Skip to content
Cannot retrieve contributors at this time
import copy
import gym
from gym.spaces import Box, Discrete, MultiDiscrete, Space
import logging
import numpy as np
import platform
import os
import tree # pip install dm_tree
from typing import Any, Callable, Dict, List, Optional, Tuple, Type, \
import ray
from ray import ObjectRef
from ray import cloudpickle as pickle
from ray.rllib.env.base_env import BaseEnv, convert_to_base_env
from ray.rllib.env.env_context import EnvContext
from ray.rllib.env.external_env import ExternalEnv
from ray.rllib.env.multi_agent_env import MultiAgentEnv
from ray.rllib.env.external_multi_agent_env import ExternalMultiAgentEnv
from ray.rllib.env.utils import record_env_wrapper
from ray.rllib.env.vector_env import VectorEnv
from ray.rllib.env.wrappers.atari_wrappers import wrap_deepmind, is_atari
from ray.rllib.evaluation.sampler import AsyncSampler, SyncSampler
from ray.rllib.evaluation.metrics import RolloutMetrics
from ray.rllib.models import ModelCatalog
from ray.rllib.models.preprocessors import Preprocessor
from ray.rllib.offline import NoopOutput, IOContext, OutputWriter, InputReader
from ray.rllib.offline.off_policy_estimator import OffPolicyEstimator, \
from ray.rllib.offline.is_estimator import ImportanceSamplingEstimator
from ray.rllib.offline.wis_estimator import WeightedImportanceSamplingEstimator
from ray.rllib.policy.sample_batch import MultiAgentBatch, DEFAULT_POLICY_ID
from ray.rllib.policy.policy import Policy, PolicySpec
from ray.rllib.policy.policy_map import PolicyMap
from ray.rllib.policy.torch_policy import TorchPolicy
from ray.rllib.utils import force_list, merge_dicts
from ray.rllib.utils.annotations import Deprecated, DeveloperAPI
from ray.rllib.utils.debug import summarize, update_global_seed_if_necessary
from ray.rllib.utils.deprecation import deprecation_warning
from ray.rllib.utils.error import EnvError, ERR_MSG_NO_GPUS, \
from ray.rllib.utils.filter import get_filter, Filter
from ray.rllib.utils.framework import try_import_tf, try_import_torch
from ray.rllib.utils.sgd import do_minibatch_sgd
from ray.rllib.utils.tf_utils import get_gpu_devices as get_tf_gpu_devices
from ray.rllib.utils.tf_run_builder import TFRunBuilder
from ray.rllib.utils.typing import AgentID, EnvConfigDict, EnvCreator, \
EnvType, ModelConfigDict, ModelGradients, ModelWeights, \
MultiAgentPolicyConfigDict, PartialTrainerConfigDict, PolicyID, \
PolicyState, SampleBatchType, T
from ray.util.debug import log_once, disable_log_once_globally, \
from ray.util.iter import ParallelIteratorWorker
from ray.rllib.evaluation.episode import Episode
from ray.rllib.evaluation.observation_function import ObservationFunction
from ray.rllib.agents.callbacks import DefaultCallbacks # noqa
tf1, tf, tfv = try_import_tf()
torch, _ = try_import_torch()
logger = logging.getLogger(__name__)
# Handle to the current rollout worker, which will be set to the most recently
# created RolloutWorker in this process. This can be helpful to access in
# custom env or policy classes for debugging or advanced use cases.
_global_worker: "RolloutWorker" = None
def get_global_worker() -> "RolloutWorker":
"""Returns a handle to the active rollout worker in this process."""
global _global_worker
return _global_worker
def _update_env_seed_if_necessary(env: EnvType, seed: int, worker_idx: int,
vector_idx: int):
"""Set a deterministic random seed on environment.
NOTE: this may not work with remote environments (issue #18154).
if not seed:
# A single RL job is unlikely to have more than 10K
# rollout workers.
max_num_envs_per_workers: int = 1000
assert worker_idx < max_num_envs_per_workers, \
"Too many envs per worker. Random seeds may collide."
computed_seed: int = (
worker_idx * max_num_envs_per_workers + vector_idx + seed)
# Gym.env.
# This will silently fail for most OpenAI gyms
# (they do nothing and return None per default)
if not hasattr(env, "seed"):"Env doesn't support env.seed(): {}".format(env))
class RolloutWorker(ParallelIteratorWorker):
"""Common experience collection class.
This class wraps a policy instance and an environment class to
collect experiences from the environment. You can create many replicas of
this class as Ray actors to scale RL training.
This class supports vectorized and multi-agent policy evaluation (e.g.,
VectorEnv, MultiAgentEnv, etc.)
>>> # Create a rollout worker and using it to collect experiences.
>>> worker = RolloutWorker(
... env_creator=lambda _: gym.make("CartPole-v0"),
... policy_spec=PGTFPolicy)
>>> print(worker.sample())
"obs": [[...]], "actions": [[...]], "rewards": [[...]],
"dones": [[...]], "new_obs": [[...]]})
>>> # Creating a multi-agent rollout worker
>>> worker = RolloutWorker(
... env_creator=lambda _: MultiAgentTrafficGrid(num_cars=25),
... policy_spec={
... # Use an ensemble of two policies for car agents
... "car_policy1":
... (PGTFPolicy, Box(...), Discrete(...), {"gamma": 0.99}),
... "car_policy2":
... (PGTFPolicy, Box(...), Discrete(...), {"gamma": 0.95}),
... # Use a single shared policy for all traffic lights
... "traffic_light_policy":
... (PGTFPolicy, Box(...), Discrete(...), {}),
... },
... policy_mapping_fn=lambda agent_id, episode, **kwargs:
... random.choice(["car_policy1", "car_policy2"])
... if agent_id.startswith("car_") else "traffic_light_policy")
>>> print(worker.sample())
"car_policy1": SampleBatch(...),
"car_policy2": SampleBatch(...),
"traffic_light_policy": SampleBatch(...)})
def as_remote(cls,
num_cpus: Optional[int] = None,
num_gpus: Optional[Union[int, float]] = None,
memory: Optional[int] = None,
object_store_memory: Optional[int] = None,
resources: Optional[dict] = None) -> type:
"""Returns RolloutWorker class as a `@ray.remote using given options`.
The returned class can then be used to instantiate ray actors.
num_cpus: The number of CPUs to allocate for the remote actor.
num_gpus: The number of GPUs to allocate for the remote actor.
This could be a fraction as well.
memory: The heap memory request for the remote actor.
object_store_memory: The object store memory for the remote actor.
resources: The default custom resources to allocate for the remote
The `@ray.remote` decorated RolloutWorker class.
return ray.remote(
def __init__(
env_creator: EnvCreator,
validate_env: Optional[Callable[[EnvType, EnvContext],
None]] = None,
policy_spec: Optional[Union[type, Dict[PolicyID,
PolicySpec]]] = None,
policy_mapping_fn: Optional[Callable[[AgentID, "Episode"],
PolicyID]] = None,
policies_to_train: Optional[List[PolicyID]] = None,
tf_session_creator: Optional[Callable[[], "tf1.Session"]] = None,
rollout_fragment_length: int = 100,
count_steps_by: str = "env_steps",
batch_mode: str = "truncate_episodes",
episode_horizon: Optional[int] = None,
preprocessor_pref: str = "deepmind",
sample_async: bool = False,
compress_observations: bool = False,
num_envs: int = 1,
observation_fn: Optional["ObservationFunction"] = None,
observation_filter: str = "NoFilter",
clip_rewards: Optional[Union[bool, float]] = None,
normalize_actions: bool = True,
clip_actions: bool = False,
env_config: Optional[EnvConfigDict] = None,
model_config: Optional[ModelConfigDict] = None,
policy_config: Optional[PartialTrainerConfigDict] = None,
worker_index: int = 0,
num_workers: int = 0,
record_env: Union[bool, str] = False,
log_dir: Optional[str] = None,
log_level: Optional[str] = None,
callbacks: Type["DefaultCallbacks"] = None,
input_creator: Callable[[
], 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: Optional[dict] = None,
fake_sampler: bool = False,
spaces: Optional[Dict[PolicyID, Tuple[Space, Space]]] = None,
"""Initializes a RolloutWorker instance.
env_creator: Function that returns a gym.Env given an EnvContext
wrapped configuration.
validate_env: Optional callable to validate the generated
environment (only on worker=0).
policy_spec: The MultiAgentPolicyConfigDict mapping policy IDs
(str) to PolicySpec's or a single policy class to use.
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
policy_mapping_fn: 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
policies_to_train: Optional list of policies to train, or None
for all policies.
tf_session_creator: A function that returns a TF session.
This is optional and only useful with TFPolicy.
rollout_fragment_length: The target number of steps
(maesured in `count_steps_by`) to include in each sample
batch returned from this worker.
count_steps_by: The unit in which to count fragment
lengths. One of env_steps or agent_steps.
batch_mode: 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: Horizon at which to stop episodes (even if the
environment itself has not retured a "done" signal).
preprocessor_pref: Whether to use RLlib preprocessors
("rllib") or deepmind ("deepmind"), when applicable.
sample_async: Whether to compute samples asynchronously in
the background, which improves throughput but can cause samples
to be slightly off-policy.
compress_observations: If true, compress the observations.
They can be decompressed with rllib/utils/compression.
num_envs: 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: Optional multi-agent observation function.
observation_filter: Name of observation filter to use.
clip_rewards: True for clipping rewards to [-1.0, 1.0] prior
to experience postprocessing. None: Clip for Atari only.
float: Clip to [-clip_rewards; +clip_rewards].
normalize_actions: Whether to normalize actions to the
action space's bounds.
clip_actions: Whether to clip action values to the range
specified by the policy action space.
env_config: Config to pass to the env creator.
model_config: Config to use when creating the policy model.
policy_config: 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: 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: For remote workers, how many workers altogether
have been created?
record_env: Write out episode stats and videos
using gym.wrappers.Monitor to this directory if specified. If
True, use the default output dir in ~/ray_results/.... If
False, do not record anything.
log_dir: Directory where logs can be placed.
log_level: Set the root log level on creation.
callbacks: Custom sub-class of
DefaultCallbacks for training/policy/rollout-worker callbacks.
input_creator: Function that returns an InputReader object for
loading previous generated experiences.
input_evaluation: 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: Function that returns an OutputWriter object for
saving generated experiences.
remote_worker_envs: If using num_envs_per_worker > 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 are expensive to step/reset (e.g., for StarCraft).
Use this cautiously, overheads are significant!
remote_env_batch_wait_ms: 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: Calculate rewards but don't reset the
environment when the horizon is hit.
no_done_at_end: Ignore the done=True at the end of the
episode and instead record done=False.
seed: Set the seed of both np and tf to this value to
to ensure each remote worker has unique exploration behavior.
extra_python_environs: Extra python environments need to be set.
fake_sampler: Use a fake (inf speed) sampler for testing.
spaces: 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.
monitor_path: Obsoleted arg. Use `record_env` instead.
# Deprecated args.
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!"
# Do quick translation into MultiAgentPolicyConfigDict.
if not isinstance(policy_spec, dict):
policy_spec = {
DEFAULT_POLICY_ID: PolicySpec(policy_class=policy_spec)
policy_spec = {
pid: spec if isinstance(spec, PolicySpec) else PolicySpec(*spec)
for pid, spec in policy_spec.copy().items()
if monitor_path is not None:
deprecation_warning("monitor_path", "record_env", error=False)
record_env = monitor_path
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 = 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()):
if log_level:
if worker_index > 1:
disable_log_once_globally() # only need 1 worker to log
elif log_level == "DEBUG":
env_context = EnvContext(
env_config or {},
self.env_context = env_context
self.policy_config: PartialTrainerConfigDict = policy_config
if callbacks:
self.callbacks: "DefaultCallbacks" = callbacks()
from ray.rllib.agents.callbacks import DefaultCallbacks # noqa
self.callbacks: DefaultCallbacks = DefaultCallbacks()
self.worker_index: int = worker_index
self.num_workers: int = num_workers
model_config: ModelConfigDict = \
model_config or self.policy_config.get("model") or {}
# Default policy mapping fn is to always return DEFAULT_POLICY_ID,
# independent on the agent ID and the episode passed in.
self.policy_mapping_fn = \
lambda agent_id, episode, worker, **kwargs: DEFAULT_POLICY_ID
# If provided, set it here.
self.env_creator: EnvCreator = env_creator
self.rollout_fragment_length: int = rollout_fragment_length * num_envs
self.count_steps_by: str = count_steps_by
self.batch_mode: str = batch_mode
self.compress_observations: bool = compress_observations
self.preprocessing_enabled: bool = False \
if policy_config.get("_disable_preprocessor_api") else True
self.observation_filter = observation_filter
self.last_batch: Optional[SampleBatchType] = None
self.global_vars: Optional[dict] = None
self.fake_sampler: bool = fake_sampler
# Update the global seed for numpy/random/tf-eager/torch if we are not
# the local worker, otherwise, this was already done in the Trainer
# object itself.
if self.worker_index > 0:
policy_config.get("framework"), seed)
# A single environment provided by the user (via config.env). This may
# also remain None.
# 1) Create the env using the user provided env_creator. This may
# return a gym.Env (incl. MultiAgentEnv), an already vectorized
# VectorEnv, BaseEnv, ExternalEnv, or an ActorHandle (remote env).
# 2) Wrap - if applicable - with Atari/recording/rendering wrappers.
# 3) Seed the env, if necessary.
# 4) Vectorize the existing single env by creating more clones of
# this env and wrapping it with the RLlib BaseEnv class.
self.env = None
# Create a (single) env for this worker.
if not (worker_index == 0 and num_workers > 0
and not policy_config.get("create_env_on_driver")):
# Run the `env_creator` function passing the EnvContext.
self.env = env_creator(copy.deepcopy(self.env_context))
if self.env is not None:
# Validate environment (general validation function).
_validate_env(self.env, env_context=self.env_context)
# Custom validation function given.
if validate_env is not None:
validate_env(self.env, self.env_context)
# We can't auto-wrap a BaseEnv.
if isinstance(self.env, (BaseEnv,
def wrap(env):
return env
# Atari type env and "deepmind" preprocessor pref.
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
# Framestacking is used.
use_framestack = model_config.get("framestack") is True
def wrap(env):
env = wrap_deepmind(
env = record_env_wrapper(env, record_env, log_dir,
return env
# gym.Env -> Wrap with gym Monitor.
def wrap(env):
return record_env_wrapper(env, record_env, log_dir,
# Wrap env through the correct wrapper.
self.env: EnvType = wrap(self.env)
# Ideally, we would use the same make_sub_env() function below
# to create self.env, but wrap(env) and self.env has a cyclic
# dependency on each other right now, so we would settle on
# duplicating the random seed setting logic for now.
_update_env_seed_if_necessary(self.env, seed, worker_index, 0)
def make_sub_env(vector_index):
# Used to created additional environments during environment
# vectorization.
# Create the env context (config dict + meta-data) for
# this particular sub-env within the vectorized one.
env_ctx = env_context.copy_with_overrides(
# Create the sub-env.
env = env_creator(env_ctx)
# Validate first.
_validate_env(env, env_context=env_ctx)
# Custom validation function given by user.
if validate_env is not None:
validate_env(env, env_ctx)
# Use our wrapper, defined above.
env = wrap(env)
# Make sure a deterministic random seed is set on
# all the sub-environments if specified.
_update_env_seed_if_necessary(env, seed, worker_index,
return env
self.make_sub_env_fn = make_sub_env
self.spaces = spaces
self.policy_dict = _determine_spaces_for_multi_agent_dict(
# List of IDs of those policies, which should be trained.
# By default, these are all policies found in `self.policy_dict`.
self.policies_to_train: List[PolicyID] = policies_to_train or list(
self.policy_map: PolicyMap = None
self.preprocessors: Dict[PolicyID, Preprocessor] = None
# Check available number of GPUs.
num_gpus = policy_config.get("num_gpus", 0) if \
self.worker_index == 0 else \
policy_config.get("num_gpus_per_worker", 0)
# Error if we don't find enough GPUs.
if ray.is_initialized() and \
ray.worker._mode() != ray.worker.LOCAL_MODE and \
not policy_config.get("_fake_gpus"):
devices = []
if policy_config.get("framework") in ["tf2", "tf", "tfe"]:
devices = get_tf_gpu_devices()
elif policy_config.get("framework") == "torch":
devices = list(range(torch.cuda.device_count()))
if len(devices) < num_gpus:
raise RuntimeError(
ERR_MSG_NO_GPUS.format(len(devices), devices) +
# Warn, if running in local-mode and actual GPUs (not faked) are
# requested.
elif ray.is_initialized() and \
ray.worker._mode() == ray.worker.LOCAL_MODE and \
num_gpus > 0 and not policy_config.get("_fake_gpus"):
"You are running ray with `local_mode=True`, but have "
f"configured {num_gpus} GPUs to be used! In local mode, "
f"Policies are placed on the CPU and the `num_gpus` setting "
f"is ignored.")
# Update Policy's view requirements from Model, only if Policy directly
# inherited from base `Policy` class. At this point here, the Policy
# must have it's Model (if any) defined and ready to output an initial
# state.
for pol in self.policy_map.values():
if not pol._model_init_state_automatically_added:
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,
(BaseEnv, ExternalMultiAgentEnv, MultiAgentEnv,
raise ValueError(
f"Have multiple policies {self.policy_map}, but the "
f"env {self.env} is not a subclass of BaseEnv, "
f"MultiAgentEnv, ActorHandle, or ExternalMultiAgentEnv!")
self.filters: Dict[PolicyID, Filter] = {}
for (policy_id, policy) in self.policy_map.items():
filter_shape = tree.map_structure(
lambda s: (None if isinstance( # noqa
s, (Discrete, MultiDiscrete)) else np.array(s.shape)),
self.filters[policy_id] = get_filter(self.observation_filter,
if self.worker_index == 0:"Built filter map: {}".format(self.filters))
# Vectorize environment, if any.
self.num_envs: int = num_envs
# This RolloutWorker has no env.
if self.env is None:
self.async_env = None
# Use a custom env-vectorizer and call it providing self.env.
elif "custom_vector_env" in policy_config:
self.async_env = policy_config["custom_vector_env"](self.env)
# Default: Vectorize self.env via the make_sub_env function. This adds
# further clones of self.env and creates a RLlib BaseEnv (which is
# vectorized under the hood).
# Always use vector env for consistency even if num_envs = 1.
self.async_env: BaseEnv = convert_to_base_env(
# `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
raise ValueError("Unsupported batch mode: {}".format(
# Create the IOContext for this worker.
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":
"Requested 'simulation' input evaluation method: "
"will discard all sampler outputs and keep only metrics.")
sample_async = True
elif method == "is":
ise = ImportanceSamplingEstimator.\
elif method == "wis":
wise = WeightedImportanceSamplingEstimator.\
raise ValueError(
"Unknown evaluation method: {}".format(method))
render = False
if policy_config.get("render_env") is True and \
(num_workers == 0 or worker_index == 1):
render = True
if self.env is None:
self.sampler = None
elif sample_async:
self.sampler = AsyncSampler(
blackhole_outputs="simulation" in input_evaluation,
# Start the Sampler thread.
self.sampler = SyncSampler(
self.input_reader: InputReader = input_creator(self.io_context)
self.output_writer: OutputWriter = output_creator(self.io_context)
"Created rollout worker with env {} ({}), policies {}".format(
self.async_env, self.env, self.policy_map))
def sample(self) -> SampleBatchType:
"""Returns a batch of experience sampled from this worker.
This method must be implemented by subclasses.
A columnar batch of experiences (e.g., tensors).
>>> print(worker.sample())
SampleBatch({"obs": [1, 2, 3], "action": [0, 1, 0], ...})
if self.fake_sampler and self.last_batch is not None:
return self.last_batch
elif self.input_reader is None:
raise ValueError("RolloutWorker has no `input_reader` object! "
"Cannot call `sample()`. You can try setting "
"`create_env_on_driver` to True.")
if log_once("sample_start"):"Generating sample batch of size {}".format(
batches = []
steps_so_far = batches[0].count if \
self.count_steps_by == "env_steps" else \
# In truncate_episodes mode, never pull more than 1 batch per env.
# This avoids over-running the target batch size.
if self.batch_mode == "truncate_episodes":
max_batches = self.num_envs
max_batches = float("inf")
while (steps_so_far < self.rollout_fragment_length
and len(batches) < max_batches):
batch =
steps_so_far += batch.count if \
self.count_steps_by == "env_steps" else \
batch = batches[0].concat_samples(batches) if len(batches) > 1 else \
self.callbacks.on_sample_end(worker=self, samples=batch)
# Always do writes prior to compression for consistency and to allow
# for better compression inside the writer.
# Do off-policy estimation, if needed.
if self.reward_estimators:
for sub_batch in batch.split_by_episode():
for estimator in self.reward_estimators:
if log_once("sample_end"):"Completed sample batch:\n\n{}\n".format(
if self.compress_observations:
batch.compress(bulk=self.compress_observations == "bulk")
if self.fake_sampler:
self.last_batch = batch
return batch
def sample_with_count(self) -> Tuple[SampleBatchType, int]:
"""Same as sample() but returns the count as a separate value.
A columnar batch of experiences (e.g., tensors) and the
size of the collected batch.
>>> print(worker.sample_with_count())
(SampleBatch({"obs": [1, 2, 3], "action": [0, 1, 0], ...}), 3)
batch = self.sample()
return batch, batch.count
def learn_on_batch(self, samples: SampleBatchType) -> Dict:
"""Update policies based on the given batch.
This is the equivalent to apply_gradients(compute_gradients(samples)),
but can be optimized to avoid pulling gradients into CPU memory.
samples: The SampleBatch or MultiAgentBatch to learn on.
Dictionary of extra metadata from compute_gradients().
>>> batch = worker.sample()
>>> info = worker.learn_on_batch(samples)
if log_once("learn_on_batch"):
"Training on concatenated sample batches:\n\n{}\n".format(
if isinstance(samples, MultiAgentBatch):
info_out = {}
builders = {}
to_fetch = {}
for pid, batch in samples.policy_batches.items():
if pid not in self.policies_to_train:
# Decompress SampleBatch, in case some columns are compressed.
policy = self.policy_map[pid]
tf_session = policy.get_session()
if tf_session and hasattr(policy, "_build_learn_on_batch"):
builders[pid] = TFRunBuilder(tf_session, "learn_on_batch")
to_fetch[pid] = policy._build_learn_on_batch(
builders[pid], batch)
info_out[pid] = policy.learn_on_batch(batch)
{pid: builders[pid].get(v)
for pid, v in to_fetch.items()})
info_out = {
if log_once("learn_out"):
logger.debug("Training out:\n\n{}\n".format(summarize(info_out)))
return info_out
def sample_and_learn(self, expected_batch_size: int, num_sgd_iter: int,
sgd_minibatch_size: str,
standardize_fields: List[str]) -> Tuple[dict, int]:
"""Sample and batch and learn on it.
This is typically used in combination with distributed allreduce.
expected_batch_size: Expected number of samples to learn on.
num_sgd_iter: Number of SGD iterations.
sgd_minibatch_size: SGD minibatch size.
standardize_fields: List of sample fields to normalize.
A tuple consisting of a dictionary of extra metadata returned from
the policies' `learn_on_batch()` and the number of samples
learned on.
batch = self.sample()
assert batch.count == expected_batch_size, \
("Batch size possibly out of sync between workers, expected:",
expected_batch_size, "got:", batch.count)"Executing distributed minibatch SGD "
"with epoch size {}, minibatch size {}".format(
batch.count, sgd_minibatch_size))
info = do_minibatch_sgd(batch, self.policy_map, self, num_sgd_iter,
sgd_minibatch_size, standardize_fields)
return info, batch.count
def compute_gradients(
self, samples: SampleBatchType) -> Tuple[ModelGradients, dict]:
"""Returns a gradient computed w.r.t the specified samples.
Uses the Policy's/ies' compute_gradients method(s) to perform the
samples: The SampleBatch or MultiAgentBatch to compute gradients
for using this worker's policies.
In the single-agent case, a tuple consisting of ModelGradients and
info dict of the worker's policy.
In the multi-agent case, a tuple consisting of a dict mapping
PolicyID to ModelGradients and a dict mapping PolicyID to extra
metadata info.
Note that the first return value (grads) can be applied as is to a
compatible worker using the worker's `apply_gradients()` method.
>>> batch = worker.sample()
>>> grads, info = worker.compute_gradients(samples)
if log_once("compute_gradients"):"Compute gradients on:\n\n{}\n".format(
# MultiAgentBatch -> Calculate gradients for all policies.
if isinstance(samples, MultiAgentBatch):
grad_out, info_out = {}, {}
if self.policy_config.get("framework") == "tf":
for pid, batch in samples.policy_batches.items():
if pid not in self.policies_to_train:
policy = self.policy_map[pid]
builder = TFRunBuilder(policy.get_session(),
grad_out[pid], info_out[pid] = (
policy._build_compute_gradients(builder, batch))
grad_out = {k: builder.get(v) for k, v in grad_out.items()}
info_out = {k: builder.get(v) for k, v in info_out.items()}
for pid, batch in samples.policy_batches.items():
if pid not in self.policies_to_train:
grad_out[pid], info_out[pid] = (
# SampleBatch -> Calculate gradients for the default policy.
grad_out, info_out = (
info_out["batch_count"] = samples.count
if log_once("grad_out"):"Compute grad info:\n\n{}\n".format(
return grad_out, info_out
def apply_gradients(
grads: Union[ModelGradients, Dict[PolicyID, ModelGradients]],
) -> None:
"""Applies the given gradients to this worker's models.
Uses the Policy's/ies' apply_gradients method(s) to perform the
grads: Single ModelGradients (single-agent case) or a dict
mapping PolicyIDs to the respective model gradients
>>> samples = worker.sample()
>>> grads, info = worker.compute_gradients(samples)
>>> worker.apply_gradients(grads)
if log_once("apply_gradients"):"Apply gradients:\n\n{}\n".format(summarize(grads)))
# Grads is a dict (mapping PolicyIDs to ModelGradients).
# Multi-agent case.
if isinstance(grads, dict):
for pid, g in grads.items():
if pid in self.policies_to_train:
# Grads is a ModelGradients type. Single-agent case.
elif DEFAULT_POLICY_ID in self.policies_to_train:
def get_metrics(self) -> List[Union[RolloutMetrics, OffPolicyEstimate]]:
"""Returns the thus-far collected metrics from this worker's rollouts.
List of RolloutMetrics and/or OffPolicyEstimate objects
collected thus-far.
# Get metrics from sampler (if any).
if self.sampler is not None:
out = self.sampler.get_metrics()
out = []
# Get metrics from our reward-estimators (if any).
for m in self.reward_estimators:
return out
def foreach_env(self, func: Callable[[EnvType], T]) -> List[T]:
"""Calls the given function with each sub-environment as arg.
func: The function to call for each underlying
sub-environment (as only arg).
The list of return values of all calls to `func([env])`.
if self.async_env is None:
return []
envs = self.async_env.get_sub_environments()
# Empty list (not implemented): Call function directly on the
# BaseEnv.
if not envs:
return [func(self.async_env)]
# Call function on all underlying (vectorized) sub environments.
return [func(e) for e in envs]
def foreach_env_with_context(
self, func: Callable[[EnvType, EnvContext], T]) -> List[T]:
"""Calls given function with each sub-env plus env_ctx as args.
func: The function to call for each underlying
sub-environment and its EnvContext (as the args).
The list of return values of all calls to `func([env, ctx])`.
if self.async_env is None:
return []
envs = self.async_env.get_sub_environments()
# Empty list (not implemented): Call function directly on the
# BaseEnv.
if not envs:
return [func(self.async_env, self.env_context)]
# Call function on all underlying (vectorized) sub environments.
ret = []
for i, e in enumerate(envs):
ctx = self.env_context.copy_with_overrides(vector_index=i)
ret.append(func(e, ctx))
return ret
def get_policy(self, policy_id: PolicyID = DEFAULT_POLICY_ID) -> \
"""Return policy for the specified id, or None.
policy_id: ID of the policy to return. None for DEFAULT_POLICY_ID
(in the single agent case).
The policy under the given ID (or None if not found).
return self.policy_map.get(policy_id)
def add_policy(
policy_id: PolicyID,
policy_cls: Type[Policy],
observation_space: Optional[Space] = None,
action_space: Optional[Space] = None,
config: Optional[PartialTrainerConfigDict] = None,
policy_state: Optional[PolicyState] = None,
policy_mapping_fn: Optional[Callable[[AgentID, "Episode"],
PolicyID]] = None,
policies_to_train: Optional[List[PolicyID]] = None,
) -> Policy:
"""Adds a new policy to this RolloutWorker.
policy_id: ID of the policy to add.
policy_cls: The Policy class to use for constructing the new
observation_space: The observation space of the policy to add.
action_space: The action space of the policy to add.
config: The config overrides for the policy to add.
policy_state: Optional state dict to apply to the new
policy instance, right after its construction.
policy_mapping_fn: An optional (updated) policy mapping function
to use from here on. Note that already ongoing episodes will
not change their mapping but will use the old mapping till
the end of the episode.
policies_to_train: An optional list of policy IDs to be trained.
If None, will keep the existing list in place. Policies,
whose IDs are not in the list will not be updated.
The newly added policy.
KeyError: If the given `policy_id` already exists in this worker's
if policy_id in self.policy_map:
raise KeyError(f"Policy ID '{policy_id}' already in policy map!")
policy_dict_to_add = _determine_spaces_for_multi_agent_dict(
policy_id: PolicySpec(policy_cls, observation_space,
action_space, config or {})
new_policy = self.policy_map[policy_id]
# Set the state of the newly created policy.
if policy_state:
self.filters[policy_id] = get_filter(
self.observation_filter, new_policy.observation_space.shape)
return new_policy
def remove_policy(
policy_id: PolicyID = DEFAULT_POLICY_ID,
policy_mapping_fn: Optional[Callable[[AgentID], PolicyID]] = None,
policies_to_train: Optional[List[PolicyID]] = None,
) -> None:
"""Removes a policy from this RolloutWorker.
policy_id: ID of the policy to be removed. None for
policy_mapping_fn: An optional (updated) policy mapping function
to use from here on. Note that already ongoing episodes will
not change their mapping but will use the old mapping till
the end of the episode.
policies_to_train: An optional list of policy IDs to be trained.
If None, will keep the existing list in place. Policies,
whose IDs are not in the list will not be updated.
if policy_id not in self.policy_map:
raise ValueError(f"Policy ID '{policy_id}' not in policy map!")
del self.policy_map[policy_id]
del self.preprocessors[policy_id]
def set_policy_mapping_fn(
policy_mapping_fn: Optional[Callable[[AgentID, "Episode"],
PolicyID]] = None,
) -> None:
"""Sets `self.policy_mapping_fn` to a new callable (if provided).
policy_mapping_fn: The new mapping function to use. If None,
will keep the existing mapping function in place.
if policy_mapping_fn is not None:
self.policy_mapping_fn = policy_mapping_fn
if not callable(self.policy_mapping_fn):
raise ValueError("`policy_mapping_fn` must be a callable!")
def set_policies_to_train(
self, policies_to_train: Optional[List[PolicyID]] = None) -> None:
"""Sets `self.policies_to_train` to a new list of PolicyIDs.
policies_to_train: The new list of policy IDs to train with.
If None, will keep the existing list in place.
if policies_to_train is not None:
self.policies_to_train = policies_to_train
def for_policy(self,
func: Callable[[Policy, Optional[Any]], T],
policy_id: Optional[PolicyID] = DEFAULT_POLICY_ID,
**kwargs) -> T:
"""Calls the given function with the specified policy as first arg.
func: The function to call with the policy as first arg.
policy_id: The PolicyID of the policy to call the function with.
Keyword Args:
kwargs: Additional kwargs to be passed to the call.
The return value of the function call.
return func(self.policy_map[policy_id], **kwargs)
def foreach_policy(self,
func: Callable[[Policy, PolicyID, Optional[Any]], T],
**kwargs) -> List[T]:
"""Calls the given function with each (policy, policy_id) tuple.
func: The function to call with each (policy, policy ID) tuple.
Keyword Args:
kwargs: Additional kwargs to be passed to the call.
The list of return values of all calls to
`func([policy, pid, **kwargs])`.
return [
func(policy, pid, **kwargs)
for pid, policy in self.policy_map.items()
def foreach_trainable_policy(
self, func: Callable[[Policy, PolicyID, Optional[Any]], T],
**kwargs) -> List[T]:
Calls the given function with each (policy, policy_id) tuple.
Only those policies/IDs will be called on, which can be found in
func: The function to call with each (policy, policy ID) tuple,
for only those policies that are in `self.policies_to_train`.
Keyword Args:
kwargs: Additional kwargs to be passed to the call.
The list of return values of all calls to
`func([policy, pid, **kwargs])`.
return [
func(policy, pid, **kwargs)
for pid, policy in self.policy_map.items()
if pid in self.policies_to_train
def sync_filters(self, new_filters: dict) -> None:
"""Changes self's filter to given and rebases any accumulated delta.
new_filters: Filters with new state to update local copy.
assert all(k in new_filters for k in self.filters)
for k in self.filters:
def get_filters(self, flush_after: bool = False) -> Dict:
"""Returns a snapshot of filters.
flush_after: Clears the filter buffer state.
Dict for serializable filters
return_filters = {}
for k, f in self.filters.items():
return_filters[k] = f.as_serializable()
if flush_after:
return return_filters
def save(self) -> bytes:
"""Serializes this RolloutWorker's current state and returns it.
The current state of this RolloutWorker as a serialized, pickled
byte sequence.
filters = self.get_filters(flush_after=True)
state = {}
policy_specs = {}
for pid in self.policy_map:
state[pid] = self.policy_map[pid].get_state()
policy_specs[pid] = self.policy_map.policy_specs[pid]
return pickle.dumps({
"filters": filters,
"state": state,
"policy_specs": policy_specs,
def restore(self, objs: bytes) -> None:
"""Restores this RolloutWorker's state from a sequence of bytes.
objs: The byte sequence to restore this worker's state from.
>>> state =
>>> new_worker = RolloutWorker(...)
>>> new_worker.restore(state)
objs = pickle.loads(objs)
for pid, state in objs["state"].items():
if pid not in self.policy_map:
pol_spec = objs.get("policy_specs", {}).get(pid)
if not pol_spec:
f"PolicyID '{pid}' was probably added on-the-fly (not"
" part of the static `multagent.policies` config) and"
" no PolicySpec objects found in the pickled policy "
"state. Will not add `{pid}`, but ignore it for now.")
def get_weights(
policies: Optional[List[PolicyID]] = None,
) -> Dict[PolicyID, ModelWeights]:
"""Returns each policies' model weights of this worker.
policies: List of PolicyIDs to get the weights from.
Use None for all policies.
Dict mapping PolicyIDs to ModelWeights.
>>> weights = worker.get_weights()
>>> print(weights)
{"default_policy": {"layer1": array(...), "layer2": ...}}
if policies is None:
policies = list(self.policy_map.keys())
policies = force_list(policies)
return {
pid: policy.get_weights()
for pid, policy in self.policy_map.items() if pid in policies
def set_weights(self,
weights: Dict[PolicyID, ModelWeights],
global_vars: Optional[Dict] = None) -> None:
"""Sets each policies' model weights of this worker.
weights: Dict mapping PolicyIDs to the new weights to be used.
global_vars: An optional global vars dict to set this
worker to. If None, do not update the global_vars.
>>> weights = worker.get_weights()
>>> # Set `global_vars` (timestep) as well.
>>> worker.set_weights(weights, {"timestep": 42})
# If per-policy weights are object refs, `ray.get()` them first.
if weights and isinstance(next(iter(weights.values())), ObjectRef):
actual_weights = ray.get(list(weights.values()))
weights = {
pid: actual_weights[i]
for i, pid in enumerate(weights.keys())
for pid, w in weights.items():
if global_vars:
def get_global_vars(self) -> dict:
"""Returns the current global_vars dict of this worker.
The current global_vars dict of this worker.
>>> global_vars = worker.get_global_vars()
>>> print(global_vars)
{"timestep": 424242}
return self.global_vars
def set_global_vars(self, global_vars: dict) -> None:
"""Updates this worker's and all its policies' global vars.
global_vars: The new global_vars dict.
>>> global_vars = worker.set_global_vars({"timestep": 4242})
self.foreach_policy(lambda p, _: p.on_global_var_update(global_vars))
self.global_vars = global_vars
def stop(self) -> None:
"""Releases all resources used by this RolloutWorker."""
# If we have an env -> Release its resources.
if self.env is not None:
# Close all policies' sessions (if tf static graph).
for policy in self.policy_map.values():
sess = policy.get_session()
# Closes the tf session, if any.
if sess is not None:
def apply(
func: Callable[["RolloutWorker", Optional[Any], Optional[Any]], T],
*args, **kwargs) -> T:
"""Calls the given function with this rollout worker instance.
Useful for when the RolloutWorker class has been converted into a
ActorHandle and the user needs to execute some functionality (e.g.
add a property) on the underlying policy object.
func: The function to call, with this RolloutWorker as first
argument, followed by args, and kwargs.
args: Optional additional args to pass to the function call.
kwargs: Optional additional kwargs to pass to the function call.
The return value of the function call.
return func(self, *args, **kwargs)
def setup_torch_data_parallel(self, url: str, world_rank: int,
world_size: int, backend: str) -> None:
"""Join a torch process group for distributed SGD.""""Joining process group, url={}, world_rank={}, "
"world_size={}, backend={}".format(url, world_rank,
world_size, backend))
for pid, policy in self.policy_map.items():
if not isinstance(policy, TorchPolicy):
raise ValueError(
"This policy does not support torch distributed", policy)
policy.distributed_world_size = world_size
def creation_args(self) -> dict:
"""Returns the kwargs dict used to create this worker."""
return self._original_kwargs
def get_host(self) -> str:
"""Returns the hostname of the process running this evaluator."""
return platform.node()
def get_node_ip(self) -> str:
"""Returns the IP address of the node that this worker runs on."""
return ray.util.get_node_ip_address()
def find_free_port(self) -> int:
"""Finds a free port on the node that this worker runs on."""
from ray.util.ml_utils.util import find_free_port
return find_free_port()
def __del__(self):
"""If this worker is deleted, clears all resources used by it."""
# In case we have-an AsyncSampler, kill its sampling thread.
if hasattr(self, "sampler") and isinstance(self.sampler, AsyncSampler):
self.sampler.shutdown = True
def _build_policy_map(
policy_dict: MultiAgentPolicyConfigDict,
policy_config: PartialTrainerConfigDict,
session_creator: Optional[Callable[[], "tf1.Session"]] = None,
seed: Optional[int] = None,
) -> None:
"""Adds the given policy_dict to `self.policy_map`.
policy_dict: The MultiAgentPolicyConfigDict to be added to this
worker's PolicyMap.
policy_config: The general policy config to use. May be updated
by individual policy condig overrides in the given
multi-agent `policy_dict`.
session_creator: A callable that creates a tf session
(if applicable).
seed: An optional random seed to pass to PolicyMap's
ma_config = policy_config.get("multiagent", {})
# If our policy_map does not exist yet, create it here.
self.policy_map = self.policy_map or PolicyMap(
# If our preprocessors dict does not exist yet, create it here.
self.preprocessors = self.preprocessors or {}
# Loop through given policy-dict and add each entry to our map.
for name, (orig_cls, obs_space, act_space,
conf) in sorted(policy_dict.items()):
logger.debug("Creating policy for {}".format(name))
# Update the general policy_config with the specific config
# for this particular policy.
merged_conf = merge_dicts(policy_config, conf or {})
# Update num_workers and worker_index.
merged_conf["num_workers"] = self.num_workers
merged_conf["worker_index"] = self.worker_index
# Preprocessors.
if self.preprocessing_enabled:
preprocessor = ModelCatalog.get_preprocessor_for_space(
obs_space, merged_conf.get("model"))
self.preprocessors[name] = preprocessor
if preprocessor is not None:
obs_space = preprocessor.observation_space
self.preprocessors[name] = None
# Create the actual policy object.
self.policy_map.create_policy(name, orig_cls, obs_space, act_space,
conf, merged_conf)
if self.worker_index == 0:"Built policy map: {self.policy_map}")"Built preprocessor map: {self.preprocessors}")
new="Trainer.get_policy().export_model([export_dir], [onnx]?)",
def export_policy_model(self,
export_dir: str,
policy_id: PolicyID = DEFAULT_POLICY_ID,
onnx: Optional[int] = None):
self.policy_map[policy_id].export_model(export_dir, onnx=onnx)
def import_policy_model_from_h5(self,
import_file: str,
policy_id: PolicyID = DEFAULT_POLICY_ID):
new="Trainer.get_policy().export_checkpoint([export_dir], "
def export_policy_checkpoint(self,
export_dir: str,
filename_prefix: str = "model",
policy_id: PolicyID = DEFAULT_POLICY_ID):
def _determine_spaces_for_multi_agent_dict(
multi_agent_dict: MultiAgentPolicyConfigDict,
env: Optional[EnvType] = None,
spaces: Optional[Dict[PolicyID, Tuple[Space, Space]]] = None,
policy_config: Optional[PartialTrainerConfigDict] = None,
) -> MultiAgentPolicyConfigDict:
policy_config = policy_config or {}
# Try extracting spaces from env or from given spaces dict.
env_obs_space = None
env_act_space = None
# Env is a ray.remote: Get spaces via its (automatically added)
# `_get_spaces()` method.
if isinstance(env,
env_obs_space, env_act_space = ray.get(env._get_spaces.remote())
# Normal env (gym.Env or MultiAgentEnv): These should have the
# `observation_space` and `action_space` properties.
elif env is not None:
if hasattr(env, "observation_space") and isinstance(
env.observation_space, gym.Space):
env_obs_space = env.observation_space
if hasattr(env, "action_space") and isinstance(env.action_space,
env_act_space = env.action_space
# Last resort: Try getting the env's spaces from the spaces
# dict's special __env__ key.
if spaces is not None:
if env_obs_space is None:
env_obs_space = spaces.get("__env__", [None])[0]
if env_act_space is None:
env_act_space = spaces.get("__env__", [None, None])[1]
for pid, policy_spec in multi_agent_dict.copy().items():
if policy_spec.observation_space is None:
if spaces is not None and pid in spaces:
obs_space = spaces[pid][0]
elif env_obs_space is not None:
obs_space = env_obs_space
elif policy_config.get("observation_space"):
obs_space = policy_config["observation_space"]
raise ValueError(
"`observation_space` not provided in PolicySpec for "
f"{pid} and env does not have an observation space OR "
"no spaces received from other workers' env(s) OR no "
"`observation_space` specified in config!")
multi_agent_dict[pid] = multi_agent_dict[pid]._replace(
if policy_spec.action_space is None:
if spaces is not None and pid in spaces:
act_space = spaces[pid][1]
elif env_act_space is not None:
act_space = env_act_space
elif policy_config.get("action_space"):
act_space = policy_config["action_space"]
raise ValueError(
"`action_space` not provided in PolicySpec for "
f"{pid} and env does not have an action space OR "
"no spaces received from other workers' env(s) OR no "
"`action_space` specified in config!")
multi_agent_dict[pid] = multi_agent_dict[pid]._replace(
return multi_agent_dict
def _validate_env(env: EnvType, env_context: EnvContext = None):
# Base message for checking the env for vector-index=0
msg = f"Validating sub-env at vector index={env_context.vector_index} ..."
allowed_types = [
gym.Env, ExternalEnv, VectorEnv, BaseEnv,
if not any(isinstance(env, tpe) for tpe in allowed_types):
# Allow this as a special case (assumed gym.Env).
# TODO: Disallow this early-out. Everything should conform to a few
# supported classes, i.e. gym.Env/MultiAgentEnv/etc...
if hasattr(env, "observation_space") and hasattr(env, "action_space"):
logger.warning(msg + f" (warning; invalid env-type={type(env)})")
logger.warning(msg + " (NOT OK)")
raise EnvError(
"Returned env should be an instance of gym.Env (incl. "
"MultiAgentEnv), ExternalEnv, VectorEnv, or BaseEnv. "
f"The provided env creator function returned {env} "
# Do some test runs with the provided env.
if isinstance(env, gym.Env) and not isinstance(env, MultiAgentEnv):
# Make sure the gym.Env has the two space attributes properly set.
assert hasattr(env, "observation_space") and hasattr(
env, "action_space")
# Get a dummy observation by resetting the env.
dummy_obs = env.reset()
# Convert lists to np.ndarrays.
if type(dummy_obs) is list and isinstance(env.observation_space, Box):
dummy_obs = np.array(dummy_obs)
# Ignore float32/float64 diffs.
if isinstance(env.observation_space, Box) and \
env.observation_space.dtype != dummy_obs.dtype:
dummy_obs = dummy_obs.astype(env.observation_space.dtype)
# Check, if observation is ok (part of the observation space). If not,
# error.
if not env.observation_space.contains(dummy_obs):
logger.warning(msg + " (NOT OK)")
raise EnvError(
f"Env's `observation_space` {env.observation_space} does not "
f"contain returned observation after a reset ({dummy_obs})!")
# Log that everything is ok. + " (ok)")