from collections import defaultdict
import concurrent
import copy
from datetime import datetime
import functools
import gymnasium as gym
import importlib
import json
import logging
import numpy as np
import os
from packaging import version
import pkg_resources
import re
import tempfile
import time
import tree # pip install dm_tree
from typing import (
Callable,
Container,
DefaultDict,
Dict,
List,
Optional,
Set,
Tuple,
Type,
Union,
)
import ray
from ray._private.usage.usage_lib import TagKey, record_extra_usage_tag
from ray.actor import ActorHandle
from ray.air.checkpoint import Checkpoint
import ray.cloudpickle as pickle
from ray.rllib.algorithms.algorithm_config import AlgorithmConfig
from ray.rllib.algorithms.registry import ALGORITHMS_CLASS_TO_NAME as ALL_ALGORITHMS
from ray.rllib.connectors.agent.obs_preproc import ObsPreprocessorConnector
from ray.rllib.core.rl_module.rl_module import SingleAgentRLModuleSpec
from ray.rllib.env.env_context import EnvContext
from ray.rllib.env.utils import _gym_env_creator
from ray.rllib.evaluation.episode import Episode
from ray.rllib.evaluation.metrics import (
collect_episodes,
collect_metrics,
summarize_episodes,
)
from ray.rllib.evaluation.rollout_worker import RolloutWorker
from ray.rllib.evaluation.worker_set import WorkerSet
from ray.rllib.execution.common import (
STEPS_TRAINED_THIS_ITER_COUNTER, # TODO: Backward compatibility.
)
from ray.rllib.execution.rollout_ops import synchronous_parallel_sample
from ray.rllib.execution.train_ops import multi_gpu_train_one_step, train_one_step
from ray.rllib.offline import get_dataset_and_shards
from ray.rllib.offline.estimators import (
OffPolicyEstimator,
ImportanceSampling,
WeightedImportanceSampling,
DirectMethod,
DoublyRobust,
)
from ray.rllib.offline.offline_evaluation_utils import remove_time_dim
from ray.rllib.offline.offline_evaluator import OfflineEvaluator
from ray.rllib.policy.policy import Policy
from ray.rllib.policy.sample_batch import DEFAULT_POLICY_ID, SampleBatch, concat_samples
from ray.rllib.utils import deep_update, FilterManager
from ray.rllib.utils.annotations import (
DeveloperAPI,
ExperimentalAPI,
OverrideToImplementCustomLogic,
OverrideToImplementCustomLogic_CallToSuperRecommended,
PublicAPI,
override,
)
from ray.rllib.utils.checkpoints import (
CHECKPOINT_VERSION,
CHECKPOINT_VERSION_LEARNER,
get_checkpoint_info,
try_import_msgpack,
)
from ray.rllib.utils.debug import update_global_seed_if_necessary
from ray.rllib.utils.deprecation import (
DEPRECATED_VALUE,
Deprecated,
deprecation_warning,
)
from ray.rllib.utils.error import ERR_MSG_INVALID_ENV_DESCRIPTOR, EnvError
from ray.rllib.utils.framework import try_import_tf
from ray.rllib.utils.from_config import from_config
from ray.rllib.utils.metrics import (
NUM_AGENT_STEPS_SAMPLED,
NUM_AGENT_STEPS_SAMPLED_THIS_ITER,
NUM_AGENT_STEPS_TRAINED,
NUM_ENV_STEPS_SAMPLED,
NUM_ENV_STEPS_SAMPLED_THIS_ITER,
NUM_ENV_STEPS_TRAINED,
SYNCH_WORKER_WEIGHTS_TIMER,
TRAINING_ITERATION_TIMER,
SAMPLE_TIMER,
)
from ray.rllib.utils.metrics.learner_info import LEARNER_INFO
from ray.rllib.utils.policy import validate_policy_id
from ray.rllib.utils.replay_buffers import MultiAgentReplayBuffer, ReplayBuffer
from ray.rllib.utils.serialization import deserialize_type, NOT_SERIALIZABLE
from ray.rllib.utils.spaces import space_utils
from ray.rllib.utils.typing import (
AgentConnectorDataType,
AgentID,
AlgorithmConfigDict,
EnvCreator,
EnvInfoDict,
EnvType,
EpisodeID,
PartialAlgorithmConfigDict,
PolicyID,
PolicyState,
ResultDict,
SampleBatchType,
TensorStructType,
TensorType,
)
from ray.tune.execution.placement_groups import PlacementGroupFactory
from ray.tune.experiment.trial import ExportFormat
from ray.tune.logger import Logger, UnifiedLogger
from ray.tune.registry import ENV_CREATOR, _global_registry
from ray.tune.resources import Resources
from ray.tune.result import DEFAULT_RESULTS_DIR
from ray.tune.trainable import Trainable
from ray.util import log_once
from ray.util.timer import _Timer
from ray.tune.registry import get_trainable_cls
tf1, tf, tfv = try_import_tf()
logger = logging.getLogger(__name__)
@Deprecated(
new="config = AlgorithmConfig().update_from_dict({'a': 1, 'b': 2}); ... ; "
"print(config.lr) -> 0.001; if config.a > 0: [do something];",
error=True,
)
def with_common_config(*args, **kwargs):
pass
[docs]@PublicAPI
class Algorithm(Trainable):
"""An RLlib algorithm responsible for optimizing one or more Policies.
Algorithms contain a WorkerSet under `self.workers`. A WorkerSet is
normally composed of a single local worker
(self.workers.local_worker()), used to compute and apply learning updates,
and optionally one or more remote workers used to generate environment
samples in parallel.
WorkerSet is fault tolerant and elastic. It tracks health states for all
the managed remote worker actors. As a result, Algorithm should never
access the underlying actor handles directly. Instead, always access them
via all the foreach APIs with assigned IDs of the underlying workers.
Each worker (remotes or local) contains a PolicyMap, which itself
may contain either one policy for single-agent training or one or more
policies for multi-agent training. Policies are synchronized
automatically from time to time using ray.remote calls. The exact
synchronization logic depends on the specific algorithm used,
but this usually happens from local worker to all remote workers and
after each training update.
You can write your own Algorithm classes by sub-classing from `Algorithm`
or any of its built-in sub-classes.
This allows you to override the `training_step` method to implement
your own algorithm logic. You can find the different built-in
algorithms' `training_step()` methods in their respective main .py files,
e.g. rllib.algorithms.dqn.dqn.py or rllib.algorithms.impala.impala.py.
The most important API methods a Algorithm exposes are `train()`,
`evaluate()`, `save()` and `restore()`.
"""
# Whether to allow unknown top-level config keys.
_allow_unknown_configs = False
# List of top-level keys with value=dict, for which new sub-keys are
# allowed to be added to the value dict.
_allow_unknown_subkeys = [
"tf_session_args",
"local_tf_session_args",
"env_config",
"model",
"optimizer",
"multiagent",
"custom_resources_per_worker",
"evaluation_config",
"exploration_config",
"replay_buffer_config",
"extra_python_environs_for_worker",
"input_config",
"output_config",
]
# List of top level keys with value=dict, for which we always override the
# entire value (dict), iff the "type" key in that value dict changes.
_override_all_subkeys_if_type_changes = [
"exploration_config",
"replay_buffer_config",
]
# List of keys that are always fully overridden if present in any dict or sub-dict
_override_all_key_list = ["off_policy_estimation_methods", "policies"]
_progress_metrics = (
"num_env_steps_sampled",
"num_env_steps_trained",
"episodes_total",
"sampler_results/episode_len_mean",
"sampler_results/episode_reward_mean",
"evaluation/sampler_results/episode_reward_mean",
)
[docs] @staticmethod
def from_checkpoint(
checkpoint: Union[str, Checkpoint],
policy_ids: Optional[Container[PolicyID]] = None,
policy_mapping_fn: Optional[Callable[[AgentID, EpisodeID], PolicyID]] = None,
policies_to_train: Optional[
Union[
Container[PolicyID],
Callable[[PolicyID, Optional[SampleBatchType]], bool],
]
] = None,
) -> "Algorithm":
"""Creates a new algorithm instance from a given checkpoint.
Note: This method must remain backward compatible from 2.0.0 on.
Args:
checkpoint: The path (str) to the checkpoint directory to use
or an AIR Checkpoint instance to restore from.
policy_ids: Optional list of PolicyIDs to recover. This allows users to
restore an Algorithm with only a subset of the originally present
Policies.
policy_mapping_fn: An optional (updated) policy mapping function
to use from here on.
policies_to_train: An optional list of policy IDs to be trained
or a callable taking PolicyID and SampleBatchType and
returning a bool (trainable or not?).
If None, will keep the existing setup in place. Policies,
whose IDs are not in the list (or for which the callable
returns False) will not be updated.
Returns:
The instantiated Algorithm.
"""
checkpoint_info = get_checkpoint_info(checkpoint)
# Not possible for (v0.1) (algo class and config information missing
# or very hard to retrieve).
if checkpoint_info["checkpoint_version"] == version.Version("0.1"):
raise ValueError(
"Cannot restore a v0 checkpoint using `Algorithm.from_checkpoint()`!"
"In this case, do the following:\n"
"1) Create a new Algorithm object using your original config.\n"
"2) Call the `restore()` method of this algo object passing it"
" your checkpoint dir or AIR Checkpoint object."
)
elif checkpoint_info["checkpoint_version"] < version.Version("1.0"):
raise ValueError(
"`checkpoint_info['checkpoint_version']` in `Algorithm.from_checkpoint"
"()` must be 1.0 or later! You are using a checkpoint with "
f"version v{checkpoint_info['checkpoint_version']}."
)
# This is a msgpack checkpoint.
if checkpoint_info["format"] == "msgpack":
# User did not provide unserializable function with this call
# (`policy_mapping_fn`). Note that if `policies_to_train` is None, it
# defaults to training all policies (so it's ok to not provide this here).
if policy_mapping_fn is None:
# Only DEFAULT_POLICY_ID present in this algorithm, provide default
# implementations of these two functions.
if checkpoint_info["policy_ids"] == {DEFAULT_POLICY_ID}:
policy_mapping_fn = AlgorithmConfig.DEFAULT_POLICY_MAPPING_FN
# Provide meaningful error message.
else:
raise ValueError(
"You are trying to restore a multi-agent algorithm from a "
"`msgpack` formatted checkpoint, which do NOT store the "
"`policy_mapping_fn` or `policies_to_train` "
"functions! Make sure that when using the "
"`Algorithm.from_checkpoint()` utility, you also pass the "
"args: `policy_mapping_fn` and `policies_to_train` with your "
"call. You might leave `policies_to_train=None` in case "
"you would like to train all policies anyways."
)
state = Algorithm._checkpoint_info_to_algorithm_state(
checkpoint_info=checkpoint_info,
policy_ids=policy_ids,
policy_mapping_fn=policy_mapping_fn,
policies_to_train=policies_to_train,
)
return Algorithm.from_state(state)
[docs] @staticmethod
def from_state(state: Dict) -> "Algorithm":
"""Recovers an Algorithm from a state object.
The `state` of an instantiated Algorithm can be retrieved by calling its
`get_state` method. It contains all information necessary
to create the Algorithm from scratch. No access to the original code (e.g.
configs, knowledge of the Algorithm's class, etc..) is needed.
Args:
state: The state to recover a new Algorithm instance from.
Returns:
A new Algorithm instance.
"""
algorithm_class: Type[Algorithm] = state.get("algorithm_class")
if algorithm_class is None:
raise ValueError(
"No `algorithm_class` key was found in given `state`! "
"Cannot create new Algorithm."
)
# algo_class = get_trainable_cls(algo_class_name)
# Create the new algo.
config = state.get("config")
if not config:
raise ValueError("No `config` found in given Algorithm state!")
new_algo = algorithm_class(config=config)
# Set the new algo's state.
new_algo.__setstate__(state)
# Return the new algo.
return new_algo
[docs] @PublicAPI
def __init__(
self,
config: Optional[AlgorithmConfig] = None,
env=None, # deprecated arg
logger_creator: Optional[Callable[[], Logger]] = None,
**kwargs,
):
"""Initializes an Algorithm instance.
Args:
config: Algorithm-specific configuration object.
logger_creator: Callable that creates a ray.tune.Logger
object. If unspecified, a default logger is created.
**kwargs: Arguments passed to the Trainable base class.
"""
config = config or self.get_default_config()
# Translate possible dict into an AlgorithmConfig object, as well as,
# resolving generic config objects into specific ones (e.g. passing
# an `AlgorithmConfig` super-class instance into a PPO constructor,
# which normally would expect a PPOConfig object).
if isinstance(config, dict):
default_config = self.get_default_config()
# `self.get_default_config()` also returned a dict ->
# Last resort: Create core AlgorithmConfig from merged dicts.
if isinstance(default_config, dict):
config = AlgorithmConfig.from_dict(
config_dict=self.merge_trainer_configs(default_config, config, True)
)
# Default config is an AlgorithmConfig -> update its properties
# from the given config dict.
else:
config = default_config.update_from_dict(config)
else:
default_config = self.get_default_config()
# Given AlgorithmConfig is not of the same type as the default config:
# This could be the case e.g. if the user is building an algo from a
# generic AlgorithmConfig() object.
if not isinstance(config, type(default_config)):
config = default_config.update_from_dict(config.to_dict())
# In case this algo is using a generic config (with no algo_class set), set it
# here.
if config.algo_class is None:
config.algo_class = type(self)
if env is not None:
deprecation_warning(
old=f"algo = Algorithm(env='{env}', ...)",
new=f"algo = AlgorithmConfig().environment('{env}').build()",
error=False,
)
config.environment(env)
# Validate and freeze our AlgorithmConfig object (no more changes possible).
config.validate()
config.freeze()
# Convert `env` provided in config into a concrete env creator callable, which
# takes an EnvContext (config dict) as arg and returning an RLlib supported Env
# type (e.g. a gym.Env).
self._env_id, self.env_creator = self._get_env_id_and_creator(
config.env, config
)
env_descr = (
self._env_id.__name__ if isinstance(self._env_id, type) else self._env_id
)
# Placeholder for a local replay buffer instance.
self.local_replay_buffer = None
# Create a default logger creator if no logger_creator is specified
if logger_creator is None:
# Default logdir prefix containing the agent's name and the
# env id.
timestr = datetime.today().strftime("%Y-%m-%d_%H-%M-%S")
env_descr_for_dir = re.sub("[/\\\\]", "-", str(env_descr))
logdir_prefix = f"{str(self)}_{env_descr_for_dir}_{timestr}"
if not os.path.exists(DEFAULT_RESULTS_DIR):
# Possible race condition if dir is created several times on
# rollout workers
os.makedirs(DEFAULT_RESULTS_DIR, exist_ok=True)
logdir = tempfile.mkdtemp(prefix=logdir_prefix, dir=DEFAULT_RESULTS_DIR)
# Allow users to more precisely configure the created logger
# via "logger_config.type".
if config.logger_config and "type" in config.logger_config:
def default_logger_creator(config):
"""Creates a custom logger with the default prefix."""
cfg = config["logger_config"].copy()
cls = cfg.pop("type")
# Provide default for logdir, in case the user does
# not specify this in the "logger_config" dict.
logdir_ = cfg.pop("logdir", logdir)
return from_config(cls=cls, _args=[cfg], logdir=logdir_)
# If no `type` given, use tune's UnifiedLogger as last resort.
else:
def default_logger_creator(config):
"""Creates a Unified logger with the default prefix."""
return UnifiedLogger(config, logdir, loggers=None)
logger_creator = default_logger_creator
# Metrics-related properties.
self._timers = defaultdict(_Timer)
self._counters = defaultdict(int)
self._episode_history = []
self._episodes_to_be_collected = []
# The fully qualified AlgorithmConfig used for evaluation
# (or None if evaluation not setup).
self.evaluation_config: Optional[AlgorithmConfig] = None
# Evaluation WorkerSet and metrics last returned by `self.evaluate()`.
self.evaluation_workers: Optional[WorkerSet] = None
# Initialize common evaluation_metrics to nan, before they become
# available. We want to make sure the metrics are always present
# (although their values may be nan), so that Tune does not complain
# when we use these as stopping criteria.
self.evaluation_metrics = {
# TODO: Don't dump sampler results into top-level.
"evaluation": {
"episode_reward_max": np.nan,
"episode_reward_min": np.nan,
"episode_reward_mean": np.nan,
"sampler_results": {
"episode_reward_max": np.nan,
"episode_reward_min": np.nan,
"episode_reward_mean": np.nan,
},
},
}
super().__init__(
config=config,
logger_creator=logger_creator,
**kwargs,
)
# Check, whether `training_iteration` is still a tune.Trainable property
# and has not been overridden by the user in the attempt to implement the
# algos logic (this should be done now inside `training_step`).
try:
assert isinstance(self.training_iteration, int)
except AssertionError:
raise AssertionError(
"Your Algorithm's `training_iteration` seems to be overridden by your "
"custom training logic! To solve this problem, simply rename your "
"`self.training_iteration()` method into `self.training_step`."
)
@OverrideToImplementCustomLogic
@classmethod
def get_default_config(cls) -> AlgorithmConfig:
return AlgorithmConfig()
@OverrideToImplementCustomLogic
def _remote_worker_ids_for_metrics(self) -> List[int]:
"""Returns a list of remote worker IDs to fetch metrics from.
Specific Algorithm implementations can override this method to
use a subset of the workers for metrics collection.
Returns:
List of remote worker IDs to fetch metrics from.
"""
return self.workers.healthy_worker_ids()
@OverrideToImplementCustomLogic_CallToSuperRecommended
@override(Trainable)
def setup(self, config: AlgorithmConfig) -> None:
# Setup our config: Merge the user-supplied config dict (which could
# be a partial config dict) with the class' default.
if not isinstance(config, AlgorithmConfig):
assert isinstance(config, PartialAlgorithmConfigDict)
config_obj = self.get_default_config()
if not isinstance(config_obj, AlgorithmConfig):
assert isinstance(config, PartialAlgorithmConfigDict)
config_obj = AlgorithmConfig().from_dict(config_obj)
config_obj.update_from_dict(config)
config_obj.env = self._env_id
self.config = config_obj
# Set Algorithm's seed after we have - if necessary - enabled
# tf eager-execution.
update_global_seed_if_necessary(self.config.framework_str, self.config.seed)
self._record_usage(self.config)
# Create the callbacks object.
self.callbacks = self.config.callbacks_class()
if self.config.log_level in ["WARN", "ERROR"]:
logger.info(
f"Current log_level is {self.config.log_level}. For more information, "
"set 'log_level': 'INFO' / 'DEBUG' or use the -v and "
"-vv flags."
)
if self.config.log_level:
logging.getLogger("ray.rllib").setLevel(self.config.log_level)
# Create local replay buffer if necessary.
self.local_replay_buffer = self._create_local_replay_buffer_if_necessary(
self.config
)
# Create a dict, mapping ActorHandles to sets of open remote
# requests (object refs). This way, we keep track, of which actors
# inside this Algorithm (e.g. a remote RolloutWorker) have
# already been sent how many (e.g. `sample()`) requests.
self.remote_requests_in_flight: DefaultDict[
ActorHandle, Set[ray.ObjectRef]
] = defaultdict(set)
self.workers: Optional[WorkerSet] = None
self.train_exec_impl = None
# Offline RL settings.
input_evaluation = self.config.get("input_evaluation")
if input_evaluation is not None and input_evaluation is not DEPRECATED_VALUE:
ope_dict = {str(ope): {"type": ope} for ope in input_evaluation}
deprecation_warning(
old="config.input_evaluation={}".format(input_evaluation),
new="config.evaluation(evaluation_config=config.overrides("
f"off_policy_estimation_methods={ope_dict}"
"))",
error=True,
help="Running OPE during training is not recommended.",
)
self.config.off_policy_estimation_methods = ope_dict
# Deprecated way of implementing Trainer sub-classes (or "templates"
# via the `build_trainer` utility function).
# Instead, sub-classes should override the Trainable's `setup()`
# method and call super().setup() from within that override at some
# point.
# Old design: Override `Trainer._init`.
_init = False
try:
self._init(self.config, self.env_creator)
_init = True
# New design: Override `Trainable.setup()` (as indented by tune.Trainable)
# and do or don't call `super().setup()` from within your override.
# By default, `super().setup()` will create both worker sets:
# "rollout workers" for collecting samples for training and - if
# applicable - "evaluation workers" for evaluation runs in between or
# parallel to training.
# TODO: Deprecate `_init()` and remove this try/except block.
except NotImplementedError:
pass
# Only if user did not override `_init()`:
if _init is False:
# - Create rollout workers here automatically.
# - Run the execution plan to create the local iterator to `next()`
# in each training iteration.
# This matches the behavior of using `build_trainer()`, which
# has been deprecated.
self.workers = WorkerSet(
env_creator=self.env_creator,
validate_env=self.validate_env,
default_policy_class=self.get_default_policy_class(self.config),
config=self.config,
num_workers=self.config.num_rollout_workers,
local_worker=True,
logdir=self.logdir,
)
# TODO (avnishn): Remove the execution plan API by q1 2023
# Function defining one single training iteration's behavior.
if self.config._disable_execution_plan_api:
# Ensure remote workers are initially in sync with the local worker.
self.workers.sync_weights()
# LocalIterator-creating "execution plan".
# Only call this once here to create `self.train_exec_impl`,
# which is a ray.util.iter.LocalIterator that will be `next`'d
# on each training iteration.
else:
self.train_exec_impl = self.execution_plan(
self.workers, self.config, **self._kwargs_for_execution_plan()
)
# Now that workers have been created, update our policies
# dict in config[multiagent] (with the correct original/
# unpreprocessed spaces).
self.config["multiagent"][
"policies"
] = self.workers.local_worker().policy_dict
# Compile, validate, and freeze an evaluation config.
self.evaluation_config = self.config.get_evaluation_config_object()
self.evaluation_config.validate()
self.evaluation_config.freeze()
# Evaluation WorkerSet setup.
# User would like to setup a separate evaluation worker set.
# Note: We skip workerset creation if we need to do offline evaluation
if self._should_create_evaluation_rollout_workers(self.evaluation_config):
_, env_creator = self._get_env_id_and_creator(
self.evaluation_config.env, self.evaluation_config
)
# Create a separate evaluation worker set for evaluation.
# If evaluation_num_workers=0, use the evaluation set's local
# worker for evaluation, otherwise, use its remote workers
# (parallelized evaluation).
self.evaluation_workers: WorkerSet = WorkerSet(
env_creator=env_creator,
validate_env=None,
default_policy_class=self.get_default_policy_class(self.config),
config=self.evaluation_config,
num_workers=self.config.evaluation_num_workers,
logdir=self.logdir,
)
if self.config.enable_async_evaluation:
self._evaluation_weights_seq_number = 0
self.evaluation_dataset = None
if (
self.evaluation_config.off_policy_estimation_methods
and not self.evaluation_config.ope_split_batch_by_episode
):
# the num worker is set to 0 to avoid creating shards. The dataset will not
# be repartioned to num_workers blocks.
logger.info("Creating evaluation dataset ...")
ds, _ = get_dataset_and_shards(self.evaluation_config, num_workers=0)
# Dataset should be in form of one episode per row. in case of bandits each
# row is just one time step. To make the computation more efficient later
# we remove the time dimension here.
parallelism = self.evaluation_config.evaluation_num_workers or 1
batch_size = max(ds.count() // parallelism, 1)
self.evaluation_dataset = ds.map_batches(
remove_time_dim,
batch_size=batch_size,
batch_format="pandas",
)
logger.info("Evaluation dataset created")
self.reward_estimators: Dict[str, OffPolicyEstimator] = {}
ope_types = {
"is": ImportanceSampling,
"wis": WeightedImportanceSampling,
"dm": DirectMethod,
"dr": DoublyRobust,
}
for name, method_config in self.config.off_policy_estimation_methods.items():
method_type = method_config.pop("type")
if method_type in ope_types:
deprecation_warning(
old=method_type,
new=str(ope_types[method_type]),
error=True,
)
method_type = ope_types[method_type]
elif isinstance(method_type, str):
logger.log(0, "Trying to import from string: " + method_type)
mod, obj = method_type.rsplit(".", 1)
mod = importlib.import_module(mod)
method_type = getattr(mod, obj)
if isinstance(method_type, type) and issubclass(
method_type, OfflineEvaluator
):
# TODO(kourosh) : Add an integration test for all these
# offline evaluators.
policy = self.get_policy()
if issubclass(method_type, OffPolicyEstimator):
method_config["gamma"] = self.config.gamma
self.reward_estimators[name] = method_type(policy, **method_config)
else:
raise ValueError(
f"Unknown off_policy_estimation type: {method_type}! Must be "
"either a class path or a sub-class of ray.rllib."
"offline.offline_evaluator::OfflineEvaluator"
)
# TODO (Rohan138): Refactor this and remove deprecated methods
# Need to add back method_type in case Algorithm is restored from checkpoint
method_config["type"] = method_type
self.learner_group = None
if self.config._enable_learner_api:
# TODO (Kourosh): This is an interim solution where policies and modules
# co-exist. In this world we have both policy_map and MARLModule that need
# to be consistent with one another. To make a consistent parity between
# the two we need to loop through the policy modules and create a simple
# MARLModule from the RLModule within each policy.
local_worker = self.workers.local_worker()
module_spec = local_worker.marl_module_spec
learner_group_config = self.config.get_learner_group_config(module_spec)
self.learner_group = learner_group_config.build()
# sync the weights from local rollout worker to trainers
weights = local_worker.get_weights()
self.learner_group.set_weights(weights)
# Run `on_algorithm_init` callback after initialization is done.
self.callbacks.on_algorithm_init(algorithm=self)
# TODO: Deprecated: In your sub-classes of Trainer, override `setup()`
# directly and call super().setup() from within it if you would like the
# default setup behavior plus some own setup logic.
# If you don't need the env/workers/config/etc.. setup for you by super,
# simply do not call super().setup() from your overridden method.
def _init(self, config: AlgorithmConfigDict, env_creator: EnvCreator) -> None:
raise NotImplementedError
[docs] @OverrideToImplementCustomLogic
@classmethod
def get_default_policy_class(
cls,
config: AlgorithmConfig,
) -> Optional[Type[Policy]]:
"""Returns a default Policy class to use, given a config.
This class will be used by an Algorithm in case
the policy class is not provided by the user in any single- or
multi-agent PolicySpec.
"""
return None
[docs] @override(Trainable)
def step(self) -> ResultDict:
"""Implements the main `Trainer.train()` logic.
Takes n attempts to perform a single training step. Thereby
catches RayErrors resulting from worker failures. After n attempts,
fails gracefully.
Override this method in your Trainer sub-classes if you would like to
handle worker failures yourself.
Otherwise, override only `training_step()` to implement the core
algorithm logic.
Returns:
The results dict with stats/infos on sampling, training,
and - if required - evaluation.
"""
# Do we have to run `self.evaluate()` this iteration?
# `self.iteration` gets incremented after this function returns,
# meaning that e. g. the first time this function is called,
# self.iteration will be 0.
evaluate_this_iter = (
self.config.evaluation_interval is not None
and (self.iteration + 1) % self.config.evaluation_interval == 0
)
# Results dict for training (and if appolicable: evaluation).
results: ResultDict = {}
# Parallel eval + training: Kick off evaluation-loop and parallel train() call.
if evaluate_this_iter and self.config.evaluation_parallel_to_training:
(
results,
train_iter_ctx,
) = self._run_one_training_iteration_and_evaluation_in_parallel()
# - No evaluation necessary, just run the next training iteration.
# - We have to evaluate in this training iteration, but no parallelism ->
# evaluate after the training iteration is entirely done.
else:
results, train_iter_ctx = self._run_one_training_iteration()
# Sequential: Train (already done above), then evaluate.
if evaluate_this_iter and not self.config.evaluation_parallel_to_training:
results.update(self._run_one_evaluation(train_future=None))
# Attach latest available evaluation results to train results,
# if necessary.
if not evaluate_this_iter and self.config.always_attach_evaluation_results:
assert isinstance(
self.evaluation_metrics, dict
), "Trainer.evaluate() needs to return a dict."
results.update(self.evaluation_metrics)
if hasattr(self, "workers") and isinstance(self.workers, WorkerSet):
# Sync filters on workers.
self._sync_filters_if_needed(
from_worker=self.workers.local_worker(),
workers=self.workers,
timeout_seconds=self.config.sync_filters_on_rollout_workers_timeout_s,
)
# TODO (avnishn): Remove the execution plan API by q1 2023
# Collect worker metrics and add combine them with `results`.
if self.config._disable_execution_plan_api:
episodes_this_iter = collect_episodes(
self.workers,
self._remote_worker_ids_for_metrics(),
timeout_seconds=self.config.metrics_episode_collection_timeout_s,
)
results = self._compile_iteration_results(
episodes_this_iter=episodes_this_iter,
step_ctx=train_iter_ctx,
iteration_results=results,
)
# Check `env_task_fn` for possible update of the env's task.
if self.config.env_task_fn is not None:
if not callable(self.config.env_task_fn):
raise ValueError(
"`env_task_fn` must be None or a callable taking "
"[train_results, env, env_ctx] as args!"
)
def fn(env, env_context, task_fn):
new_task = task_fn(results, env, env_context)
cur_task = env.get_task()
if cur_task != new_task:
env.set_task(new_task)
fn = functools.partial(fn, task_fn=self.config.env_task_fn)
self.workers.foreach_env_with_context(fn)
return results
[docs] @PublicAPI
def evaluate(
self,
duration_fn: Optional[Callable[[int], int]] = None,
) -> dict:
"""Evaluates current policy under `evaluation_config` settings.
Note that this default implementation does not do anything beyond
merging evaluation_config with the normal trainer config.
Args:
duration_fn: An optional callable taking the already run
num episodes as only arg and returning the number of
episodes left to run. It's used to find out whether
evaluation should continue.
"""
# Call the `_before_evaluate` hook.
self._before_evaluate()
if self.evaluation_dataset is not None:
return {"evaluation": self._run_offline_evaluation()}
# Sync weights to the evaluation WorkerSet.
if self.evaluation_workers is not None:
self.evaluation_workers.sync_weights(
from_worker_or_trainer=self.workers.local_worker()
)
self._sync_filters_if_needed(
from_worker=self.workers.local_worker(),
workers=self.evaluation_workers,
timeout_seconds=self.config.sync_filters_on_rollout_workers_timeout_s,
)
self.callbacks.on_evaluate_start(algorithm=self)
if self.config.custom_evaluation_function:
logger.info(
"Running custom eval function {}".format(
self.config.custom_evaluation_function
)
)
metrics = self.config.custom_evaluation_function(
self, self.evaluation_workers
)
if not metrics or not isinstance(metrics, dict):
raise ValueError(
"Custom eval function must return "
"dict of metrics, got {}.".format(metrics)
)
else:
if (
self.evaluation_workers is None
and self.workers.local_worker().input_reader is None
):
raise ValueError(
"Cannot evaluate w/o an evaluation worker set in "
"the Trainer or w/o an env on the local worker!\n"
"Try one of the following:\n1) Set "
"`evaluation_interval` >= 0 to force creating a "
"separate evaluation worker set.\n2) Set "
"`create_env_on_driver=True` to force the local "
"(non-eval) worker to have an environment to "
"evaluate on."
)
# How many episodes/timesteps do we need to run?
# In "auto" mode (only for parallel eval + training): Run as long
# as training lasts.
unit = self.config.evaluation_duration_unit
eval_cfg = self.evaluation_config
rollout = eval_cfg.rollout_fragment_length
num_envs = eval_cfg.num_envs_per_worker
auto = self.config.evaluation_duration == "auto"
duration = (
self.config.evaluation_duration
if not auto
else (self.config.evaluation_num_workers or 1)
* (1 if unit == "episodes" else rollout)
)
agent_steps_this_iter = 0
env_steps_this_iter = 0
# Default done-function returns True, whenever num episodes
# have been completed.
if duration_fn is None:
def duration_fn(num_units_done):
return duration - num_units_done
logger.info(f"Evaluating current state of {self} for {duration} {unit}.")
metrics = None
all_batches = []
# No evaluation worker set ->
# Do evaluation using the local worker. Expect error due to the
# local worker not having an env.
if self.evaluation_workers is None:
# If unit=episodes -> Run n times `sample()` (each sample
# produces exactly 1 episode).
# If unit=ts -> Run 1 `sample()` b/c the
# `rollout_fragment_length` is exactly the desired ts.
iters = duration if unit == "episodes" else 1
for _ in range(iters):
batch = self.workers.local_worker().sample()
agent_steps_this_iter += batch.agent_steps()
env_steps_this_iter += batch.env_steps()
if self.reward_estimators:
all_batches.append(batch)
metrics = collect_metrics(
self.workers,
keep_custom_metrics=eval_cfg.keep_per_episode_custom_metrics,
timeout_seconds=eval_cfg.metrics_episode_collection_timeout_s,
)
# Evaluation worker set only has local worker.
elif self.evaluation_workers.num_remote_workers() == 0:
# If unit=episodes -> Run n times `sample()` (each sample
# produces exactly 1 episode).
# If unit=ts -> Run 1 `sample()` b/c the
# `rollout_fragment_length` is exactly the desired ts.
iters = duration if unit == "episodes" else 1
for _ in range(iters):
batch = self.evaluation_workers.local_worker().sample()
agent_steps_this_iter += batch.agent_steps()
env_steps_this_iter += batch.env_steps()
if self.reward_estimators:
all_batches.append(batch)
# Evaluation worker set has n remote workers.
elif self.evaluation_workers.num_healthy_remote_workers() > 0:
# How many episodes have we run (across all eval workers)?
num_units_done = 0
_round = 0
# In case all of the remote evaluation workers die during a round
# of evaluation, we need to stop.
while True and self.evaluation_workers.num_healthy_remote_workers() > 0:
units_left_to_do = duration_fn(num_units_done)
if units_left_to_do <= 0:
break
_round += 1
unit_per_remote_worker = (
1 if unit == "episodes" else rollout * num_envs
)
# Select proper number of evaluation workers for this round.
selected_eval_worker_ids = [
worker_id
for i, worker_id in enumerate(
self.evaluation_workers.healthy_worker_ids()
)
if i * unit_per_remote_worker < units_left_to_do
]
batches = self.evaluation_workers.foreach_worker(
func=lambda w: w.sample(),
local_worker=False,
remote_worker_ids=selected_eval_worker_ids,
timeout_seconds=self.config.evaluation_sample_timeout_s,
)
if len(batches) != len(selected_eval_worker_ids):
logger.warning(
"Calling `sample()` on your remote evaluation worker(s) "
"resulted in a timeout (after the configured "
f"{self.config.evaluation_sample_timeout_s} seconds)! "
"Try to set `evaluation_sample_timeout_s` in your config"
" to a larger value."
+ (
" If your episodes don't terminate easily, you may "
"also want to set `evaluation_duration_unit` to "
"'timesteps' (instead of 'episodes')."
if unit == "episodes"
else ""
)
)
break
_agent_steps = sum(b.agent_steps() for b in batches)
_env_steps = sum(b.env_steps() for b in batches)
# 1 episode per returned batch.
if unit == "episodes":
num_units_done += len(batches)
# Make sure all batches are exactly one episode.
for ma_batch in batches:
ma_batch = ma_batch.as_multi_agent()
for batch in ma_batch.policy_batches.values():
assert batch.is_terminated_or_truncated()
# n timesteps per returned batch.
else:
num_units_done += (
_agent_steps
if self.config.count_steps_by == "agent_steps"
else _env_steps
)
if self.reward_estimators:
# TODO: (kourosh) This approach will cause an OOM issue when
# the dataset gets huge (should be ok for now).
all_batches.extend(batches)
agent_steps_this_iter += _agent_steps
env_steps_this_iter += _env_steps
logger.info(
f"Ran round {_round} of non-parallel evaluation "
f"({num_units_done}/{duration if not auto else '?'} "
f"{unit} done)"
)
else:
# Can't find a good way to run this evaluation.
# Wait for next iteration.
pass
if metrics is None:
metrics = collect_metrics(
self.evaluation_workers,
keep_custom_metrics=self.config.keep_per_episode_custom_metrics,
timeout_seconds=eval_cfg.metrics_episode_collection_timeout_s,
)
# TODO: Don't dump sampler results into top-level.
if not self.config.custom_evaluation_function:
metrics = dict({"sampler_results": metrics}, **metrics)
metrics[NUM_AGENT_STEPS_SAMPLED_THIS_ITER] = agent_steps_this_iter
metrics[NUM_ENV_STEPS_SAMPLED_THIS_ITER] = env_steps_this_iter
# TODO: Remove this key at some point. Here for backward compatibility.
metrics["timesteps_this_iter"] = env_steps_this_iter
# Compute off-policy estimates
estimates = defaultdict(list)
# for each batch run the estimator's fwd pass
for name, estimator in self.reward_estimators.items():
for batch in all_batches:
estimate_result = estimator.estimate(
batch,
split_batch_by_episode=self.config.ope_split_batch_by_episode,
)
estimates[name].append(estimate_result)
# collate estimates from all batches
if estimates:
metrics["off_policy_estimator"] = {}
for name, estimate_list in estimates.items():
avg_estimate = tree.map_structure(
lambda *x: np.mean(x, axis=0), *estimate_list
)
metrics["off_policy_estimator"][name] = avg_estimate
# Evaluation does not run for every step.
# Save evaluation metrics on trainer, so it can be attached to
# subsequent step results as latest evaluation result.
self.evaluation_metrics = {"evaluation": metrics}
# Trigger `on_evaluate_end` callback.
self.callbacks.on_evaluate_end(
algorithm=self, evaluation_metrics=self.evaluation_metrics
)
# Also return the results here for convenience.
return self.evaluation_metrics
@ExperimentalAPI
def _evaluate_async(
self,
duration_fn: Optional[Callable[[int], int]] = None,
) -> dict:
"""Evaluates current policy under `evaluation_config` settings.
Uses the AsyncParallelRequests manager to send frequent `sample.remote()`
requests to the evaluation RolloutWorkers and collect the results of these
calls. Handles worker failures (or slowdowns) gracefully due to the asynch'ness
and the fact that other eval RolloutWorkers can thus cover the workload.
Important Note: This will replace the current `self.evaluate()` method as the
default in the future.
Args:
duration_fn: An optional callable taking the already run
num episodes as only arg and returning the number of
episodes left to run. It's used to find out whether
evaluation should continue.
"""
# How many episodes/timesteps do we need to run?
# In "auto" mode (only for parallel eval + training): Run as long
# as training lasts.
unit = self.config.evaluation_duration_unit
eval_cfg = self.evaluation_config
rollout = eval_cfg.rollout_fragment_length
num_envs = eval_cfg.num_envs_per_worker
auto = self.config.evaluation_duration == "auto"
duration = (
self.config.evaluation_duration
if not auto
else (self.config.evaluation_num_workers or 1)
* (1 if unit == "episodes" else rollout)
)
# Call the `_before_evaluate` hook.
self._before_evaluate()
# TODO(Jun): Implement solution via connectors.
self._sync_filters_if_needed(
from_worker=self.workers.local_worker(),
workers=self.evaluation_workers,
timeout_seconds=eval_cfg.sync_filters_on_rollout_workers_timeout_s,
)
if self.config.custom_evaluation_function:
raise ValueError(
"`config.custom_evaluation_function` not supported in combination "
"with `enable_async_evaluation=True` config setting!"
)
if self.evaluation_workers is None and (
self.workers.local_worker().input_reader is None
or self.config.evaluation_num_workers == 0
):
raise ValueError(
"Evaluation w/o eval workers (calling Algorithm.evaluate() w/o "
"evaluation specifically set up) OR evaluation without input reader "
"OR evaluation with only a local evaluation worker "
"(`evaluation_num_workers=0`) not supported in combination "
"with `enable_async_evaluation=True` config setting!"
)
agent_steps_this_iter = 0
env_steps_this_iter = 0
logger.info(f"Evaluating current state of {self} for {duration} {unit}.")
all_batches = []
# Default done-function returns True, whenever num episodes
# have been completed.
if duration_fn is None:
def duration_fn(num_units_done):
return duration - num_units_done
# Put weights only once into object store and use same object
# ref to synch to all workers.
self._evaluation_weights_seq_number += 1
weights_ref = ray.put(self.workers.local_worker().get_weights())
weights_seq_no = self._evaluation_weights_seq_number
def remote_fn(worker):
# Pass in seq-no so that eval workers may ignore this call if no update has
# happened since the last call to `remote_fn` (sample).
worker.set_weights(
weights=ray.get(weights_ref), weights_seq_no=weights_seq_no
)
batch = worker.sample()
metrics = worker.get_metrics()
return batch, metrics, weights_seq_no
rollout_metrics = []
# How many episodes have we run (across all eval workers)?
num_units_done = 0
_round = 0
while self.evaluation_workers.num_healthy_remote_workers() > 0:
units_left_to_do = duration_fn(num_units_done)
if units_left_to_do <= 0:
break
_round += 1
# Get ready evaluation results and metrics asynchronously.
self.evaluation_workers.foreach_worker_async(
func=remote_fn,
healthy_only=True,
)
eval_results = self.evaluation_workers.fetch_ready_async_reqs()
batches = []
i = 0
for _, result in eval_results:
batch, metrics, seq_no = result
# Ignore results, if the weights seq-number does not match (is
# from a previous evaluation step) OR if we have already reached
# the configured duration (e.g. number of episodes to evaluate
# for).
if seq_no == self._evaluation_weights_seq_number and (
i * (1 if unit == "episodes" else rollout * num_envs)
< units_left_to_do
):
batches.append(batch)
rollout_metrics.extend(metrics)
i += 1
_agent_steps = sum(b.agent_steps() for b in batches)
_env_steps = sum(b.env_steps() for b in batches)
# 1 episode per returned batch.
if unit == "episodes":
num_units_done += len(batches)
# Make sure all batches are exactly one episode.
for ma_batch in batches:
ma_batch = ma_batch.as_multi_agent()
for batch in ma_batch.policy_batches.values():
assert batch.is_terminated_or_truncated()
# n timesteps per returned batch.
else:
num_units_done += (
_agent_steps
if self.config.count_steps_by == "agent_steps"
else _env_steps
)
if self.reward_estimators:
all_batches.extend(batches)
agent_steps_this_iter += _agent_steps
env_steps_this_iter += _env_steps
logger.info(
f"Ran round {_round} of parallel evaluation "
f"({num_units_done}/{duration if not auto else '?'} "
f"{unit} done)"
)
sampler_results = summarize_episodes(
rollout_metrics,
keep_custom_metrics=eval_cfg["keep_per_episode_custom_metrics"],
)
# TODO: Don't dump sampler results into top-level.
metrics = dict({"sampler_results": sampler_results}, **sampler_results)
metrics[NUM_AGENT_STEPS_SAMPLED_THIS_ITER] = agent_steps_this_iter
metrics[NUM_ENV_STEPS_SAMPLED_THIS_ITER] = env_steps_this_iter
# TODO: Remove this key at some point. Here for backward compatibility.
metrics["timesteps_this_iter"] = env_steps_this_iter
if self.reward_estimators:
# Compute off-policy estimates
metrics["off_policy_estimator"] = {}
total_batch = concat_samples(all_batches)
for name, estimator in self.reward_estimators.items():
estimates = estimator.estimate(total_batch)
metrics["off_policy_estimator"][name] = estimates
# Evaluation does not run for every step.
# Save evaluation metrics on trainer, so it can be attached to
# subsequent step results as latest evaluation result.
self.evaluation_metrics = {"evaluation": metrics}
# Trigger `on_evaluate_end` callback.
self.callbacks.on_evaluate_end(
algorithm=self, evaluation_metrics=self.evaluation_metrics
)
# Return evaluation results.
return self.evaluation_metrics
[docs] @OverrideToImplementCustomLogic
@DeveloperAPI
def restore_workers(self, workers: WorkerSet):
"""Try to restore failed workers if necessary.
Algorithms that use custom RolloutWorkers may override this method to
disable default, and create custom restoration logics.
Args:
workers: The WorkerSet to restore. This may be Rollout or Evaluation
workers.
"""
# If `workers` is None, or
# 1. `workers` (WorkerSet) does not have a local worker, and
# 2. `self.workers` (WorkerSet used for training) does not have a local worker
# -> we don't have a local worker to get state from, so we can't recover
# remote worker in this case.
if not workers or (
not workers.local_worker() and not self.workers.local_worker()
):
return
# This is really cheap, since probe_unhealthy_workers() is a no-op
# if there are no unhealthy workers.
restored = workers.probe_unhealthy_workers()
if restored:
from_worker = workers.local_worker() or self.workers.local_worker()
# Get the state of the correct (reference) worker. E.g. The local worker
# of the main WorkerSet.
state_ref = ray.put(from_worker.get_state())
# By default, entire local worker state is synced after restoration
# to bring these workers up to date.
workers.foreach_worker(
func=lambda w: w.set_state(ray.get(state_ref)),
remote_worker_ids=restored,
# Don't update the local_worker, b/c it's the one we are synching from.
local_worker=False,
timeout_seconds=self.config.worker_restore_timeout_s,
# Bring back actor after successful state syncing.
mark_healthy=True,
)
[docs] @OverrideToImplementCustomLogic
@DeveloperAPI
def training_step(self) -> ResultDict:
"""Default single iteration logic of an algorithm.
- Collect on-policy samples (SampleBatches) in parallel using the
Trainer's RolloutWorkers (@ray.remote).
- Concatenate collected SampleBatches into one train batch.
- Note that we may have more than one policy in the multi-agent case:
Call the different policies' `learn_on_batch` (simple optimizer) OR
`load_batch_into_buffer` + `learn_on_loaded_batch` (multi-GPU
optimizer) methods to calculate loss and update the model(s).
- Return all collected metrics for the iteration.
Returns:
The results dict from executing the training iteration.
"""
# Collect SampleBatches from sample workers until we have a full batch.
with self._timers[SAMPLE_TIMER]:
if self.config.count_steps_by == "agent_steps":
train_batch = synchronous_parallel_sample(
worker_set=self.workers,
max_agent_steps=self.config.train_batch_size,
)
else:
train_batch = synchronous_parallel_sample(
worker_set=self.workers, max_env_steps=self.config.train_batch_size
)
train_batch = train_batch.as_multi_agent()
self._counters[NUM_AGENT_STEPS_SAMPLED] += train_batch.agent_steps()
self._counters[NUM_ENV_STEPS_SAMPLED] += train_batch.env_steps()
# Only train if train_batch is not empty.
# In an extreme situation, all rollout workers die during the
# synchronous_parallel_sample() call above.
# In which case, we should skip training, wait a little bit, then probe again.
train_results = {}
if train_batch.agent_steps() > 0:
# Use simple optimizer (only for multi-agent or tf-eager; all other
# cases should use the multi-GPU optimizer, even if only using 1 GPU).
# TODO: (sven) rename MultiGPUOptimizer into something more
# meaningful.
if self.config._enable_learner_api:
is_module_trainable = self.workers.local_worker().is_policy_to_train
self.learner_group.set_is_module_trainable(is_module_trainable)
train_results = self.learner_group.update(train_batch)
elif self.config.get("simple_optimizer") is True:
train_results = train_one_step(self, train_batch)
else:
train_results = multi_gpu_train_one_step(self, train_batch)
else:
# Wait 1 sec before probing again via weight syncing.
time.sleep(1)
# Update weights and global_vars - after learning on the local worker - on all
# remote workers (only those policies that were actually trained).
global_vars = {
"timestep": self._counters[NUM_ENV_STEPS_SAMPLED],
}
with self._timers[SYNCH_WORKER_WEIGHTS_TIMER]:
# TODO (Avnish): Implement this on learner_group.get_weights().
# TODO (Kourosh): figure out how we are going to sync MARLModule
# weights to MARLModule weights under the policy_map objects?
from_worker_or_trainer = None
if self.config._enable_learner_api:
from_worker_or_trainer = self.learner_group
self.workers.sync_weights(
from_worker_or_trainer=from_worker_or_trainer,
policies=list(train_results.keys()),
global_vars=global_vars,
)
return train_results
@staticmethod
def execution_plan(workers, config, **kwargs):
raise NotImplementedError(
"It is not longer recommended to use Trainer's `execution_plan` method/API."
" Set `_disable_execution_plan_api=True` in your config and override the "
"`Trainer.training_step()` method with your algo's custom "
"execution logic."
)
[docs] @PublicAPI
def compute_single_action(
self,
observation: Optional[TensorStructType] = None,
state: Optional[List[TensorStructType]] = None,
*,
prev_action: Optional[TensorStructType] = None,
prev_reward: Optional[float] = None,
info: Optional[EnvInfoDict] = None,
input_dict: Optional[SampleBatch] = None,
policy_id: PolicyID = DEFAULT_POLICY_ID,
full_fetch: bool = False,
explore: Optional[bool] = None,
timestep: Optional[int] = None,
episode: Optional[Episode] = None,
unsquash_action: Optional[bool] = None,
clip_action: Optional[bool] = None,
# Deprecated args.
unsquash_actions=DEPRECATED_VALUE,
clip_actions=DEPRECATED_VALUE,
# Kwargs placeholder for future compatibility.
**kwargs,
) -> Union[
TensorStructType,
Tuple[TensorStructType, List[TensorType], Dict[str, TensorType]],
]:
"""Computes an action for the specified policy on the local worker.
Note that you can also access the policy object through
self.get_policy(policy_id) and call compute_single_action() on it
directly.
Args:
observation: Single (unbatched) observation from the
environment.
state: List of all RNN hidden (single, unbatched) state tensors.
prev_action: Single (unbatched) previous action value.
prev_reward: Single (unbatched) previous reward value.
info: Env info dict, if any.
input_dict: An optional SampleBatch that holds all the values
for: obs, state, prev_action, and prev_reward, plus maybe
custom defined views of the current env trajectory. Note
that only one of `obs` or `input_dict` must be non-None.
policy_id: Policy to query (only applies to multi-agent).
Default: "default_policy".
full_fetch: Whether to return extra action fetch results.
This is always set to True if `state` is specified.
explore: Whether to apply exploration to the action.
Default: None -> use self.config.explore.
timestep: The current (sampling) time step.
episode: This provides access to all of the internal episodes'
state, which may be useful for model-based or multi-agent
algorithms.
unsquash_action: Should actions be unsquashed according to the
env's/Policy's action space? If None, use the value of
self.config.normalize_actions.
clip_action: Should actions be clipped according to the
env's/Policy's action space? If None, use the value of
self.config.clip_actions.
Keyword Args:
kwargs: forward compatibility placeholder
Returns:
The computed action if full_fetch=False, or a tuple of a) the
full output of policy.compute_actions() if full_fetch=True
or we have an RNN-based Policy.
Raises:
KeyError: If the `policy_id` cannot be found in this Trainer's
local worker.
"""
if clip_actions != DEPRECATED_VALUE:
deprecation_warning(
old="Trainer.compute_single_action(`clip_actions`=...)",
new="Trainer.compute_single_action(`clip_action`=...)",
error=True,
)
clip_action = clip_actions
if unsquash_actions != DEPRECATED_VALUE:
deprecation_warning(
old="Trainer.compute_single_action(`unsquash_actions`=...)",
new="Trainer.compute_single_action(`unsquash_action`=...)",
error=True,
)
unsquash_action = unsquash_actions
# `unsquash_action` is None: Use value of config['normalize_actions'].
if unsquash_action is None:
unsquash_action = self.config.normalize_actions
# `clip_action` is None: Use value of config['clip_actions'].
elif clip_action is None:
clip_action = self.config.clip_actions
# User provided an input-dict: Assert that `obs`, `prev_a|r`, `state`
# are all None.
err_msg = (
"Provide either `input_dict` OR [`observation`, ...] as "
"args to Trainer.compute_single_action!"
)
if input_dict is not None:
assert (
observation is None
and prev_action is None
and prev_reward is None
and state is None
), err_msg
observation = input_dict[SampleBatch.OBS]
else:
assert observation is not None, err_msg
# Get the policy to compute the action for (in the multi-agent case,
# Trainer may hold >1 policies).
policy = self.get_policy(policy_id)
if policy is None:
raise KeyError(
f"PolicyID '{policy_id}' not found in PolicyMap of the "
f"Trainer's local worker!"
)
local_worker = self.workers.local_worker()
if not self.config.get("enable_connectors"):
# Check the preprocessor and preprocess, if necessary.
pp = local_worker.preprocessors[policy_id]
if pp and type(pp).__name__ != "NoPreprocessor":
observation = pp.transform(observation)
observation = local_worker.filters[policy_id](observation, update=False)
else:
# Just preprocess observations, similar to how it used to be done before.
pp = policy.agent_connectors[ObsPreprocessorConnector]
# convert the observation to array if possible
if not isinstance(observation, (np.ndarray, dict, tuple)):
try:
observation = np.asarray(observation)
except Exception:
raise ValueError(
f"Observation type {type(observation)} cannot be converted to "
f"np.ndarray."
)
if pp:
assert len(pp) == 1, "Only one preprocessor should be in the pipeline"
pp = pp[0]
if not pp.is_identity():
# Note(Kourosh): This call will leave the policy's connector
# in eval mode. would that be a problem?
pp.in_eval()
if observation is not None:
_input_dict = {SampleBatch.OBS: observation}
elif input_dict is not None:
_input_dict = {SampleBatch.OBS: input_dict[SampleBatch.OBS]}
else:
raise ValueError(
"Either observation or input_dict must be provided."
)
# TODO (Kourosh): Create a new util method for algorithm that
# computes actions based on raw inputs from env and can keep track
# of its own internal state.
acd = AgentConnectorDataType("0", "0", _input_dict)
# make sure the state is reset since we are only applying the
# preprocessor
pp.reset(env_id="0")
ac_o = pp([acd])[0]
observation = ac_o.data[SampleBatch.OBS]
# Input-dict.
if input_dict is not None:
input_dict[SampleBatch.OBS] = observation
action, state, extra = policy.compute_single_action(
input_dict=input_dict,
explore=explore,
timestep=timestep,
episode=episode,
)
# Individual args.
else:
action, state, extra = policy.compute_single_action(
obs=observation,
state=state,
prev_action=prev_action,
prev_reward=prev_reward,
info=info,
explore=explore,
timestep=timestep,
episode=episode,
)
# If we work in normalized action space (normalize_actions=True),
# we re-translate here into the env's action space.
if unsquash_action:
action = space_utils.unsquash_action(action, policy.action_space_struct)
# Clip, according to env's action space.
elif clip_action:
action = space_utils.clip_action(action, policy.action_space_struct)
# Return 3-Tuple: Action, states, and extra-action fetches.
if state or full_fetch:
return action, state, extra
# Ensure backward compatibility.
else:
return action
[docs] @PublicAPI
def compute_actions(
self,
observations: TensorStructType,
state: Optional[List[TensorStructType]] = None,
*,
prev_action: Optional[TensorStructType] = None,
prev_reward: Optional[TensorStructType] = None,
info: Optional[EnvInfoDict] = None,
policy_id: PolicyID = DEFAULT_POLICY_ID,
full_fetch: bool = False,
explore: Optional[bool] = None,
timestep: Optional[int] = None,
episodes: Optional[List[Episode]] = None,
unsquash_actions: Optional[bool] = None,
clip_actions: Optional[bool] = None,
# Deprecated.
normalize_actions=None,
**kwargs,
):
"""Computes an action for the specified policy on the local Worker.
Note that you can also access the policy object through
self.get_policy(policy_id) and call compute_actions() on it directly.
Args:
observation: Observation from the environment.
state: RNN hidden state, if any. If state is not None,
then all of compute_single_action(...) is returned
(computed action, rnn state(s), logits dictionary).
Otherwise compute_single_action(...)[0] is returned
(computed action).
prev_action: Previous action value, if any.
prev_reward: Previous reward, if any.
info: Env info dict, if any.
policy_id: Policy to query (only applies to multi-agent).
full_fetch: Whether to return extra action fetch results.
This is always set to True if RNN state is specified.
explore: Whether to pick an exploitation or exploration
action (default: None -> use self.config.explore).
timestep: The current (sampling) time step.
episodes: This provides access to all of the internal episodes'
state, which may be useful for model-based or multi-agent
algorithms.
unsquash_actions: Should actions be unsquashed according
to the env's/Policy's action space? If None, use
self.config.normalize_actions.
clip_actions: Should actions be clipped according to the
env's/Policy's action space? If None, use
self.config.clip_actions.
Keyword Args:
kwargs: forward compatibility placeholder
Returns:
The computed action if full_fetch=False, or a tuple consisting of
the full output of policy.compute_actions_from_input_dict() if
full_fetch=True or we have an RNN-based Policy.
"""
if normalize_actions is not None:
deprecation_warning(
old="Trainer.compute_actions(`normalize_actions`=...)",
new="Trainer.compute_actions(`unsquash_actions`=...)",
error=True,
)
unsquash_actions = normalize_actions
# `unsquash_actions` is None: Use value of config['normalize_actions'].
if unsquash_actions is None:
unsquash_actions = self.config.normalize_actions
# `clip_actions` is None: Use value of config['clip_actions'].
elif clip_actions is None:
clip_actions = self.config.clip_actions
# Preprocess obs and states.
state_defined = state is not None
policy = self.get_policy(policy_id)
filtered_obs, filtered_state = [], []
for agent_id, ob in observations.items():
worker = self.workers.local_worker()
preprocessed = worker.preprocessors[policy_id].transform(ob)
filtered = worker.filters[policy_id](preprocessed, update=False)
filtered_obs.append(filtered)
if state is None:
continue
elif agent_id in state:
filtered_state.append(state[agent_id])
else:
filtered_state.append(policy.get_initial_state())
# Batch obs and states
obs_batch = np.stack(filtered_obs)
if state is None:
state = []
else:
state = list(zip(*filtered_state))
state = [np.stack(s) for s in state]
input_dict = {SampleBatch.OBS: obs_batch}
# prev_action and prev_reward can be None, np.ndarray, or tensor-like structure.
# Explicitly check for None here to avoid the error message "The truth value of
# an array with more than one element is ambiguous.", when np arrays are passed
# as arguments.
if prev_action is not None:
input_dict[SampleBatch.PREV_ACTIONS] = prev_action
if prev_reward is not None:
input_dict[SampleBatch.PREV_REWARDS] = prev_reward
if info:
input_dict[SampleBatch.INFOS] = info
for i, s in enumerate(state):
input_dict[f"state_in_{i}"] = s
# Batch compute actions
actions, states, infos = policy.compute_actions_from_input_dict(
input_dict=input_dict,
explore=explore,
timestep=timestep,
episodes=episodes,
)
# Unbatch actions for the environment into a multi-agent dict.
single_actions = space_utils.unbatch(actions)
actions = {}
for key, a in zip(observations, single_actions):
# If we work in normalized action space (normalize_actions=True),
# we re-translate here into the env's action space.
if unsquash_actions:
a = space_utils.unsquash_action(a, policy.action_space_struct)
# Clip, according to env's action space.
elif clip_actions:
a = space_utils.clip_action(a, policy.action_space_struct)
actions[key] = a
# Unbatch states into a multi-agent dict.
unbatched_states = {}
for idx, agent_id in enumerate(observations):
unbatched_states[agent_id] = [s[idx] for s in states]
# Return only actions or full tuple
if state_defined or full_fetch:
return actions, unbatched_states, infos
else:
return actions
[docs] @PublicAPI
def get_policy(self, policy_id: PolicyID = DEFAULT_POLICY_ID) -> Policy:
"""Return policy for the specified id, or None.
Args:
policy_id: ID of the policy to return.
"""
return self.workers.local_worker().get_policy(policy_id)
[docs] @PublicAPI
def get_weights(self, policies: Optional[List[PolicyID]] = None) -> dict:
"""Return a dictionary of policy ids to weights.
Args:
policies: Optional list of policies to return weights for,
or None for all policies.
"""
return self.workers.local_worker().get_weights(policies)
[docs] @PublicAPI
def set_weights(self, weights: Dict[PolicyID, dict]):
"""Set policy weights by policy id.
Args:
weights: Map of policy ids to weights to set.
"""
self.workers.local_worker().set_weights(weights)
[docs] @PublicAPI
def add_policy(
self,
policy_id: PolicyID,
policy_cls: Optional[Type[Policy]] = None,
policy: Optional[Policy] = None,
*,
observation_space: Optional[gym.spaces.Space] = None,
action_space: Optional[gym.spaces.Space] = None,
config: Optional[Union[AlgorithmConfig, PartialAlgorithmConfigDict]] = None,
policy_state: Optional[PolicyState] = None,
policy_mapping_fn: Optional[Callable[[AgentID, EpisodeID], PolicyID]] = None,
policies_to_train: Optional[
Union[
Container[PolicyID],
Callable[[PolicyID, Optional[SampleBatchType]], bool],
]
] = None,
evaluation_workers: bool = True,
module_spec: Optional[SingleAgentRLModuleSpec] = None,
# Deprecated.
workers: Optional[List[Union[RolloutWorker, ActorHandle]]] = DEPRECATED_VALUE,
) -> Optional[Policy]:
"""Adds a new policy to this Algorithm.
Args:
policy_id: ID of the policy to add.
IMPORTANT: Must not contain characters that
are also not allowed in Unix/Win filesystems, such as: `<>:"/|?*`,
or a dot, space or backslash at the end of the ID.
policy_cls: The Policy class to use for constructing the new Policy.
Note: Only one of `policy_cls` or `policy` must be provided.
policy: The Policy instance to add to this algorithm. If not None, the
given Policy object will be directly inserted into the Algorithm's
local worker and clones of that Policy will be created on all remote
workers as well as all evaluation workers.
Note: Only one of `policy_cls` or `policy` must be provided.
observation_space: The observation space of the policy to add.
If None, try to infer this space from the environment.
action_space: The action space of the policy to add.
If None, try to infer this space from the environment.
config: The config object or 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
or a callable taking PolicyID and SampleBatchType and
returning a bool (trainable or not?).
If None, will keep the existing setup in place. Policies,
whose IDs are not in the list (or for which the callable
returns False) will not be updated.
evaluation_workers: Whether to add the new policy also
to the evaluation WorkerSet.
module_spec: In the new RLModule API we need to pass in the module_spec for
the new module that is supposed to be added. Knowing the policy spec is
not sufficient.
workers: A list of RolloutWorker/ActorHandles (remote
RolloutWorkers) to add this policy to. If defined, will only
add the given policy to these workers.
Returns:
The newly added policy (the copy that got added to the local
worker). If `workers` was provided, None is returned.
"""
validate_policy_id(policy_id, error=True)
if workers is not DEPRECATED_VALUE:
deprecation_warning(
old="Algorithm.add_policy(.., workers=..)",
help=(
"The `workers` argument to `Algorithm.add_policy()` is deprecated! "
"Please do not use it anymore."
),
error=True,
)
self.workers.add_policy(
policy_id,
policy_cls,
policy,
observation_space=observation_space,
action_space=action_space,
config=config,
policy_state=policy_state,
policy_mapping_fn=policy_mapping_fn,
policies_to_train=policies_to_train,
module_spec=module_spec,
)
# If learner API is enabled, we need to also add the underlying module
# to the learner group.
if self.config._enable_learner_api:
policy = self.get_policy(policy_id)
module = policy.model
self.learner_group.add_module(
module_id=policy_id,
module_spec=SingleAgentRLModuleSpec.from_module(module),
)
weights = policy.get_weights()
self.learner_group.set_weights({policy_id: weights})
# Add to evaluation workers, if necessary.
if evaluation_workers is True and self.evaluation_workers is not None:
self.evaluation_workers.add_policy(
policy_id,
policy_cls,
policy,
observation_space=observation_space,
action_space=action_space,
config=config,
policy_state=policy_state,
policy_mapping_fn=policy_mapping_fn,
policies_to_train=policies_to_train,
module_spec=module_spec,
)
# Return newly added policy (from the local rollout worker).
return self.get_policy(policy_id)
[docs] @PublicAPI
def remove_policy(
self,
policy_id: PolicyID = DEFAULT_POLICY_ID,
*,
policy_mapping_fn: Optional[Callable[[AgentID], PolicyID]] = None,
policies_to_train: Optional[
Union[
Container[PolicyID],
Callable[[PolicyID, Optional[SampleBatchType]], bool],
]
] = None,
evaluation_workers: bool = True,
) -> None:
"""Removes a new policy from this Algorithm.
Args:
policy_id: ID of the policy to be removed.
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
or a callable taking PolicyID and SampleBatchType and
returning a bool (trainable or not?).
If None, will keep the existing setup in place. Policies,
whose IDs are not in the list (or for which the callable
returns False) will not be updated.
evaluation_workers: Whether to also remove the policy from the
evaluation WorkerSet.
"""
def fn(worker):
worker.remove_policy(
policy_id=policy_id,
policy_mapping_fn=policy_mapping_fn,
policies_to_train=policies_to_train,
)
self.workers.foreach_worker(fn, local_worker=True, healthy_only=True)
if evaluation_workers and self.evaluation_workers is not None:
self.evaluation_workers.foreach_worker(
fn,
local_worker=True,
healthy_only=True,
)
[docs] @DeveloperAPI
def export_policy_model(
self,
export_dir: str,
policy_id: PolicyID = DEFAULT_POLICY_ID,
onnx: Optional[int] = None,
) -> None:
"""Exports policy model with given policy_id to a local directory.
Args:
export_dir: Writable local directory.
policy_id: Optional policy id to export.
onnx: If given, will export model in ONNX format. The
value of this parameter set the ONNX OpSet version to use.
If None, the output format will be DL framework specific.
Example:
>>> from ray.rllib.algorithms.ppo import PPO
>>> # Use an Algorithm from RLlib or define your own.
>>> algo = PPO(...) # doctest: +SKIP
>>> for _ in range(10): # doctest: +SKIP
>>> algo.train() # doctest: +SKIP
>>> algo.export_policy_model("/tmp/dir") # doctest: +SKIP
>>> algo.export_policy_model("/tmp/dir/onnx", onnx=1) # doctest: +SKIP
"""
self.get_policy(policy_id).export_model(export_dir, onnx)
[docs] @DeveloperAPI
def export_policy_checkpoint(
self,
export_dir: str,
filename_prefix=DEPRECATED_VALUE, # deprecated arg, do not use anymore
policy_id: PolicyID = DEFAULT_POLICY_ID,
) -> None:
"""Exports Policy checkpoint to a local directory and returns an AIR Checkpoint.
Args:
export_dir: Writable local directory to store the AIR Checkpoint
information into.
policy_id: Optional policy ID to export. If not provided, will export
"default_policy". If `policy_id` does not exist in this Algorithm,
will raise a KeyError.
Raises:
KeyError if `policy_id` cannot be found in this Algorithm.
Example:
>>> from ray.rllib.algorithms.ppo import PPO
>>> # Use an Algorithm from RLlib or define your own.
>>> algo = PPO(...) # doctest: +SKIP
>>> for _ in range(10): # doctest: +SKIP
>>> algo.train() # doctest: +SKIP
>>> algo.export_policy_checkpoint("/tmp/export_dir") # doctest: +SKIP
"""
# `filename_prefix` should not longer be used as new Policy checkpoints
# contain more than one file with a fixed filename structure.
if filename_prefix != DEPRECATED_VALUE:
deprecation_warning(
old="Algorithm.export_policy_checkpoint(filename_prefix=...)",
error=True,
)
policy = self.get_policy(policy_id)
if policy is None:
raise KeyError(f"Policy with ID {policy_id} not found in Algorithm!")
policy.export_checkpoint(export_dir)
[docs] @DeveloperAPI
def import_policy_model_from_h5(
self,
import_file: str,
policy_id: PolicyID = DEFAULT_POLICY_ID,
) -> None:
"""Imports a policy's model with given policy_id from a local h5 file.
Args:
import_file: The h5 file to import from.
policy_id: Optional policy id to import into.
Example:
>>> from ray.rllib.algorithms.ppo import PPO
>>> algo = PPO(...) # doctest: +SKIP
>>> algo.import_policy_model_from_h5("/tmp/weights.h5") # doctest: +SKIP
>>> for _ in range(10): # doctest: +SKIP
>>> algo.train() # doctest: +SKIP
"""
self.get_policy(policy_id).import_model_from_h5(import_file)
# Sync new weights to remote workers.
self._sync_weights_to_workers(worker_set=self.workers)
[docs] @override(Trainable)
def save_checkpoint(self, checkpoint_dir: str) -> str:
"""Exports AIR Checkpoint to a local directory and returns its directory path.
The structure of an Algorithm checkpoint dir will be as follows::
policies/
pol_1/
policy_state.pkl
pol_2/
policy_state.pkl
learner/
learner_state.json
module_state/
module_1/
...
optimizer_state/
optimizers_module_1/
...
rllib_checkpoint.json
algorithm_state.pkl
Note: `rllib_checkpoint.json` contains a "version" key (e.g. with value 0.1)
helping RLlib to remain backward compatible wrt. restoring from checkpoints from
Ray 2.0 onwards.
Args:
checkpoint_dir: The directory where the checkpoint files will be stored.
Returns:
The path to the created AIR Checkpoint directory.
"""
state = self.__getstate__()
# Extract policy states from worker state (Policies get their own
# checkpoint sub-dirs).
policy_states = {}
if "worker" in state and "policy_states" in state["worker"]:
policy_states = state["worker"].pop("policy_states", {})
# Add RLlib checkpoint version.
if self.config._enable_learner_api:
state["checkpoint_version"] = CHECKPOINT_VERSION_LEARNER
else:
state["checkpoint_version"] = CHECKPOINT_VERSION
# Write state (w/o policies) to disk.
state_file = os.path.join(checkpoint_dir, "algorithm_state.pkl")
with open(state_file, "wb") as f:
pickle.dump(state, f)
# Write rllib_checkpoint.json.
with open(os.path.join(checkpoint_dir, "rllib_checkpoint.json"), "w") as f:
json.dump(
{
"type": "Algorithm",
"checkpoint_version": str(state["checkpoint_version"]),
"format": "cloudpickle",
"state_file": state_file,
"policy_ids": list(policy_states.keys()),
"ray_version": ray.__version__,
"ray_commit": ray.__commit__,
},
f,
)
# Write individual policies to disk, each in their own sub-directory.
for pid, policy_state in policy_states.items():
# From here on, disallow policyIDs that would not work as directory names.
validate_policy_id(pid, error=True)
policy_dir = os.path.join(checkpoint_dir, "policies", pid)
os.makedirs(policy_dir, exist_ok=True)
policy = self.get_policy(pid)
policy.export_checkpoint(policy_dir, policy_state=policy_state)
# if we are using the learner API, save the learner group state
if self.config._enable_learner_api:
learner_state_dir = os.path.join(checkpoint_dir, "learner")
self.learner_group.save_state(learner_state_dir)
return checkpoint_dir
@override(Trainable)
def load_checkpoint(self, checkpoint: str) -> None:
# Checkpoint is provided as a directory name.
# Restore from the checkpoint file or dir.
checkpoint_info = get_checkpoint_info(checkpoint)
checkpoint_data = Algorithm._checkpoint_info_to_algorithm_state(checkpoint_info)
self.__setstate__(checkpoint_data)
if self.config._enable_learner_api:
learner_state_dir = os.path.join(checkpoint, "learner")
self.learner_group.load_state(learner_state_dir)
@override(Trainable)
def log_result(self, result: ResultDict) -> None:
# Log after the callback is invoked, so that the user has a chance
# to mutate the result.
# TODO: Remove `trainer` arg at some point to fully deprecate the old signature.
self.callbacks.on_train_result(algorithm=self, result=result)
# Then log according to Trainable's logging logic.
Trainable.log_result(self, result)
@override(Trainable)
def cleanup(self) -> None:
# Stop all workers.
if hasattr(self, "workers") and self.workers is not None:
self.workers.stop()
if hasattr(self, "evaluation_workers") and self.evaluation_workers is not None:
self.evaluation_workers.stop()
@OverrideToImplementCustomLogic
@classmethod
@override(Trainable)
def default_resource_request(
cls, config: Union[AlgorithmConfig, PartialAlgorithmConfigDict]
) -> Union[Resources, PlacementGroupFactory]:
# Default logic for RLlib Algorithms:
# Create one bundle per individual worker (local or remote).
# Use `num_cpus_for_local_worker` and `num_gpus` for the local worker and
# `num_cpus_per_worker` and `num_gpus_per_worker` for the remote
# workers to determine their CPU/GPU resource needs.
# Convenience config handles.
cf = cls.get_default_config().update_from_dict(config)
cf.validate()
cf.freeze()
# get evaluation config
eval_cf = cf.get_evaluation_config_object()
eval_cf.validate()
eval_cf.freeze()
# resources for the driver of this trainable
if cf._enable_learner_api:
if cf.num_learner_workers == 0:
# in this case local_worker only does sampling and training is done on
# local learner worker
driver = {
# sampling and training is not done concurrently when local is
# used, so pick the max.
"CPU": max(
cf.num_cpus_per_learner_worker, cf.num_cpus_for_local_worker
),
"GPU": cf.num_gpus_per_learner_worker,
}
else:
# in this case local_worker only does sampling and training is done on
# remote learner workers
driver = {"CPU": cf.num_cpus_for_local_worker, "GPU": 0}
else:
# Without Learner API, the local_worker can do both sampling and training.
# So, we need to allocate the same resources for the driver as for the
# local_worker.
driver = {
"CPU": cf.num_cpus_for_local_worker,
"GPU": 0 if cf._fake_gpus else cf.num_gpus,
}
# resources for remote rollout env samplers
rollout_bundles = [
{
"CPU": cf.num_cpus_per_worker,
"GPU": cf.num_gpus_per_worker,
**cf.custom_resources_per_worker,
}
for _ in range(cf.num_rollout_workers)
]
# resources for remote evaluation env samplers or datasets (if any)
if cls._should_create_evaluation_rollout_workers(eval_cf):
# Evaluation workers.
# Note: The local eval worker is located on the driver CPU.
evaluation_bundles = [
{
"CPU": eval_cf.num_cpus_per_worker,
"GPU": eval_cf.num_gpus_per_worker,
**eval_cf.custom_resources_per_worker,
}
for _ in range(eval_cf.evaluation_num_workers)
]
else:
# resources for offline dataset readers during evaluation
# Note (Kourosh): we should not claim extra workers for
# training on the offline dataset, since rollout workers have already
# claimed it.
# Another Note (Kourosh): dataset reader will not use placement groups so
# whatever we specify here won't matter because dataset won't even use it.
# Disclaimer: using ray dataset in tune may cause deadlock when multiple
# tune trials get scheduled on the same node and do not leave any spare
# resources for dataset operations. The workaround is to limit the
# max_concurrent trials so that some spare cpus are left for dataset
# operations. This behavior should get fixed by the dataset team. more info
# found here:
# https://docs.ray.io/en/master/data/dataset-internals.html#datasets-tune
evaluation_bundles = []
# resources for remote learner workers
learner_bundles = []
if cf._enable_learner_api and cf.num_learner_workers > 0:
# can't specify cpus for learner workers at the same
# time as gpus
if cf.num_gpus_per_learner_worker:
learner_bundles = [
{
"GPU": cf.num_gpus_per_learner_worker,
}
for _ in range(cf.num_learner_workers)
]
elif cf.num_cpus_per_learner_worker:
learner_bundles = [
{
"CPU": cf.num_cpus_per_learner_worker,
}
for _ in range(cf.num_learner_workers)
]
bundles = [driver] + rollout_bundles + evaluation_bundles + learner_bundles
# Return PlacementGroupFactory containing all needed resources
# (already properly defined as device bundles).
return PlacementGroupFactory(
bundles=bundles,
strategy=config.get("placement_strategy", "PACK"),
)
@DeveloperAPI
def _before_evaluate(self):
"""Pre-evaluation callback."""
pass
@staticmethod
def _get_env_id_and_creator(
env_specifier: Union[str, EnvType, None], config: AlgorithmConfig
) -> Tuple[Optional[str], EnvCreator]:
"""Returns env_id and creator callable given original env id from config.
Args:
env_specifier: An env class, an already tune registered env ID, a known
gym env name, or None (if no env is used).
config: The AlgorithmConfig object.
Returns:
Tuple consisting of a) env ID string and b) env creator callable.
"""
# Environment is specified via a string.
if isinstance(env_specifier, str):
# An already registered env.
if _global_registry.contains(ENV_CREATOR, env_specifier):
return env_specifier, _global_registry.get(ENV_CREATOR, env_specifier)
# A class path specifier.
elif "." in env_specifier:
def env_creator_from_classpath(env_context):
try:
env_obj = from_config(env_specifier, env_context)
except ValueError:
raise EnvError(
ERR_MSG_INVALID_ENV_DESCRIPTOR.format(env_specifier)
)
return env_obj
return env_specifier, env_creator_from_classpath
# Try gym/PyBullet/Vizdoom.
else:
return env_specifier, functools.partial(
_gym_env_creator, env_descriptor=env_specifier
)
elif isinstance(env_specifier, type):
env_id = env_specifier # .__name__
if config["remote_worker_envs"]:
# Check gym version (0.22 or higher?).
# If > 0.21, can't perform auto-wrapping of the given class as this
# would lead to a pickle error.
gym_version = pkg_resources.get_distribution("gym").version
if version.parse(gym_version) >= version.parse("0.22"):
raise ValueError(
"Cannot specify a gym.Env class via `config.env` while setting "
"`config.remote_worker_env=True` AND your gym version is >= "
"0.22! Try installing an older version of gym or set `config."
"remote_worker_env=False`."
)
@ray.remote(num_cpus=1)
class _wrapper(env_specifier):
# Add convenience `_get_spaces` and `_is_multi_agent`
# methods:
def _get_spaces(self):
return self.observation_space, self.action_space
def _is_multi_agent(self):
from ray.rllib.env.multi_agent_env import MultiAgentEnv
return isinstance(self, MultiAgentEnv)
return env_id, lambda cfg: _wrapper.remote(cfg)
# gym.Env-subclass: Also go through our RLlib gym-creator.
elif issubclass(env_specifier, gym.Env):
return env_id, functools.partial(
_gym_env_creator,
env_descriptor=env_specifier,
auto_wrap_old_gym_envs=config.get("auto_wrap_old_gym_envs", True),
)
# All other env classes: Call c'tor directly.
else:
return env_id, lambda cfg: env_specifier(cfg)
# No env -> Env creator always returns None.
elif env_specifier is None:
return None, lambda env_config: None
else:
raise ValueError(
"{} is an invalid env specifier. ".format(env_specifier)
+ "You can specify a custom env as either a class "
'(e.g., YourEnvCls) or a registered env id (e.g., "your_env").'
)
def _sync_filters_if_needed(
self,
from_worker: RolloutWorker,
workers: WorkerSet,
timeout_seconds: Optional[float] = None,
):
if (
from_worker
and self.config.get("observation_filter", "NoFilter") != "NoFilter"
):
FilterManager.synchronize(
from_worker.filters,
workers,
update_remote=self.config.synchronize_filters,
timeout_seconds=timeout_seconds,
)
logger.debug("synchronized filters: {}".format(from_worker.filters))
@DeveloperAPI
def _sync_weights_to_workers(
self,
*,
worker_set: WorkerSet,
) -> None:
"""Sync "main" weights to given WorkerSet or list of workers."""
# Broadcast the new policy weights to all remote workers in worker_set.
logger.info("Synchronizing weights to workers.")
worker_set.sync_weights()
@classmethod
@override(Trainable)
def resource_help(cls, config: Union[AlgorithmConfig, AlgorithmConfigDict]) -> str:
return (
"\n\nYou can adjust the resource requests of RLlib Algorithms by calling "
"`AlgorithmConfig.resources("
"num_gpus=.., num_cpus_per_worker=.., num_gpus_per_worker=.., ..)` or "
"`AgorithmConfig.rollouts(num_rollout_workers=..)`. See "
"the `ray.rllib.algorithms.algorithm_config.AlgorithmConfig` classes "
"(each Algorithm has its own subclass of this class) for more info.\n\n"
f"The config of this Algorithm is: {config}"
)
@override(Trainable)
def get_auto_filled_metrics(
self,
now: Optional[datetime] = None,
time_this_iter: Optional[float] = None,
timestamp: Optional[int] = None,
debug_metrics_only: bool = False,
) -> dict:
# Override this method to make sure, the `config` key of the returned results
# contains the proper Tune config dict (instead of an AlgorithmConfig object).
auto_filled = super().get_auto_filled_metrics(
now, time_this_iter, timestamp, debug_metrics_only
)
if "config" not in auto_filled:
raise KeyError("`config` key not found in auto-filled results dict!")
# If `config` key is no dict (but AlgorithmConfig object) ->
# make sure, it's a dict to not break Tune APIs.
if not isinstance(auto_filled["config"], dict):
assert isinstance(auto_filled["config"], AlgorithmConfig)
auto_filled["config"] = auto_filled["config"].to_dict()
return auto_filled
[docs] @classmethod
def merge_trainer_configs(
cls,
config1: AlgorithmConfigDict,
config2: PartialAlgorithmConfigDict,
_allow_unknown_configs: Optional[bool] = None,
) -> AlgorithmConfigDict:
"""Merges a complete Algorithm config dict with a partial override dict.
Respects nested structures within the config dicts. The values in the
partial override dict take priority.
Args:
config1: The complete Algorithm's dict to be merged (overridden)
with `config2`.
config2: The partial override config dict to merge on top of
`config1`.
_allow_unknown_configs: If True, keys in `config2` that don't exist
in `config1` are allowed and will be added to the final config.
Returns:
The merged full algorithm config dict.
"""
config1 = copy.deepcopy(config1)
if "callbacks" in config2 and type(config2["callbacks"]) is dict:
deprecation_warning(
"callbacks dict interface",
"a class extending rllib.algorithms.callbacks.DefaultCallbacks; "
"see `rllib/examples/custom_metrics_and_callbacks.py` for an example.",
error=True,
)
if _allow_unknown_configs is None:
_allow_unknown_configs = cls._allow_unknown_configs
return deep_update(
config1,
config2,
_allow_unknown_configs,
cls._allow_unknown_subkeys,
cls._override_all_subkeys_if_type_changes,
cls._override_all_key_list,
)
[docs] @staticmethod
@ExperimentalAPI
def validate_env(env: EnvType, env_context: EnvContext) -> None:
"""Env validator function for this Algorithm class.
Override this in child classes to define custom validation
behavior.
Args:
env: The (sub-)environment to validate. This is normally a
single sub-environment (e.g. a gym.Env) within a vectorized
setup.
env_context: The EnvContext to configure the environment.
Raises:
Exception in case something is wrong with the given environment.
"""
pass
@override(Trainable)
def _export_model(
self, export_formats: List[str], export_dir: str
) -> Dict[str, str]:
ExportFormat.validate(export_formats)
exported = {}
if ExportFormat.CHECKPOINT in export_formats:
path = os.path.join(export_dir, ExportFormat.CHECKPOINT)
self.export_policy_checkpoint(path)
exported[ExportFormat.CHECKPOINT] = path
if ExportFormat.MODEL in export_formats:
path = os.path.join(export_dir, ExportFormat.MODEL)
self.export_policy_model(path)
exported[ExportFormat.MODEL] = path
if ExportFormat.ONNX in export_formats:
path = os.path.join(export_dir, ExportFormat.ONNX)
self.export_policy_model(path, onnx=int(os.getenv("ONNX_OPSET", "11")))
exported[ExportFormat.ONNX] = path
return exported
[docs] def import_model(self, import_file: str):
"""Imports a model from import_file.
Note: Currently, only h5 files are supported.
Args:
import_file: The file to import the model from.
Returns:
A dict that maps ExportFormats to successfully exported models.
"""
# Check for existence.
if not os.path.exists(import_file):
raise FileNotFoundError(
"`import_file` '{}' does not exist! Can't import Model.".format(
import_file
)
)
# Get the format of the given file.
import_format = "h5" # TODO(sven): Support checkpoint loading.
ExportFormat.validate([import_format])
if import_format != ExportFormat.H5:
raise NotImplementedError
else:
return self.import_policy_model_from_h5(import_file)
@PublicAPI
def __getstate__(self) -> Dict:
"""Returns current state of Algorithm, sufficient to restore it from scratch.
Returns:
The current state dict of this Algorithm, which can be used to sufficiently
restore the algorithm from scratch without any other information.
"""
# Add config to state so complete Algorithm can be reproduced w/o it.
state = {
"algorithm_class": type(self),
"config": self.config,
}
if hasattr(self, "workers"):
state["worker"] = self.workers.local_worker().get_state()
# TODO: Experimental functionality: Store contents of replay buffer
# to checkpoint, only if user has configured this.
if self.local_replay_buffer is not None and self.config.get(
"store_buffer_in_checkpoints"
):
state["local_replay_buffer"] = self.local_replay_buffer.get_state()
if self.train_exec_impl is not None:
state["train_exec_impl"] = self.train_exec_impl.shared_metrics.get().save()
else:
state["counters"] = self._counters
return state
@PublicAPI
def __setstate__(self, state) -> None:
"""Sets the algorithm to the provided state.
Args:
state: The state dict to restore this Algorithm instance to. `state` may
have been returned by a call to an Algorithm's `__getstate__()` method.
"""
# TODO (sven): Validate that our config and the config in state are compatible.
# For example, the model architectures may differ.
# Also, what should the behavior be if e.g. some training parameter
# (e.g. lr) changed?
if hasattr(self, "workers") and "worker" in state:
self.workers.local_worker().set_state(state["worker"])
remote_state = ray.put(state["worker"])
self.workers.foreach_worker(
lambda w: w.set_state(ray.get(remote_state)),
local_worker=False,
healthy_only=False,
)
if self.evaluation_workers:
# If evaluation workers are used, also restore the policies
# there in case they are used for evaluation purpose.
self.evaluation_workers.foreach_worker(
lambda w: w.set_state(ray.get(remote_state)),
healthy_only=False,
)
# If necessary, restore replay data as well.
if self.local_replay_buffer is not None:
# TODO: Experimental functionality: Restore contents of replay
# buffer from checkpoint, only if user has configured this.
if self.config.get("store_buffer_in_checkpoints"):
if "local_replay_buffer" in state:
self.local_replay_buffer.set_state(state["local_replay_buffer"])
else:
logger.warning(
"`store_buffer_in_checkpoints` is True, but no replay "
"data found in state!"
)
elif "local_replay_buffer" in state and log_once(
"no_store_buffer_in_checkpoints_but_data_found"
):
logger.warning(
"`store_buffer_in_checkpoints` is False, but some replay "
"data found in state!"
)
if self.train_exec_impl is not None:
self.train_exec_impl.shared_metrics.get().restore(state["train_exec_impl"])
elif "counters" in state:
self._counters = state["counters"]
@staticmethod
def _checkpoint_info_to_algorithm_state(
checkpoint_info: dict,
policy_ids: Optional[Container[PolicyID]] = None,
policy_mapping_fn: Optional[Callable[[AgentID, EpisodeID], PolicyID]] = None,
policies_to_train: Optional[
Union[
Container[PolicyID],
Callable[[PolicyID, Optional[SampleBatchType]], bool],
]
] = None,
) -> Dict:
"""Converts a checkpoint info or object to a proper Algorithm state dict.
The returned state dict can be used inside self.__setstate__().
Args:
checkpoint_info: A checkpoint info dict as returned by
`ray.rllib.utils.checkpoints.get_checkpoint_info(
[checkpoint dir or AIR Checkpoint])`.
policy_ids: Optional list/set of PolicyIDs. If not None, only those policies
listed here will be included in the returned state. Note that
state items such as filters, the `is_policy_to_train` function, as
well as the multi-agent `policy_ids` dict will be adjusted as well,
based on this arg.
policy_mapping_fn: An optional (updated) policy mapping function
to include in the returned state.
policies_to_train: An optional list of policy IDs to be trained
or a callable taking PolicyID and SampleBatchType and
returning a bool (trainable or not?) to include in the returned state.
Returns:
The state dict usable within the `self.__setstate__()` method.
"""
if checkpoint_info["type"] != "Algorithm":
raise ValueError(
"`checkpoint` arg passed to "
"`Algorithm._checkpoint_info_to_algorithm_state()` must be an "
f"Algorithm checkpoint (but is {checkpoint_info['type']})!"
)
msgpack = None
if checkpoint_info.get("format") == "msgpack":
msgpack = try_import_msgpack(error=True)
with open(checkpoint_info["state_file"], "rb") as f:
if msgpack is not None:
state = msgpack.load(f)
else:
state = pickle.load(f)
# New checkpoint format: Policies are in separate sub-dirs.
# Note: Algorithms like ES/ARS don't have a WorkerSet, so we just return
# the plain state here.
if (
checkpoint_info["checkpoint_version"] > version.Version("0.1")
and state.get("worker") is not None
):
worker_state = state["worker"]
# Retrieve the set of all required policy IDs.
policy_ids = set(
policy_ids if policy_ids is not None else worker_state["policy_ids"]
)
# Remove those policies entirely from filters that are not in
# `policy_ids`.
worker_state["filters"] = {
pid: filter
for pid, filter in worker_state["filters"].items()
if pid in policy_ids
}
# Get Algorithm class.
if isinstance(state["algorithm_class"], str):
# Try deserializing from a full classpath.
# Or as a last resort: Tune registered algorithm name.
state["algorithm_class"] = deserialize_type(
state["algorithm_class"]
) or get_trainable_cls(state["algorithm_class"])
# Compile actual config object.
default_config = state["algorithm_class"].get_default_config()
if isinstance(default_config, AlgorithmConfig):
new_config = default_config.update_from_dict(state["config"])
else:
new_config = Algorithm.merge_trainer_configs(
default_config, state["config"]
)
# Remove policies from multiagent dict that are not in `policy_ids`.
new_policies = new_config.policies
if isinstance(new_policies, (set, list, tuple)):
new_policies = {pid for pid in new_policies if pid in policy_ids}
else:
new_policies = {
pid: spec for pid, spec in new_policies.items() if pid in policy_ids
}
new_config.multi_agent(
policies=new_policies,
policies_to_train=policies_to_train,
**(
{"policy_mapping_fn": policy_mapping_fn}
if policy_mapping_fn is not None
else {}
),
)
state["config"] = new_config
# Prepare local `worker` state to add policies' states into it,
# read from separate policy checkpoint files.
worker_state["policy_states"] = {}
for pid in policy_ids:
policy_state_file = os.path.join(
checkpoint_info["checkpoint_dir"],
"policies",
pid,
"policy_state."
+ ("msgpck" if checkpoint_info["format"] == "msgpack" else "pkl"),
)
if not os.path.isfile(policy_state_file):
raise ValueError(
"Given checkpoint does not seem to be valid! No policy "
f"state file found for PID={pid}. "
f"The file not found is: {policy_state_file}."
)
with open(policy_state_file, "rb") as f:
if msgpack is not None:
worker_state["policy_states"][pid] = msgpack.load(f)
else:
worker_state["policy_states"][pid] = pickle.load(f)
# These two functions are never serialized in a msgpack checkpoint (which
# does not store code, unlike a cloudpickle checkpoint). Hence the user has
# to provide them with the `Algorithm.from_checkpoint()` call.
if policy_mapping_fn is not None:
worker_state["policy_mapping_fn"] = policy_mapping_fn
if (
policies_to_train is not None
# `policies_to_train` might be left None in case all policies should be
# trained.
or worker_state["is_policy_to_train"] == NOT_SERIALIZABLE
):
worker_state["is_policy_to_train"] = policies_to_train
return state
@DeveloperAPI
def _create_local_replay_buffer_if_necessary(
self, config: PartialAlgorithmConfigDict
) -> Optional[MultiAgentReplayBuffer]:
"""Create a MultiAgentReplayBuffer instance if necessary.
Args:
config: Algorithm-specific configuration data.
Returns:
MultiAgentReplayBuffer instance based on algorithm config.
None, if local replay buffer is not needed.
"""
if not config.get("replay_buffer_config") or config["replay_buffer_config"].get(
"no_local_replay_buffer"
):
return
return from_config(ReplayBuffer, config["replay_buffer_config"])
@DeveloperAPI
def _kwargs_for_execution_plan(self):
kwargs = {}
if self.local_replay_buffer is not None:
kwargs["local_replay_buffer"] = self.local_replay_buffer
return kwargs
def _run_one_training_iteration(self) -> Tuple[ResultDict, "TrainIterCtx"]:
"""Runs one training iteration (self.iteration will be +1 after this).
Calls `self.training_step()` repeatedly until the minimum time (sec),
sample- or training steps have been reached.
Returns:
The results dict from the training iteration.
"""
# In case we are training (in a thread) parallel to evaluation,
# we may have to re-enable eager mode here (gets disabled in the
# thread).
if self.config.get("framework") == "tf2" and not tf.executing_eagerly():
tf1.enable_eager_execution()
results = None
# Create a step context ...
with TrainIterCtx(algo=self) as train_iter_ctx:
# .. so we can query it whether we should stop the iteration loop (e.g.
# when we have reached `min_time_s_per_iteration`).
while not train_iter_ctx.should_stop(results):
# Try to train one step.
# TODO (avnishn): Remove the execution plan API by q1 2023
with self._timers[TRAINING_ITERATION_TIMER]:
if self.config._disable_execution_plan_api:
results = self.training_step()
else:
results = next(self.train_exec_impl)
# With training step done. Try to bring failed workers back.
self.restore_workers(self.workers)
return results, train_iter_ctx
def _run_one_evaluation(
self,
train_future: Optional[concurrent.futures.ThreadPoolExecutor] = None,
) -> ResultDict:
"""Runs evaluation step via `self.evaluate()` and handling worker failures.
Args:
train_future: In case, we are training and avaluating in parallel,
this arg carries the currently running ThreadPoolExecutor
object that runs the training iteration
Returns:
The results dict from the evaluation call.
"""
eval_func_to_use = (
self._evaluate_async
if self.config.enable_async_evaluation
else self.evaluate
)
if self.config.evaluation_duration == "auto":
assert (
train_future is not None and self.config.evaluation_parallel_to_training
)
unit = self.config.evaluation_duration_unit
eval_results = eval_func_to_use(
duration_fn=functools.partial(
self._automatic_evaluation_duration_fn,
unit,
self.config.evaluation_num_workers,
self.evaluation_config,
train_future,
)
)
# Run `self.evaluate()` only once per training iteration.
else:
eval_results = eval_func_to_use()
if self.evaluation_workers is not None:
# After evaluation, do a round of health check to see if any of
# the failed workers are back.
self.restore_workers(self.evaluation_workers)
# Add number of healthy evaluation workers after this iteration.
eval_results["evaluation"][
"num_healthy_workers"
] = self.evaluation_workers.num_healthy_remote_workers()
eval_results["evaluation"][
"num_in_flight_async_reqs"
] = self.evaluation_workers.num_in_flight_async_reqs()
eval_results["evaluation"][
"num_remote_worker_restarts"
] = self.evaluation_workers.num_remote_worker_restarts()
return eval_results
def _run_one_training_iteration_and_evaluation_in_parallel(
self,
) -> Tuple[ResultDict, "TrainIterCtx"]:
"""Runs one training iteration and one evaluation step in parallel.
First starts the training iteration (via `self._run_one_training_iteration()`)
within a ThreadPoolExecutor, then runs the evaluation step in parallel.
In auto-duration mode (config.evaluation_duration=auto), makes sure the
evaluation step takes roughly the same time as the training iteration.
Returns:
The accumulated training and evaluation results.
"""
with concurrent.futures.ThreadPoolExecutor() as executor:
train_future = executor.submit(lambda: self._run_one_training_iteration())
# Pass the train_future into `self._run_one_evaluation()` to allow it
# to run exactly as long as the training iteration takes in case
# evaluation_duration=auto.
results = self._run_one_evaluation(train_future)
# Collect the training results from the future.
train_results, train_iter_ctx = train_future.result()
results.update(train_results)
return results, train_iter_ctx
def _run_offline_evaluation(self):
"""Runs offline evaluation via `OfflineEvaluator.estimate_on_dataset()` API.
This method will be used when `evaluation_dataset` is provided.
Note: This will only work if the policy is a single agent policy.
Returns:
The results dict from the offline evaluation call.
"""
assert len(self.workers.local_worker().policy_map) == 1
parallelism = self.evaluation_config.evaluation_num_workers or 1
offline_eval_results = {"off_policy_estimator": {}}
for evaluator_name, offline_evaluator in self.reward_estimators.items():
offline_eval_results["off_policy_estimator"][
evaluator_name
] = offline_evaluator.estimate_on_dataset(
self.evaluation_dataset,
n_parallelism=parallelism,
)
return offline_eval_results
@classmethod
def _should_create_evaluation_rollout_workers(cls, eval_config: "AlgorithmConfig"):
"""Determines whether we need to create evaluation workers.
Returns False if we need to run offline evaluation
(with ope.estimate_on_dastaset API) or when local worker is to be used for
evaluation. Note: We only use estimate_on_dataset API with bandits for now.
That is when ope_split_batch_by_episode is False.
TODO: In future we will do the same for episodic RL OPE.
"""
run_offline_evaluation = (
eval_config.off_policy_estimation_methods
and not eval_config.ope_split_batch_by_episode
)
return not run_offline_evaluation and (
eval_config.evaluation_num_workers > 0 or eval_config.evaluation_interval
)
@staticmethod
def _automatic_evaluation_duration_fn(
unit, num_eval_workers, eval_cfg, train_future, num_units_done
):
# Training is done and we already ran at least one
# evaluation -> Nothing left to run.
if num_units_done > 0 and train_future.done():
return 0
# Count by episodes. -> Run n more
# (n=num eval workers).
elif unit == "episodes":
return num_eval_workers
# Count by timesteps. -> Run n*m*p more
# (n=num eval workers; m=rollout fragment length;
# p=num-envs-per-worker).
else:
return (
num_eval_workers
* eval_cfg["rollout_fragment_length"]
* eval_cfg["num_envs_per_worker"]
)
def _compile_iteration_results(
self, *, episodes_this_iter, step_ctx, iteration_results=None
):
# Return dict.
results: ResultDict = {}
iteration_results = iteration_results or {}
# Evaluation results.
if "evaluation" in iteration_results:
results["evaluation"] = iteration_results.pop("evaluation")
# Custom metrics and episode media.
results["custom_metrics"] = iteration_results.pop("custom_metrics", {})
results["episode_media"] = iteration_results.pop("episode_media", {})
# Learner info.
results["info"] = {LEARNER_INFO: iteration_results}
# Calculate how many (if any) of older, historical episodes we have to add to
# `episodes_this_iter` in order to reach the required smoothing window.
episodes_for_metrics = episodes_this_iter[:]
missing = self.config.metrics_num_episodes_for_smoothing - len(
episodes_this_iter
)
# We have to add some older episodes to reach the smoothing window size.
if missing > 0:
episodes_for_metrics = self._episode_history[-missing:] + episodes_this_iter
assert (
len(episodes_for_metrics)
<= self.config.metrics_num_episodes_for_smoothing
)
# Note that when there are more than `metrics_num_episodes_for_smoothing`
# episodes in `episodes_for_metrics`, leave them as-is. In this case, we'll
# compute the stats over that larger number.
# Add new episodes to our history and make sure it doesn't grow larger than
# needed.
self._episode_history.extend(episodes_this_iter)
self._episode_history = self._episode_history[
-self.config.metrics_num_episodes_for_smoothing :
]
results["sampler_results"] = summarize_episodes(
episodes_for_metrics,
episodes_this_iter,
self.config.keep_per_episode_custom_metrics,
)
# TODO: Don't dump sampler results into top-level.
results.update(results["sampler_results"])
results["num_healthy_workers"] = self.workers.num_healthy_remote_workers()
results["num_in_flight_async_reqs"] = self.workers.num_in_flight_async_reqs()
results[
"num_remote_worker_restarts"
] = self.workers.num_remote_worker_restarts()
# Train-steps- and env/agent-steps this iteration.
for c in [
NUM_AGENT_STEPS_SAMPLED,
NUM_AGENT_STEPS_TRAINED,
NUM_ENV_STEPS_SAMPLED,
NUM_ENV_STEPS_TRAINED,
]:
results[c] = self._counters[c]
time_taken_sec = step_ctx.get_time_taken_sec()
if self.config.count_steps_by == "agent_steps":
results[NUM_AGENT_STEPS_SAMPLED + "_this_iter"] = step_ctx.sampled
results[NUM_AGENT_STEPS_TRAINED + "_this_iter"] = step_ctx.trained
results[NUM_AGENT_STEPS_SAMPLED + "_throughput_per_sec"] = (
step_ctx.sampled / time_taken_sec
)
results[NUM_AGENT_STEPS_TRAINED + "_throughput_per_sec"] = (
step_ctx.trained / time_taken_sec
)
# TODO: For CQL and other algos, count by trained steps.
results["timesteps_total"] = self._counters[NUM_AGENT_STEPS_SAMPLED]
else:
results[NUM_ENV_STEPS_SAMPLED + "_this_iter"] = step_ctx.sampled
results[NUM_ENV_STEPS_TRAINED + "_this_iter"] = step_ctx.trained
results[NUM_ENV_STEPS_SAMPLED + "_throughput_per_sec"] = (
step_ctx.sampled / time_taken_sec
)
results[NUM_ENV_STEPS_TRAINED + "_throughput_per_sec"] = (
step_ctx.trained / time_taken_sec
)
# TODO: For CQL and other algos, count by trained steps.
results["timesteps_total"] = self._counters[NUM_ENV_STEPS_SAMPLED]
# TODO: Backward compatibility.
results[STEPS_TRAINED_THIS_ITER_COUNTER] = step_ctx.trained
results["agent_timesteps_total"] = self._counters[NUM_AGENT_STEPS_SAMPLED]
# Process timer results.
timers = {}
for k, timer in self._timers.items():
timers["{}_time_ms".format(k)] = round(timer.mean * 1000, 3)
if timer.has_units_processed():
timers["{}_throughput".format(k)] = round(timer.mean_throughput, 3)
results["timers"] = timers
# Process counter results.
counters = {}
for k, counter in self._counters.items():
counters[k] = counter
results["counters"] = counters
# TODO: Backward compatibility.
results["info"].update(counters)
return results
def __repr__(self):
return type(self).__name__
def _record_usage(self, config):
"""Record the framework and algorithm used.
Args:
config: Algorithm config dict.
"""
record_extra_usage_tag(TagKey.RLLIB_FRAMEWORK, config["framework"])
record_extra_usage_tag(TagKey.RLLIB_NUM_WORKERS, str(config["num_workers"]))
alg = self.__class__.__name__
# We do not want to collect user defined algorithm names.
if alg not in ALL_ALGORITHMS:
alg = "USER_DEFINED"
record_extra_usage_tag(TagKey.RLLIB_ALGORITHM, alg)
@Deprecated(new="Algorithm.compute_single_action()", error=True)
def compute_action(self, *args, **kwargs):
return self.compute_single_action(*args, **kwargs)
@Deprecated(new="construct WorkerSet(...) instance directly", error=True)
def _make_workers(self, *args, **kwargs):
pass
@Deprecated(new="AlgorithmConfig.validate()", error=False)
def validate_config(self, config):
pass
@staticmethod
@Deprecated(new="AlgorithmConfig.validate()", error=True)
def _validate_config(config, trainer_or_none):
pass
# TODO: Create a dict that throw a deprecation warning once we have fully moved
# to AlgorithmConfig() objects (some algos still missing).
COMMON_CONFIG: AlgorithmConfigDict = AlgorithmConfig(Algorithm).to_dict()
class TrainIterCtx:
def __init__(self, algo: Algorithm):
self.algo = algo
self.time_start = None
self.time_stop = None
def __enter__(self):
# Before first call to `step()`, `results` is expected to be None ->
# Start with self.failures=-1 -> set to 0 before the very first call
# to `self.step()`.
self.failures = -1
self.time_start = time.time()
self.sampled = 0
self.trained = 0
self.init_env_steps_sampled = self.algo._counters[NUM_ENV_STEPS_SAMPLED]
self.init_env_steps_trained = self.algo._counters[NUM_ENV_STEPS_TRAINED]
self.init_agent_steps_sampled = self.algo._counters[NUM_AGENT_STEPS_SAMPLED]
self.init_agent_steps_trained = self.algo._counters[NUM_AGENT_STEPS_TRAINED]
self.failure_tolerance = self.algo.config[
"num_consecutive_worker_failures_tolerance"
]
return self
def __exit__(self, *args):
self.time_stop = time.time()
def get_time_taken_sec(self) -> float:
"""Returns the time we spent in the context in seconds."""
return self.time_stop - self.time_start
def should_stop(self, results):
# Before first call to `step()`.
if results is None:
# Fail after n retries.
self.failures += 1
if self.failures > self.failure_tolerance:
raise RuntimeError(
"More than `num_consecutive_worker_failures_tolerance="
f"{self.failure_tolerance}` consecutive worker failures! "
"Exiting."
)
# Continue to very first `step()` call or retry `step()` after
# a (tolerable) failure.
return False
# Stopping criteria.
elif self.algo.config._disable_execution_plan_api:
if self.algo.config.count_steps_by == "agent_steps":
self.sampled = (
self.algo._counters[NUM_AGENT_STEPS_SAMPLED]
- self.init_agent_steps_sampled
)
self.trained = (
self.algo._counters[NUM_AGENT_STEPS_TRAINED]
- self.init_agent_steps_trained
)
else:
self.sampled = (
self.algo._counters[NUM_ENV_STEPS_SAMPLED]
- self.init_env_steps_sampled
)
self.trained = (
self.algo._counters[NUM_ENV_STEPS_TRAINED]
- self.init_env_steps_trained
)
min_t = self.algo.config["min_time_s_per_iteration"]
min_sample_ts = self.algo.config["min_sample_timesteps_per_iteration"]
min_train_ts = self.algo.config["min_train_timesteps_per_iteration"]
# Repeat if not enough time has passed or if not enough
# env|train timesteps have been processed (or these min
# values are not provided by the user).
if (
(not min_t or time.time() - self.time_start >= min_t)
and (not min_sample_ts or self.sampled >= min_sample_ts)
and (not min_train_ts or self.trained >= min_train_ts)
):
return True
else:
return False
# No errors (we got results != None) -> Return True
# (meaning: yes, should stop -> no further step attempts).
else:
return True