fragment_length,
self.callbacks,
horizon=episode_horizon,
pack=pack_episodes,
tf_sess=self.tf_sess,
clip_actions=clip_actions,
blackhole_outputs="simulation" in input_evaluation,
soft_horizon=soft_horizon,
no_done_at_end=no_done_at_end,
observation_fn=observation_fn)
self.sampler.start()
else:
self.sampler = SyncSampler(
self,
self.async_env,
self.policy_map,
policy_mapping_fn,
self.preprocessors,
self.filters,
clip_rewards,
rollout_fragment_length,
self.callbacks,
horizon=episode_horizon,
pack=pack_episodes,
tf_sess=self.tf_sess,
clip_actions=clip_actions,
soft_horizon=soft_horizon,
no_done_at_end=no_done_at_end,
observation_fn=observation_fn)
self.input_reader = input_creator(self.io_context)
assert isinstance(self.input_reader, InputReader), self.input_reader
self.output_writer = output_creator(self.io_context)
assert isinstance(self.output_writer, OutputWriter), self.output_writer
logger.debug(
"Created rollout worker with env {} ({}), policies {}".format(
self.async_env, self.env, self.policy_map))
def __init__(
self,
*,
env_creator: Callable[[EnvContext], EnvType],
validate_env: Optional[Callable[[EnvType, EnvContext],
None]] = None,
policy_spec: Union[type, Dict[
str, Tuple[Optional[type], gym.Space, gym.Space,
PartialTrainerConfigDict]]] = None,
policy_mapping_fn: Optional[Callable[[AgentID], PolicyID]] = None,
policies_to_train: Optional[List[PolicyID]] = None,
tf_session_creator: Optional[Callable[[], "tf1.Session"]] = None,
rollout_fragment_length: int = 100,
batch_mode: str = "truncate_episodes",
episode_horizon: int = None,
preprocessor_pref: str = "deepmind",
sample_async: bool = False,
compress_observations: bool = False,
num_envs: int = 1,
observation_fn: "ObservationFunction" = None,
observation_filter: str = "NoFilter",
clip_rewards: bool = None,
clip_actions: bool = True,
env_config: EnvConfigDict = None,
model_config: ModelConfigDict = None,
policy_config: TrainerConfigDict = None,
worker_index: int = 0,
num_workers: int = 0,
monitor_path: str = None,
log_dir: str = None,
log_level: str = None,
callbacks: Type["DefaultCallbacks"] = None,
input_creator: Callable[[
IOContext
], InputReader] = lambda ioctx: ioctx.default_sampler_input(),
input_evaluation: List[str] = frozenset([]),
output_creator: Callable[
[IOContext], OutputWriter] = lambda ioctx: NoopOutput(),
remote_worker_envs: bool = False,
remote_env_batch_wait_ms: int = 0,
soft_horizon: bool = False,
no_done_at_end: bool = False,
seed: int = None,
extra_python_environs: dict = None,
fake_sampler: bool = False,
spaces: Optional[Dict[PolicyID, Tuple[gym.spaces.Space,
gym.spaces.Space]]] = None,
policy: Union[type, Dict[
str, Tuple[Optional[type], gym.Space, gym.Space,
PartialTrainerConfigDict]]] = None,
):
"""Initialize a rollout worker.
Args:
env_creator (Callable[[EnvContext], EnvType]): Function that
returns a gym.Env given an EnvContext wrapped configuration.
validate_env (Optional[Callable[[EnvType, EnvContext], None]]):
Optional callable to validate the generated environment (only
on worker=0).
policy_spec (Union[type, Dict[str, Tuple[Type[Policy], gym.Space,
gym.Space, PartialTrainerConfigDict]]]): Either a Policy class
or a dict of policy id strings to
(Policy class, obs_space, action_space, config)-tuples. If a
dict is specified, then we are in multi-agent mode and a
policy_mapping_fn can also be set (if not, will map all agents
to DEFAULT_POLICY_ID).
policy_mapping_fn (Optional[Callable[[AgentID], PolicyID]]): A
callable that maps agent ids to policy ids in multi-agent mode.
This function will be called each time a new agent appears in
an episode, to bind that agent to a policy for the duration of
the episode. If not provided, will map all agents to
DEFAULT_POLICY_ID.
policies_to_train (Optional[List[PolicyID]]): Optional list of
policies to train, or None for all policies.
tf_session_creator (Optional[Callable[[], tf1.Session]]): A
function that returns a TF session. This is optional and only
useful with TFPolicy.
rollout_fragment_length (int): The target number of env transitions
to include in each sample batch returned from this worker.
batch_mode (str): One of the following batch modes:
"truncate_episodes": Each call to sample() will return a batch
of at most `rollout_fragment_length * num_envs` in size.
The batch will be exactly
`rollout_fragment_length * num_envs` in size if
postprocessing does not change batch sizes. Episodes may be
truncated in order to meet this size requirement.
"complete_episodes": Each call to sample() will return a batch
of at least `rollout_fragment_length * num_envs` in size.
Episodes will not be truncated, but multiple episodes may
be packed within one batch to meet the batch size. Note
that when `num_envs > 1`, episode steps will be buffered
until the episode completes, and hence batches may contain
significant amounts of off-policy data.
episode_horizon (int): Whether to stop episodes at this horizon.
preprocessor_pref (str): Whether to prefer RLlib preprocessors
("rllib") or deepmind ("deepmind") when applicable.
sample_async (bool): Whether to compute samples asynchronously in
the background, which improves throughput but can cause samples
to be slightly off-policy.
compress_observations (bool): If true, compress the observations.
They can be decompressed with rllib/utils/compression.
num_envs (int): If more than one, will create multiple envs
and vectorize the computation of actions. This has no effect if
if the env already implements VectorEnv.
observation_fn (ObservationFunction): Optional multi-agent
observation function.
observation_filter (str): Name of observation filter to use.
clip_rewards (bool): Whether to clip rewards to [-1, 1] prior to
experience postprocessing. Setting to None means clip for Atari
only.
clip_actions (bool): Whether to clip action values to the range
specified by the policy action space.
env_config (EnvConfigDict): Config to pass to the env creator.
model_config (ModelConfigDict): Config to use when creating the
policy model.
policy_config (TrainerConfigDict): Config to pass to the policy.
In the multi-agent case, this config will be merged with the
per-policy configs specified by `policy_spec`.
worker_index (int): For remote workers, this should be set to a
non-zero and unique value. This index is passed to created envs
through EnvContext so that envs can be configured per worker.
num_workers (int): For remote workers, how many workers altogether
have been created?
monitor_path (str): Write out episode stats and videos to this
directory if specified.
log_dir (str): Directory where logs can be placed.
log_level (str): Set the root log level on creation.
callbacks (DefaultCallbacks): Custom training callbacks.
input_creator (Callable[[IOContext], InputReader]): Function that
returns an InputReader object for loading previous generated
experiences.
input_evaluation (List[str]): How to evaluate the policy
performance. This only makes sense to set when the input is
reading offline data. The possible values include:
- "is": the step-wise importance sampling estimator.
- "wis": the weighted step-wise is estimator.
- "simulation": run the environment in the background, but
use this data for evaluation only and never for learning.
output_creator (Callable[[IOContext], OutputWriter]): Function that
returns an OutputWriter object for saving generated
experiences.
remote_worker_envs (bool): If using num_envs > 1, whether to create
those new envs in remote processes instead of in the current
process. This adds overheads, but can make sense if your envs
remote_env_batch_wait_ms (float): Timeout that remote workers
are waiting when polling environments. 0 (continue when at
least one env is ready) is a reasonable default, but optimal
value could be obtained by measuring your environment
step / reset and model inference perf.
soft_horizon (bool): Calculate rewards but don't reset the
environment when the horizon is hit.
no_done_at_end (bool): Ignore the done=True at the end of the
episode and instead record done=False.
seed (int): Set the seed of both np and tf to this value to
to ensure each remote worker has unique exploration behavior.
extra_python_environs (dict): Extra python environments need to
be set.
fake_sampler (bool): Use a fake (inf speed) sampler for testing.
spaces (Optional[Dict[PolicyID, Tuple[gym.spaces.Space,
gym.spaces.Space]]]): An optional space dict mapping policy IDs
to (obs_space, action_space)-tuples. This is used in case no
Env is created on this RolloutWorker.
policy: Obsoleted arg. Use `policy_spec` instead.
"""
# Deprecated arg.
if policy is not None:
deprecation_warning("policy", "policy_spec", error=False)
policy_spec = policy
assert policy_spec is not None, "Must provide `policy_spec` when " \
"creating RolloutWorker!"
self._original_kwargs: dict = locals().copy()
del self._original_kwargs["self"]
global _global_worker
_global_worker = self
# set extra environs first
if extra_python_environs:
for key, value in extra_python_environs.items():
os.environ[key] = str(value)
def gen_rollouts():
while True:
yield self.sample()
ParallelIteratorWorker.__init__(self, gen_rollouts, False)
policy_config: TrainerConfigDict = policy_config or {}
if (tf1 and policy_config.get("framework") in ["tf2", "tfe"]
# This eager check is necessary for certain all-framework tests
# that use tf's eager_mode() context generator.
and not tf1.executing_eagerly()):
tf1.enable_eager_execution()
if log_level:
logging.getLogger("ray.rllib").setLevel(log_level)
if worker_index > 1:
disable_log_once_globally() # only need 1 worker to log
elif log_level == "DEBUG":
enable_periodic_logging()
env_context = EnvContext(env_config or {}, worker_index)
self.env_context = env_context
self.policy_config: TrainerConfigDict = policy_config
if callbacks:
self.callbacks: "DefaultCallbacks" = callbacks()
else:
from ray.rllib.agents.callbacks import DefaultCallbacks
self.callbacks: "DefaultCallbacks" = DefaultCallbacks()
self.worker_index: int = worker_index
self.num_workers: int = num_workers
model_config: ModelConfigDict = model_config or {}
policy_mapping_fn = (policy_mapping_fn
or (lambda agent_id: DEFAULT_POLICY_ID))
if not callable(policy_mapping_fn):
raise ValueError("Policy mapping function not callable?")
self.env_creator: Callable[[EnvContext], EnvType] = env_creator
self.rollout_fragment_length: int = rollout_fragment_length * num_envs
self.batch_mode: str = batch_mode
self.compress_observations: bool = compress_observations
self.preprocessing_enabled: bool = True
self.last_batch: SampleBatchType = None
self.global_vars: dict = None
self.fake_sampler: bool = fake_sampler
# No Env will be used in this particular worker (not needed).
if worker_index == 0 and num_workers > 0 and \
policy_config["create_env_on_driver"] is False:
self.env = None
# Create an env for this worker.
else:
self.env = _validate_env(env_creator(env_context))
if validate_env is not None:
validate_env(self.env, self.env_context)
if isinstance(self.env, (BaseEnv, MultiAgentEnv)):
def wrap(env):
return env # we can't auto-wrap these env types
elif is_atari(self.env) and \
not model_config.get("custom_preprocessor") and \
preprocessor_pref == "deepmind":
# Deepmind wrappers already handle all preprocessing.
self.preprocessing_enabled = False
# If clip_rewards not explicitly set to False, switch it
# on here (clip between -1.0 and 1.0).
if clip_rewards is None:
clip_rewards = True
def wrap(env):
env = wrap_deepmind(
env,
dim=model_config.get("dim"),
framestack=model_config.get("framestack"))
if monitor_path:
from gym import wrappers
env = wrappers.Monitor(env, monitor_path, resume=True)
return env
else:
def wrap(env):
if monitor_path:
from gym import wrappers
env = wrappers.Monitor(env, monitor_path, resume=True)
return env
self.env: EnvType = wrap(self.env)
def make_env(vector_index):
return wrap(
env_creator(
env_context.copy_with_overrides(
worker_index=worker_index,
vector_index=vector_index,
remote=remote_worker_envs)))
self.make_env_fn = make_env
self.tf_sess = None
policy_dict = _validate_and_canonicalize(
policy_spec, self.env, spaces=spaces)
self.policies_to_train: List[PolicyID] = policies_to_train or list(
policy_dict.keys())
self.policy_map: Dict[PolicyID, Policy] = None
self.preprocessors: Dict[PolicyID, Preprocessor] = None
# set numpy and python seed
if seed is not None:
np.random.seed(seed)
random.seed(seed)
if not hasattr(self.env, "seed"):
logger.info("Env doesn't support env.seed(): {}".format(
self.env))
else:
self.env.seed(seed)
try:
assert torch is not None
torch.manual_seed(seed)
except AssertionError:
logger.info("Could not seed torch")
if _has_tensorflow_graph(policy_dict) and not (
tf1 and tf1.executing_eagerly()):
if not tf1:
raise ImportError("Could not import tensorflow")
with tf1.Graph().as_default():
if tf_session_creator:
self.tf_sess = tf_session_creator()
else:
self.tf_sess = tf1.Session(
config=tf1.ConfigProto(
gpu_options=tf1.GPUOptions(allow_growth=True)))
with self.tf_sess.as_default():
# set graph-level seed
if seed is not None:
tf1.set_random_seed(seed)
self.policy_map, self.preprocessors = \
self._build_policy_map(policy_dict, policy_config)
else:
self.policy_map, self.preprocessors = self._build_policy_map(
policy_dict, policy_config)
if (ray.is_initialized()
and ray.worker._mode() != ray.worker.LOCAL_MODE):
# Check available number of GPUs
if not ray.get_gpu_ids():
logger.debug("Creating policy evaluation worker {}".format(
worker_index) +
" on CPU (please ignore any CUDA init errors)")
elif (policy_config["framework"] in ["tf2", "tf", "tfe"] and
not tf.config.experimental.list_physical_devices("GPU")) or \
(policy_config["framework"] == "torch" and
not torch.cuda.is_available()):
raise RuntimeError(
"GPUs were assigned to this worker by Ray, but "
"your DL framework ({}) reports GPU acceleration is "
"disabled. This could be due to a bad CUDA- or {} "
"installation.".format(policy_config["framework"],
policy_config["framework"]))
self.multiagent: bool = set(
self.policy_map.keys()) != {DEFAULT_POLICY_ID}
if self.multiagent and self.env is not None:
if not ((isinstance(self.env, MultiAgentEnv)
or isinstance(self.env, ExternalMultiAgentEnv))
or isinstance(self.env, BaseEnv)):
raise ValueError(
"Have multiple policies {}, but the env ".format(
self.policy_map) +
"{} is not a subclass of BaseEnv, MultiAgentEnv or "
"ExternalMultiAgentEnv?".format(self.env))
self.filters: Dict[PolicyID, Filter] = {
policy_id: get_filter(observation_filter,
policy.observation_space.shape)
for (policy_id, policy) in self.policy_map.items()
}
if self.worker_index == 0:
logger.info("Built filter map: {}".format(self.filters))
self.num_envs: int = num_envs
if self.env is None:
self.async_env = None
elif "custom_vector_env" in policy_config:
custom_vec_wrapper = policy_config["custom_vector_env"]
self.async_env = custom_vec_wrapper(self.env)
else:
# Always use vector env for consistency even if num_envs = 1.
self.async_env: BaseEnv = BaseEnv.to_base_env(
self.env,
make_env=make_env,
num_envs=num_envs,
remote_envs=remote_worker_envs,
remote_env_batch_wait_ms=remote_env_batch_wait_ms)
# `truncate_episodes`: Allow a batch to contain more than one episode
# (fragments) and always make the batch `rollout_fragment_length`
# long.
if self.batch_mode == "truncate_episodes":
pack = True
# `complete_episodes`: Never cut episodes and sampler will return
# exactly one (complete) episode per poll.
elif self.batch_mode == "complete_episodes":
rollout_fragment_length = float("inf")
pack = False
else:
raise ValueError("Unsupported batch mode: {}".format(
self.batch_mode))
self.io_context: IOContext = IOContext(log_dir, policy_config,
worker_index, self)
self.reward_estimators: List[OffPolicyEstimator] = []
for method in input_evaluation:
if method == "simulation":
logger.warning(
"Requested 'simulation' input evaluation method: "
"will discard all sampler outputs and keep only metrics.")
sample_async = True
elif method == "is":
ise = ImportanceSamplingEstimator.create(self.io_context)
self.reward_estimators.append(ise)
elif method == "wis":
wise = WeightedImportanceSamplingEstimator.create(
self.io_context)
self.reward_estimators.append(wise)
else:
raise ValueError(
"Unknown evaluation method: {}".format(method))
if self.env is None:
self.sampler = None
elif sample_async:
self.sampler = AsyncSampler(
worker=self,
env=self.async_env,
policies=self.policy_map,
policy_mapping_fn=policy_mapping_fn,
preprocessors=self.preprocessors,
obs_filters=self.filters,
clip_rewards=clip_rewards,
rollout_fragment_length=rollout_fragment_length,
callbacks=self.callbacks,
horizon=episode_horizon,
multiple_episodes_in_batch=pack,
tf_sess=self.tf_sess,
clip_actions=clip_actions,
blackhole_outputs="simulation" in input_evaluation,
soft_horizon=soft_horizon,
no_done_at_end=no_done_at_end,
observation_fn=observation_fn,
_use_trajectory_view_api=policy_config.get(
"_use_trajectory_view_api", False))
# Start the Sampler thread.
self.sampler.start()
else:
self.sampler = SyncSampler(
worker=self,
env=self.async_env,
policies=self.policy_map,
policy_mapping_fn=policy_mapping_fn,
preprocessors=self.preprocessors,
obs_filters=self.filters,
clip_rewards=clip_rewards,
rollout_fragment_length=rollout_fragment_length,
callbacks=self.callbacks,
horizon=episode_horizon,
multiple_episodes_in_batch=pack,
tf_sess=self.tf_sess,
clip_actions=clip_actions,
soft_horizon=soft_horizon,
no_done_at_end=no_done_at_end,
observation_fn=observation_fn,
_use_trajectory_view_api=policy_config.get(
"_use_trajectory_view_api", False))
self.input_reader: InputReader = input_creator(self.io_context)
self.output_writer: OutputWriter = output_creator(self.io_context)
logger.debug(
"Created rollout worker with env {} ({}), policies {}".format(
self.async_env, self.env, self.policy_map))
def __init__(self,
env_creator,
policy_graph,
policy_mapping_fn=None,
policies_to_train=None,
tf_session_creator=None,
batch_steps=100,
batch_mode="truncate_episodes",
episode_horizon=None,
preprocessor_pref="deepmind",
sample_async=False,
compress_observations=False,
num_envs=1,
observation_filter="NoFilter",
clip_rewards=False,
env_config=None,
model_config=None,
policy_config=None,
worker_index=0,
monitor_path=None):
"""Initialize a policy evaluator.
Arguments:
env_creator (func): Function that returns a gym.Env given an
EnvContext wrapped configuration.
policy_graph (class|dict): Either a class implementing
PolicyGraph, or a dictionary of policy id strings to
(PolicyGraph, obs_space, action_space, config) tuples. If a
dict is specified, then we are in multi-agent mode and a
policy_mapping_fn should also be set.
policy_mapping_fn (func): A function that maps agent ids to
policy ids in multi-agent mode. This function will be called
each time a new agent appears in an episode, to bind that agent
to a policy for the duration of the episode.
policies_to_train (list): Optional whitelist of policies to train,
or None for all policies.
tf_session_creator (func): A function that returns a TF session.
This is optional and only useful with TFPolicyGraph.
batch_steps (int): The target number of env transitions to include
in each sample batch returned from this evaluator.
batch_mode (str): One of the following batch modes:
"truncate_episodes": Each call to sample() will return a batch
of at most `batch_steps` in size. The batch will be exactly
`batch_steps` in size if postprocessing does not change
batch sizes. Episodes may be truncated in order to meet
this size requirement. When `num_envs > 1`, episodes will
be truncated to sequences of `batch_size / num_envs` in
length.
"complete_episodes": Each call to sample() will return a batch
of at least `batch_steps in size. Episodes will not be
truncated, but multiple episodes may be packed within one
batch to meet the batch size. Note that when
`num_envs > 1`, episode steps will be buffered until the
episode completes, and hence batches may contain
significant amounts of off-policy data.
episode_horizon (int): Whether to stop episodes at this horizon.
preprocessor_pref (str): Whether to prefer RLlib preprocessors
("rllib") or deepmind ("deepmind") when applicable.
sample_async (bool): Whether to compute samples asynchronously in
the background, which improves throughput but can cause samples
to be slightly off-policy.
compress_observations (bool): If true, compress the observations.
They can be decompressed with rllib/utils/compression.
num_envs (int): If more than one, will create multiple envs
and vectorize the computation of actions. This has no effect if
if the env already implements VectorEnv.
observation_filter (str): Name of observation filter to use.
clip_rewards (bool): Whether to clip rewards to [-1, 1] prior to
experience postprocessing.
env_config (dict): Config to pass to the env creator.
model_config (dict): Config to use when creating the policy model.
policy_config (dict): Config to pass to the policy. In the
multi-agent case, this config will be merged with the
per-policy configs specified by `policy_graph`.
worker_index (int): For remote evaluators, this should be set to a
non-zero and unique value. This index is passed to created envs
through EnvContext so that envs can be configured per worker.
monitor_path (str): Write out episode stats and videos to this
directory if specified.
"""
env_context = EnvContext(env_config or {}, worker_index)
policy_config = policy_config or {}
self.policy_config = policy_config
model_config = model_config or {}
policy_mapping_fn = (policy_mapping_fn
or (lambda agent_id: DEFAULT_POLICY_ID))
self.env_creator = env_creator
self.batch_steps = batch_steps
self.batch_mode = batch_mode
self.compress_observations = compress_observations
self.env = env_creator(env_context)
if isinstance(self.env, VectorEnv) or \
isinstance(self.env, ServingEnv) or \
isinstance(self.env, MultiAgentEnv) or \
isinstance(self.env, AsyncVectorEnv):
def wrap(env):
return env # we can't auto-wrap these env types
elif is_atari(self.env) and \
"custom_preprocessor" not in model_config and \
preprocessor_pref == "deepmind":
def wrap(env):
env = wrap_deepmind(env,
dim=model_config.get("dim", 84),
framestack=not model_config.get("use_lstm")
and not model_config.get("no_framestack"))
if monitor_path:
env = _monitor(env, monitor_path)
return env
else:
def wrap(env):
env = ModelCatalog.get_preprocessor_as_wrapper(
env, model_config)
if monitor_path:
env = _monitor(env, monitor_path)
return env
self.env = wrap(self.env)
def make_env(vector_index):
return wrap(
env_creator(env_context.with_vector_index(vector_index)))
self.tf_sess = None
policy_dict = _validate_and_canonicalize(policy_graph, self.env)
self.policies_to_train = policies_to_train or list(policy_dict.keys())
if _has_tensorflow_graph(policy_dict):
with tf.Graph().as_default():
if tf_session_creator:
self.tf_sess = tf_session_creator()
else:
self.tf_sess = tf.Session(config=tf.ConfigProto(
gpu_options=tf.GPUOptions(allow_growth=True)))
with self.tf_sess.as_default():
self.policy_map = self._build_policy_map(
policy_dict, policy_config)
else:
self.policy_map = self._build_policy_map(policy_dict,
policy_config)
self.multiagent = self.policy_map.keys() != {DEFAULT_POLICY_ID}
self.filters = {
policy_id: get_filter(observation_filter,
policy.observation_space.shape)
for (policy_id, policy) in self.policy_map.items()
}
# Always use vector env for consistency even if num_envs = 1
self.async_env = AsyncVectorEnv.wrap_async(self.env,
make_env=make_env,
num_envs=num_envs)
self.num_envs = num_envs
if self.batch_mode == "truncate_episodes":
if batch_steps % num_envs != 0:
raise ValueError(
"In 'truncate_episodes' batch mode, `batch_steps` must be "
"evenly divisible by `num_envs`. Got {} and {}.".format(
batch_steps, num_envs))
batch_steps = batch_steps // num_envs
pack_episodes = True
elif self.batch_mode == "complete_episodes":
batch_steps = float("inf") # never cut episodes
pack_episodes = False # sampler will return 1 episode per poll
else:
raise ValueError("Unsupported batch mode: {}".format(
self.batch_mode))
if sample_async:
self.sampler = AsyncSampler(self.async_env,
self.policy_map,
policy_mapping_fn,
self.filters,
clip_rewards,
batch_steps,
horizon=episode_horizon,
pack=pack_episodes,
tf_sess=self.tf_sess)
self.sampler.start()
else:
self.sampler = SyncSampler(self.async_env,
self.policy_map,
policy_mapping_fn,
self.filters,
clip_rewards,
batch_steps,
horizon=episode_horizon,
pack=pack_episodes,
tf_sess=self.tf_sess)
def _init(self):
policy_params = {"action_noise_std": 0.01}
self.theta_dict = []
self.policymax = []
self.curr_parent = 0
self.population = []
self.returns_n2 = []
self.ret = []
env = self.env_creator(self.config["env_config"])
self.monitor_path = self.logdir if self.config["monitor"] else None
if is_atari(env):
env = wrap_deepmind(env, dim=84, framestack=4)
env2 = self.env_creator(self.config["env_config"])
if is_atari(env2):
env2 = wrap_deepmind(env2, dim=84, framestack=4)
from ray.rllib import models
preprocessor = models.ModelCatalog.get_preprocessor(env)
for p in range(self.config["pop_size"]):
with tf.Graph().as_default():
self.sess = utils.make_session(single_threaded=False)
self.policy = policies.GenericPolicy(
self.sess, env, env2, env.action_space,
env.observation_space, preprocessor,
self.config["observation_filter"], self.config["model"],
**policy_params)
tf_util.initialize(self.sess)
theta = self.policy.get_weights()
self.theta_dict.append(theta)
self.optimizer = optimizers.Adam(self.policy, self.config["stepsize"])
self.report_length = self.config["report_length"]
# Create the shared noise table.
logger.info("Creating shared noise table.")
noise_id = create_shared_noise.remote(self.config["noise_size"])
self.noise = SharedNoiseTable(ray.get(noise_id))
# for p in range(self.config["population_size"]):
# noise_index = self.noise.sample_index(self.policy.num_params)
# self.population.append(self.noise.get(noise_index, self.policy.num_params))
# Create the actors.
logger.info("Creating actors.")
self.workers = [
Worker.remote(self.config, policy_params, self.env_creator,
noise_id, self.monitor_path)
for _ in range(self.config["num_workers"])
]
self.episodes_so_far = 0
self.reward_list1 = []
self.reward_list2 = []
self.reward_list3 = []
self.tstart = time.time()
self.noisy_rew_max1 = -1000
self.noisy_rew_mean1 = -1000
self.noisy_rew_max2 = -1000
self.noisy_rew_mean2 = -1000
self.noisy_rew_max3 = -1000
self.noisy_rew_mean3 = -1000
self.noisy_rew_max4 = -1000
self.noisy_rew_mean4 = -1000
self.reward_mean1 = -1000
self.reward_mean2 = -1000
self.reward_mean3 = -1000
self.maxrew = []
self.maxrew2 = []
def __init__(self,
env_creator,
policy_graph,
policy_mapping_fn=None,
policies_to_train=None,
tf_session_creator=None,
batch_steps=100,
batch_mode="truncate_episodes",
episode_horizon=None,
preprocessor_pref="deepmind",
sample_async=False,
compress_observations=False,
num_envs=1,
observation_filter="NoFilter",
clip_rewards=None,
clip_actions=True,
env_config=None,
model_config=None,
policy_config=None,
worker_index=0,
monitor_path=None,
log_dir=None,
log_level=None,
callbacks=None,
input_creator=lambda ioctx: ioctx.default_sampler_input(),
input_evaluation_method=None,
output_creator=lambda ioctx: NoopOutput()):
"""Initialize a policy evaluator.
Arguments:
env_creator (func): Function that returns a gym.Env given an
EnvContext wrapped configuration.
policy_graph (class|dict): Either a class implementing
PolicyGraph, or a dictionary of policy id strings to
(PolicyGraph, obs_space, action_space, config) tuples. If a
dict is specified, then we are in multi-agent mode and a
policy_mapping_fn should also be set.
policy_mapping_fn (func): A function that maps agent ids to
policy ids in multi-agent mode. This function will be called
each time a new agent appears in an episode, to bind that agent
to a policy for the duration of the episode.
policies_to_train (list): Optional whitelist of policies to train,
or None for all policies.
tf_session_creator (func): A function that returns a TF session.
This is optional and only useful with TFPolicyGraph.
batch_steps (int): The target number of env transitions to include
in each sample batch returned from this evaluator.
batch_mode (str): One of the following batch modes:
"truncate_episodes": Each call to sample() will return a batch
of at most `batch_steps * num_envs` in size. The batch will
be exactly `batch_steps * num_envs` in size if
postprocessing does not change batch sizes. Episodes may be
truncated in order to meet this size requirement.
"complete_episodes": Each call to sample() will return a batch
of at least `batch_steps * num_envs` in size. Episodes will
not be truncated, but multiple episodes may be packed
within one batch to meet the batch size. Note that when
`num_envs > 1`, episode steps will be buffered until the
episode completes, and hence batches may contain
significant amounts of off-policy data.
episode_horizon (int): Whether to stop episodes at this horizon.
preprocessor_pref (str): Whether to prefer RLlib preprocessors
("rllib") or deepmind ("deepmind") when applicable.
sample_async (bool): Whether to compute samples asynchronously in
the background, which improves throughput but can cause samples
to be slightly off-policy.
compress_observations (bool): If true, compress the observations.
They can be decompressed with rllib/utils/compression.
num_envs (int): If more than one, will create multiple envs
and vectorize the computation of actions. This has no effect if
if the env already implements VectorEnv.
observation_filter (str): Name of observation filter to use.
clip_rewards (bool): Whether to clip rewards to [-1, 1] prior to
experience postprocessing. Setting to None means clip for Atari
only.
clip_actions (bool): Whether to clip action values to the range
specified by the policy action space.
env_config (dict): Config to pass to the env creator.
model_config (dict): Config to use when creating the policy model.
policy_config (dict): Config to pass to the policy. In the
multi-agent case, this config will be merged with the
per-policy configs specified by `policy_graph`.
worker_index (int): For remote evaluators, this should be set to a
non-zero and unique value. This index is passed to created envs
through EnvContext so that envs can be configured per worker.
monitor_path (str): Write out episode stats and videos to this
directory if specified.
log_dir (str): Directory where logs can be placed.
log_level (str): Set the root log level on creation.
callbacks (dict): Dict of custom debug callbacks.
input_creator (func): Function that returns an InputReader object
for loading previous generated experiences.
input_evaluation_method (str): How to evaluate the current policy.
This only applies when the input is reading offline data.
Options are:
- None: don't evaluate the policy. The episode reward and
other metrics will be NaN.
- "simulation": run the environment in the background, but
use this data for evaluation only and never for learning.
- "counterfactual": use counterfactual policy evaluation to
estimate performance.
output_creator (func): Function that returns an OutputWriter object
for saving generated experiences.
"""
if log_level:
logging.getLogger("ray.rllib").setLevel(log_level)
env_context = EnvContext(env_config or {}, worker_index)
policy_config = policy_config or {}
self.policy_config = policy_config
self.callbacks = callbacks or {}
model_config = model_config or {}
policy_mapping_fn = (policy_mapping_fn
or (lambda agent_id: DEFAULT_POLICY_ID))
if not callable(policy_mapping_fn):
raise ValueError(
"Policy mapping function not callable. If you're using Tune, "
"make sure to escape the function with tune.function() "
"to prevent it from being evaluated as an expression.")
self.env_creator = env_creator
self.sample_batch_size = batch_steps * num_envs
self.batch_mode = batch_mode
self.compress_observations = compress_observations
self.preprocessing_enabled = True
self.env = env_creator(env_context)
if isinstance(self.env, MultiAgentEnv) or \
isinstance(self.env, AsyncVectorEnv):
def wrap(env):
return env # we can't auto-wrap these env types
elif is_atari(self.env) and \
not model_config.get("custom_preprocessor") and \
preprocessor_pref == "deepmind":
# Deepmind wrappers already handle all preprocessing
self.preprocessing_enabled = False
if clip_rewards is None:
clip_rewards = True
def wrap(env):
env = wrap_deepmind(
env,
dim=model_config.get("dim"),
framestack=model_config.get("framestack"))
if monitor_path:
env = _monitor(env, monitor_path)
return env
else:
def wrap(env):
if monitor_path:
env = _monitor(env, monitor_path)
return env
self.env = wrap(self.env)
def make_env(vector_index):
return wrap(
env_creator(env_context.with_vector_index(vector_index)))
self.tf_sess = None
policy_dict = _validate_and_canonicalize(policy_graph, self.env)
self.policies_to_train = policies_to_train or list(policy_dict.keys())
if _has_tensorflow_graph(policy_dict):
if (ray.worker._mode() != ray.worker.LOCAL_MODE
and not ray.get_gpu_ids()):
logger.info("Creating policy evaluation worker {}".format(
worker_index) +
" on CPU (please ignore any CUDA init errors)")
with tf.Graph().as_default():
if tf_session_creator:
self.tf_sess = tf_session_creator()
else:
self.tf_sess = tf.Session(
config=tf.ConfigProto(
gpu_options=tf.GPUOptions(allow_growth=True)))
with self.tf_sess.as_default():
self.policy_map, self.preprocessors = \
self._build_policy_map(policy_dict, policy_config)
else:
self.policy_map, self.preprocessors = self._build_policy_map(
policy_dict, policy_config)
self.multiagent = set(self.policy_map.keys()) != {DEFAULT_POLICY_ID}
if self.multiagent:
if not (isinstance(self.env, MultiAgentEnv)
or isinstance(self.env, AsyncVectorEnv)):
raise ValueError(
"Have multiple policy graphs {}, but the env ".format(
self.policy_map) +
"{} is not a subclass of MultiAgentEnv?".format(self.env))
self.filters = {
policy_id: get_filter(observation_filter,
policy.observation_space.shape)
for (policy_id, policy) in self.policy_map.items()
}
# Always use vector env for consistency even if num_envs = 1
self.async_env = AsyncVectorEnv.wrap_async(
self.env, make_env=make_env, num_envs=num_envs)
self.num_envs = num_envs
if self.batch_mode == "truncate_episodes":
unroll_length = batch_steps
pack_episodes = True
elif self.batch_mode == "complete_episodes":
unroll_length = float("inf") # never cut episodes
pack_episodes = False # sampler will return 1 episode per poll
else:
raise ValueError("Unsupported batch mode: {}".format(
self.batch_mode))
if input_evaluation_method == "simulation":
logger.warning(
"Requested 'simulation' input evaluation method: "
"will discard all sampler outputs and keep only metrics.")
sample_async = True
elif input_evaluation_method == "counterfactual":
raise NotImplementedError
elif input_evaluation_method is None:
pass
else:
raise ValueError("Unknown evaluation method: {}".format(
input_evaluation_method))
if sample_async:
self.sampler = AsyncSampler(
self.async_env,
self.policy_map,
policy_mapping_fn,
self.preprocessors,
self.filters,
clip_rewards,
unroll_length,
self.callbacks,
horizon=episode_horizon,
pack=pack_episodes,
tf_sess=self.tf_sess,
clip_actions=clip_actions,
blackhole_outputs=input_evaluation_method == "simulation")
self.sampler.start()
else:
self.sampler = SyncSampler(
self.async_env,
self.policy_map,
policy_mapping_fn,
self.preprocessors,
self.filters,
clip_rewards,
unroll_length,
self.callbacks,
horizon=episode_horizon,
pack=pack_episodes,
tf_sess=self.tf_sess,
clip_actions=clip_actions)
self.io_context = IOContext(log_dir, policy_config, worker_index, self)
self.input_reader = input_creator(self.io_context)
assert isinstance(self.input_reader, InputReader), self.input_reader
self.output_writer = output_creator(self.io_context)
assert isinstance(self.output_writer, OutputWriter), self.output_writer
logger.debug("Created evaluator with env {} ({}), policies {}".format(
self.async_env, self.env, self.policy_map))
def __init__(self,
env_creator,
policy_graph,
policy_mapping_fn=None,
policies_to_train=None,
tf_session_creator=None,
batch_steps=100,
batch_mode="truncate_episodes",
episode_horizon=None,
preprocessor_pref="deepmind",
sample_async=False,
compress_observations=False,
num_envs=1,
observation_filter="NoFilter",
clip_rewards=None,
clip_actions=True,
env_config=None,
model_config=None,
policy_config=None,
worker_index=0,
monitor_path=None,
log_dir=None,
log_level=None,
callbacks=None,
input_creator=lambda ioctx: ioctx.default_sampler_input(),
input_evaluation_method=None,
output_creator=lambda ioctx: NoopOutput()):
"""Initialize a policy evaluator.
Arguments:
env_creator (func): Function that returns a gym.Env given an
EnvContext wrapped configuration.
policy_graph (class|dict): Either a class implementing
PolicyGraph, or a dictionary of policy id strings to
(PolicyGraph, obs_space, action_space, config) tuples. If a
dict is specified, then we are in multi-agent mode and a
policy_mapping_fn should also be set.
policy_mapping_fn (func): A function that maps agent ids to
policy ids in multi-agent mode. This function will be called
each time a new agent appears in an episode, to bind that agent
to a policy for the duration of the episode.
policies_to_train (list): Optional whitelist of policies to train,
or None for all policies.
tf_session_creator (func): A function that returns a TF session.
This is optional and only useful with TFPolicyGraph.
batch_steps (int): The target number of env transitions to include
in each sample batch returned from this evaluator.
batch_mode (str): One of the following batch modes:
"truncate_episodes": Each call to sample() will return a batch
of at most `batch_steps * num_envs` in size. The batch will
be exactly `batch_steps * num_envs` in size if
postprocessing does not change batch sizes. Episodes may be
truncated in order to meet this size requirement.
"complete_episodes": Each call to sample() will return a batch
of at least `batch_steps * num_envs` in size. Episodes will
not be truncated, but multiple episodes may be packed
within one batch to meet the batch size. Note that when
`num_envs > 1`, episode steps will be buffered until the
episode completes, and hence batches may contain
significant amounts of off-policy data.
episode_horizon (int): Whether to stop episodes at this horizon.
preprocessor_pref (str): Whether to prefer RLlib preprocessors
("rllib") or deepmind ("deepmind") when applicable.
sample_async (bool): Whether to compute samples asynchronously in
the background, which improves throughput but can cause samples
to be slightly off-policy.
compress_observations (bool): If true, compress the observations.
They can be decompressed with rllib/utils/compression.
num_envs (int): If more than one, will create multiple envs
and vectorize the computation of actions. This has no effect if
if the env already implements VectorEnv.
observation_filter (str): Name of observation filter to use.
clip_rewards (bool): Whether to clip rewards to [-1, 1] prior to
experience postprocessing. Setting to None means clip for Atari
only.
clip_actions (bool): Whether to clip action values to the range
specified by the policy action space.
env_config (dict): Config to pass to the env creator.
model_config (dict): Config to use when creating the policy model.
policy_config (dict): Config to pass to the policy. In the
multi-agent case, this config will be merged with the
per-policy configs specified by `policy_graph`.
worker_index (int): For remote evaluators, this should be set to a
non-zero and unique value. This index is passed to created envs
through EnvContext so that envs can be configured per worker.
monitor_path (str): Write out episode stats and videos to this
directory if specified.
log_dir (str): Directory where logs can be placed.
log_level (str): Set the root log level on creation.
callbacks (dict): Dict of custom debug callbacks.
input_creator (func): Function that returns an InputReader object
for loading previous generated experiences.
input_evaluation_method (str): How to evaluate the current policy.
This only applies when the input is reading offline data.
Options are:
- None: don't evaluate the policy. The episode reward and
other metrics will be NaN.
- "simulation": run the environment in the background, but
use this data for evaluation only and never for learning.
output_creator (func): Function that returns an OutputWriter object
for saving generated experiences.
"""
if log_level:
logging.getLogger("ray.rllib").setLevel(log_level)
env_context = EnvContext(env_config or {}, worker_index)
policy_config = policy_config or {}
self.policy_config = policy_config
self.callbacks = callbacks or {}
model_config = model_config or {}
policy_mapping_fn = (policy_mapping_fn
or (lambda agent_id: DEFAULT_POLICY_ID))
if not callable(policy_mapping_fn):
raise ValueError(
"Policy mapping function not callable. If you're using Tune, "
"make sure to escape the function with tune.function() "
"to prevent it from being evaluated as an expression.")
self.env_creator = env_creator
self.sample_batch_size = batch_steps * num_envs
self.batch_mode = batch_mode
self.compress_observations = compress_observations
self.preprocessing_enabled = True
self.env = env_creator(env_context)
if isinstance(self.env, MultiAgentEnv) or \
isinstance(self.env, AsyncVectorEnv):
def wrap(env):
return env # we can't auto-wrap these env types
elif is_atari(self.env) and \
not model_config.get("custom_preprocessor") and \
preprocessor_pref == "deepmind":
# Deepmind wrappers already handle all preprocessing
self.preprocessing_enabled = False
if clip_rewards is None:
clip_rewards = True
def wrap(env):
env = wrap_deepmind(
env,
dim=model_config.get("dim"),
framestack=model_config.get("framestack"))
if monitor_path:
env = _monitor(env, monitor_path)
return env
else:
def wrap(env):
if monitor_path:
env = _monitor(env, monitor_path)
return env
self.env = wrap(self.env)
def make_env(vector_index):
return wrap(
env_creator(env_context.with_vector_index(vector_index)))
self.tf_sess = None
policy_dict = _validate_and_canonicalize(policy_graph, self.env)
self.policies_to_train = policies_to_train or list(policy_dict.keys())
if _has_tensorflow_graph(policy_dict):
if (ray.worker._mode() != ray.worker.LOCAL_MODE
and not ray.get_gpu_ids()):
logger.info("Creating policy evaluation worker {}".format(
worker_index) +
" on CPU (please ignore any CUDA init errors)")
with tf.Graph().as_default():
if tf_session_creator:
self.tf_sess = tf_session_creator()
else:
self.tf_sess = tf.Session(
config=tf.ConfigProto(
gpu_options=tf.GPUOptions(allow_growth=True)))
with self.tf_sess.as_default():
self.policy_map, self.preprocessors = \
self._build_policy_map(policy_dict, policy_config)
else:
self.policy_map, self.preprocessors = self._build_policy_map(
policy_dict, policy_config)
self.multiagent = set(self.policy_map.keys()) != {DEFAULT_POLICY_ID}
if self.multiagent:
if not (isinstance(self.env, MultiAgentEnv)
or isinstance(self.env, AsyncVectorEnv)):
raise ValueError(
"Have multiple policy graphs {}, but the env ".format(
self.policy_map) +
"{} is not a subclass of MultiAgentEnv?".format(self.env))
self.filters = {
policy_id: get_filter(observation_filter,
policy.observation_space.shape)
for (policy_id, policy) in self.policy_map.items()
}
# Always use vector env for consistency even if num_envs = 1
self.async_env = AsyncVectorEnv.wrap_async(
self.env, make_env=make_env, num_envs=num_envs)
self.num_envs = num_envs
if self.batch_mode == "truncate_episodes":
unroll_length = batch_steps
pack_episodes = True
elif self.batch_mode == "complete_episodes":
unroll_length = float("inf") # never cut episodes
pack_episodes = False # sampler will return 1 episode per poll
else:
raise ValueError("Unsupported batch mode: {}".format(
self.batch_mode))
if input_evaluation_method == "simulation":
logger.warning(
"Requested 'simulation' input evaluation method: "
"will discard all sampler outputs and keep only metrics.")
sample_async = True
elif input_evaluation_method is None:
pass
else:
raise ValueError("Unknown evaluation method: {}".format(
input_evaluation_method))
if sample_async:
self.sampler = AsyncSampler(
self.async_env,
self.policy_map,
policy_mapping_fn,
self.preprocessors,
self.filters,
clip_rewards,
unroll_length,
self.callbacks,
horizon=episode_horizon,
pack=pack_episodes,
tf_sess=self.tf_sess,
clip_actions=clip_actions,
blackhole_outputs=input_evaluation_method == "simulation")
self.sampler.start()
else:
self.sampler = SyncSampler(
self.async_env,
self.policy_map,
policy_mapping_fn,
self.preprocessors,
self.filters,
clip_rewards,
unroll_length,
self.callbacks,
horizon=episode_horizon,
pack=pack_episodes,
tf_sess=self.tf_sess,
clip_actions=clip_actions)
self.io_context = IOContext(log_dir, policy_config, worker_index, self)
self.input_reader = input_creator(self.io_context)
assert isinstance(self.input_reader, InputReader), self.input_reader
self.output_writer = output_creator(self.io_context)
assert isinstance(self.output_writer, OutputWriter), self.output_writer
logger.debug("Created evaluator with env {} ({}), policies {}".format(
self.async_env, self.env, self.policy_map))