def setup(self, config: PartialTrainerConfigDict):
super().setup(config)
# Initialize torch process group for
if self.config["_disable_execution_plan_api"] is True:
self._curr_learner_info = {}
ip = ray.get(self.workers.remote_workers()[0].get_node_ip.remote())
port = ray.get(
self.workers.remote_workers()[0].find_free_port.remote())
address = "tcp://{ip}:{port}".format(ip=ip, port=port)
logger.info(
"Creating torch process group with leader {}".format(address))
# Get setup tasks in order to throw errors on failure.
ray.get([
worker.setup_torch_data_parallel.remote(
url=address,
world_rank=i,
world_size=len(self.workers.remote_workers()),
backend=self.config["torch_distributed_backend"],
) for i, worker in enumerate(self.workers.remote_workers())
])
logger.info("Torch process group init completed")
self._ddppo_worker_manager = AsyncRequestsManager(
self.workers.remote_workers(),
max_remote_requests_in_flight_per_worker=1,
ray_wait_timeout_s=0.03,
)
def test_test_async_requests_task_doesnt_buffering(self):
"""Tests that the async manager drops"""
workers = [RemoteRLlibActor.remote(sleep_time=0.1) for _ in range(2)]
manager = AsyncRequestsManager(
workers, max_remote_requests_in_flight_per_worker=2
)
for i in range(8):
scheduled = manager.call(lambda w: w.task())
if i < 4:
assert scheduled, "We should have scheduled the task"
else:
assert not scheduled, (
"We should not have scheduled the task because"
" all workers are busy."
)
assert len(manager._pending_remotes) == 4, "We should have 4 pending requests"
time.sleep(3)
ready_requests = manager.get_ready()
for worker in workers:
if not len(ready_requests[worker]) == 2:
raise Exception(
"We should return the 2 ready requests in this case from each "
"actors."
)
for _ in range(4):
manager.call(lambda w: w.task())
# new tasks scheduled from the buffer
time.sleep(3)
ready_requests = manager.get_ready()
for worker in workers:
if not len(ready_requests[worker]) == 2:
raise Exception(
"We should return the 2 ready requests in this case from each "
"actors"
)
def test_round_robin_scheduling(self):
"""Test that the async manager schedules actors in a round robin fashion"""
workers = [RemoteRLlibActor.remote(sleep_time=0.1) for _ in range(2)]
manager = AsyncRequestsManager(
workers, max_remote_requests_in_flight_per_worker=2
)
for i in range(4):
scheduled_actor = workers[i % len(workers)]
manager.call(lambda w: w.task())
if i < 2:
assert len(manager._remote_requests_in_flight[scheduled_actor]) == 1, (
"We should have 1 request in flight for the actor that we just "
"scheduled on"
)
else:
assert len(manager._remote_requests_in_flight[scheduled_actor]) == 2, (
"We should have 2 request in flight for the actor that we just "
"scheduled on"
)
def test_high_load(self):
workers = [
RemoteRLlibActor.remote(sleep_time=random.random() * 2.0) for _ in range(60)
]
manager = AsyncRequestsManager(
workers,
max_remote_requests_in_flight_per_worker=2,
return_object_refs=True,
ray_wait_timeout_s=0.0,
)
num_ready = 0
for i in range(2000):
manager.call_on_all_available(lambda w: w.task())
time.sleep(0.01)
ready = manager.get_ready()
for reqs in ready.values():
num_ready += len(reqs)
ray.get(reqs)
for worker in ready.keys():
worker.task2.remote(1, 3)
time.sleep(20)
ready = manager.get_ready()
num_ready += sum(len(reqs) for reqs in ready.values())
actually_called = sum(
ray.get(
[worker.apply.remote(lambda w: w.num_task_called) for worker in workers]
)
)
assert actually_called == num_ready, (actually_called, num_ready)
def test_call_to_actor(self):
workers = [RemoteRLlibActor.remote(sleep_time=0.1) for _ in range(2)]
worker_not_in_manager = RemoteRLlibActor.remote(sleep_time=0.1)
manager = AsyncRequestsManager(
workers, max_remote_requests_in_flight_per_worker=2)
manager.call(lambda w: w.task(), actor=workers[0])
time.sleep(3)
results = manager.get_ready()
if not len(results) == 1 and workers[0] not in results:
raise Exception(
"We should return the 1 ready requests in this case from the worker we "
"called to")
with pytest.raises(ValueError,
match=".*has not been added to the manager.*"):
manager.call(lambda w: w.task(), actor=worker_not_in_manager)
def test_args_kwargs(self):
"""Tests that the async manager can properly handle actors with tasks that
vary in the amount of time that they take to run"""
workers = [RemoteRLlibActor.remote(sleep_time=0.1)]
manager = AsyncRequestsManager(
workers, max_remote_requests_in_flight_per_worker=2)
for _ in range(2):
manager.call(lambda w, a, b: w.task2(a, b), fn_args=[1, 2])
time.sleep(3)
if not len(manager.get_ready()[workers[0]]) == 2:
raise Exception(
"We should return the 2 ready requests in this case from the actors"
" that have shorter tasks")
for _ in range(2):
manager.call(lambda w, a, b: w.task2(a, b),
fn_kwargs=dict(a=1, b=2))
time.sleep(3)
if not len(manager.get_ready()[workers[0]]) == 2:
raise Exception(
"We should return the 2 ready requests in this case from the actors"
" that have longer tasks")
def test_async_requests_manager_num_returns(self):
"""Tests that an async manager can properly handle actors with tasks that
vary in the amount of time that they take to run"""
workers = [RemoteRLlibActor.remote(sleep_time=0.1) for _ in range(2)]
workers += [RemoteRLlibActor.remote(sleep_time=5) for _ in range(2)]
manager = AsyncRequestsManager(
workers, max_remote_requests_in_flight_per_worker=1)
for _ in range(4):
manager.call(lambda w: w.task())
time.sleep(3)
if not len(manager.get_ready()) == 2:
raise Exception(
"We should return the 2 ready requests in this case from the actors"
" that have shorter tasks")
time.sleep(7)
if not len(manager.get_ready()) == 2:
raise Exception(
"We should return the 2 ready requests in this case from the actors"
" that have longer tasks")
def setup(self, config: PartialAlgorithmConfigDict):
super().setup(config)
# Shortcut: If execution_plan, thread and buffer will be created in there.
if self.config["_disable_execution_plan_api"] is False:
return
# Tag those workers (top 1/3rd indices) that we should collect episodes from
# for metrics due to `PerWorkerEpsilonGreedy` exploration strategy.
if self.workers.remote_workers():
self._remote_workers_for_metrics = self.workers.remote_workers(
)[-len(self.workers.remote_workers()) // 3:]
num_replay_buffer_shards = self.config["optimizer"][
"num_replay_buffer_shards"]
# Create copy here so that we can modify without breaking other logic
replay_actor_config = copy.deepcopy(
self.config["replay_buffer_config"])
replay_actor_config["capacity"] = (
self.config["replay_buffer_config"]["capacity"] //
num_replay_buffer_shards)
ReplayActor = ray.remote(num_cpus=0)(replay_actor_config["type"])
# Place all replay buffer shards on the same node as the learner
# (driver process that runs this execution plan).
if replay_actor_config["replay_buffer_shards_colocated_with_driver"]:
self._replay_actors = create_colocated_actors(
actor_specs=[ # (class, args, kwargs={}, count)
(
ReplayActor,
None,
replay_actor_config,
num_replay_buffer_shards,
)
],
node=platform.node(), # localhost
)[0] # [0]=only one item in `actor_specs`.
# Place replay buffer shards on any node(s).
else:
self._replay_actors = [
ReplayActor.remote(*replay_actor_config)
for _ in range(num_replay_buffer_shards)
]
self._replay_actor_manager = AsyncRequestsManager(
self._replay_actors,
max_remote_requests_in_flight_per_worker=self.
config["max_requests_in_flight_per_replay_worker"],
ray_wait_timeout_s=self.config["timeout_s_replay_manager"],
)
self._sampling_actor_manager = AsyncRequestsManager(
self.workers.remote_workers(),
max_remote_requests_in_flight_per_worker=self.
config["max_requests_in_flight_per_sampler_worker"],
ray_wait_timeout_s=self.config["timeout_s_sampler_manager"],
)
self.learner_thread = LearnerThread(self.workers.local_worker())
self.learner_thread.start()
self.steps_since_update = defaultdict(int)
weights = self.workers.local_worker().get_weights()
self.curr_learner_weights = ray.put(weights)
self.curr_num_samples_collected = 0
self.replay_sample_batches = []
self._num_ts_trained_since_last_target_update = 0
class ApexDQN(DQN):
@override(Trainable)
def setup(self, config: PartialAlgorithmConfigDict):
super().setup(config)
# Shortcut: If execution_plan, thread and buffer will be created in there.
if self.config["_disable_execution_plan_api"] is False:
return
# Tag those workers (top 1/3rd indices) that we should collect episodes from
# for metrics due to `PerWorkerEpsilonGreedy` exploration strategy.
if self.workers.remote_workers():
self._remote_workers_for_metrics = self.workers.remote_workers(
)[-len(self.workers.remote_workers()) // 3:]
num_replay_buffer_shards = self.config["optimizer"][
"num_replay_buffer_shards"]
# Create copy here so that we can modify without breaking other logic
replay_actor_config = copy.deepcopy(
self.config["replay_buffer_config"])
replay_actor_config["capacity"] = (
self.config["replay_buffer_config"]["capacity"] //
num_replay_buffer_shards)
ReplayActor = ray.remote(num_cpus=0)(replay_actor_config["type"])
# Place all replay buffer shards on the same node as the learner
# (driver process that runs this execution plan).
if replay_actor_config["replay_buffer_shards_colocated_with_driver"]:
self._replay_actors = create_colocated_actors(
actor_specs=[ # (class, args, kwargs={}, count)
(
ReplayActor,
None,
replay_actor_config,
num_replay_buffer_shards,
)
],
node=platform.node(), # localhost
)[0] # [0]=only one item in `actor_specs`.
# Place replay buffer shards on any node(s).
else:
self._replay_actors = [
ReplayActor.remote(*replay_actor_config)
for _ in range(num_replay_buffer_shards)
]
self._replay_actor_manager = AsyncRequestsManager(
self._replay_actors,
max_remote_requests_in_flight_per_worker=self.
config["max_requests_in_flight_per_replay_worker"],
ray_wait_timeout_s=self.config["timeout_s_replay_manager"],
)
self._sampling_actor_manager = AsyncRequestsManager(
self.workers.remote_workers(),
max_remote_requests_in_flight_per_worker=self.
config["max_requests_in_flight_per_sampler_worker"],
ray_wait_timeout_s=self.config["timeout_s_sampler_manager"],
)
self.learner_thread = LearnerThread(self.workers.local_worker())
self.learner_thread.start()
self.steps_since_update = defaultdict(int)
weights = self.workers.local_worker().get_weights()
self.curr_learner_weights = ray.put(weights)
self.curr_num_samples_collected = 0
self.replay_sample_batches = []
self._num_ts_trained_since_last_target_update = 0
@classmethod
@override(DQN)
def get_default_config(cls) -> AlgorithmConfigDict:
return ApexDQNConfig().to_dict()
@override(DQN)
def validate_config(self, config):
if config["num_gpus"] > 1:
raise ValueError("`num_gpus` > 1 not yet supported for APEX-DQN!")
# Call DQN's validation method.
super().validate_config(config)
@override(DQN)
def training_step(self) -> ResultDict:
num_samples_ready_dict = self.get_samples_and_store_to_replay_buffers()
worker_samples_collected = defaultdict(int)
for worker, samples_infos in num_samples_ready_dict.items():
for samples_info in samples_infos:
self._counters[NUM_AGENT_STEPS_SAMPLED] += samples_info[
"agent_steps"]
self._counters[NUM_ENV_STEPS_SAMPLED] += samples_info[
"env_steps"]
worker_samples_collected[worker] += samples_info["agent_steps"]
# update the weights of the workers that returned samples
# only do this if there are remote workers (config["num_workers"] > 1)
if self.workers.remote_workers():
self.update_workers(worker_samples_collected)
# trigger a sample from the replay actors and enqueue operation to the
# learner thread.
self.sample_from_replay_buffer_place_on_learner_queue_non_blocking(
worker_samples_collected)
self.update_replay_sample_priority()
return copy.deepcopy(self.learner_thread.learner_info)
def get_samples_and_store_to_replay_buffers(self):
# in the case the num_workers = 0
if not self.workers.remote_workers():
with self._timers[SAMPLE_TIMER]:
local_sampling_worker = self.workers.local_worker()
batch = local_sampling_worker.sample()
actor = random.choice(self._replay_actors)
ray.get(actor.add.remote(batch))
batch_statistics = {
local_sampling_worker: [{
"agent_steps": batch.agent_steps(),
"env_steps": batch.env_steps(),
}]
}
return batch_statistics
def remote_worker_sample_and_store(worker: RolloutWorker,
replay_actors: List[ActorHandle]):
# This function is run as a remote function on sampling workers,
# and should only be used with the RolloutWorker's apply function ever.
# It is used to gather samples, and trigger the operation to store them to
# replay actors from the rollout worker instead of returning the obj
# refs for the samples to the driver process and doing the sampling
# operation on there.
_batch = worker.sample()
_actor = random.choice(replay_actors)
_actor.add.remote(_batch)
_batch_statistics = {
"agent_steps": _batch.agent_steps(),
"env_steps": _batch.env_steps(),
}
return _batch_statistics
# Sample and Store in the Replay Actors on the sampling workers.
with self._timers[SAMPLE_TIMER]:
self._sampling_actor_manager.call_on_all_available(
remote_worker_sample_and_store,
fn_kwargs={"replay_actors": self._replay_actors},
)
num_samples_ready_dict = self._sampling_actor_manager.get_ready()
return num_samples_ready_dict
def update_workers(self, _num_samples_ready: Dict[ActorHandle,
int]) -> int:
"""Update the remote workers that have samples ready.
Args:
_num_samples_ready: A mapping from ActorHandle (RolloutWorker) to
the number of samples returned by the remote worker.
Returns:
The number of remote workers whose weights were updated.
"""
max_steps_weight_sync_delay = self.config["optimizer"][
"max_weight_sync_delay"]
# Update our local copy of the weights if the learner thread has updated
# the learner worker's weights
if self.learner_thread.weights_updated:
self.learner_thread.weights_updated = False
weights = self.workers.local_worker().get_weights()
self.curr_learner_weights = ray.put(weights)
with self._timers[SYNCH_WORKER_WEIGHTS_TIMER]:
for (
remote_sampler_worker,
num_samples_collected,
) in _num_samples_ready.items():
self.steps_since_update[
remote_sampler_worker] += num_samples_collected
if (self.steps_since_update[remote_sampler_worker] >=
max_steps_weight_sync_delay):
remote_sampler_worker.set_weights.remote(
self.curr_learner_weights,
{"timestep": self._counters[STEPS_TRAINED_COUNTER]},
)
self.steps_since_update[remote_sampler_worker] = 0
self._counters["num_weight_syncs"] += 1
def sample_from_replay_buffer_place_on_learner_queue_non_blocking(
self, num_samples_collected: Dict[ActorHandle, int]) -> None:
"""Get samples from the replay buffer and place them on the learner queue.
Args:
num_samples_collected: A mapping from ActorHandle (RolloutWorker) to
number of samples returned by the remote worker. This is used to
implement training intensity which is the concept of triggering a
certain amount of training based on the number of samples that have
been collected since the last time that training was triggered.
"""
def wait_on_replay_actors() -> None:
"""Wait for the replay actors to finish sampling for timeout seconds.
If the timeout is None, then block on the actors indefinitely.
"""
_replay_samples_ready = self._replay_actor_manager.get_ready()
for _replay_actor, _sample_batches in _replay_samples_ready.items(
):
for _sample_batch in _sample_batches:
self.replay_sample_batches.append(
(_replay_actor, _sample_batch))
num_samples_collected = sum(num_samples_collected.values())
self.curr_num_samples_collected += num_samples_collected
wait_on_replay_actors()
if self.curr_num_samples_collected >= self.config["train_batch_size"]:
training_intensity = int(self.config["training_intensity"] or 1)
num_requests_to_launch = (
self.curr_num_samples_collected /
self.config["train_batch_size"]) * training_intensity
num_requests_to_launch = max(1, round(num_requests_to_launch))
self.curr_num_samples_collected = 0
for _ in range(num_requests_to_launch):
self._replay_actor_manager.call(
lambda actor, num_items: actor.sample(num_items),
fn_args=[self.config["train_batch_size"]],
)
wait_on_replay_actors()
# add the sample batches to the learner queue
while self.replay_sample_batches:
try:
item = self.replay_sample_batches[0]
# the replay buffer returns none if it has not been filled to
# the minimum threshold yet.
if item:
self.learner_thread.inqueue.put(
self.replay_sample_batches[0], timeout=0.001)
self.replay_sample_batches.pop(0)
except queue.Full:
break
def update_replay_sample_priority(self) -> None:
"""Update the priorities of the sample batches with new priorities that are
computed by the learner thread.
"""
num_samples_trained_this_itr = 0
for _ in range(self.learner_thread.outqueue.qsize()):
if self.learner_thread.is_alive():
(
replay_actor,
priority_dict,
env_steps,
agent_steps,
) = self.learner_thread.outqueue.get(timeout=0.001)
if (self.config["replay_buffer_config"].get(
"prioritized_replay_alpha") > 0):
replay_actor.update_priorities.remote(priority_dict)
num_samples_trained_this_itr += env_steps
self.update_target_networks(env_steps)
self._counters[NUM_ENV_STEPS_TRAINED] += env_steps
self._counters[NUM_AGENT_STEPS_TRAINED] += agent_steps
self.workers.local_worker().set_global_vars(
{"timestep": self._counters[NUM_ENV_STEPS_TRAINED]})
else:
raise RuntimeError("The learner thread died in while training")
self._counters[
STEPS_TRAINED_THIS_ITER_COUNTER] = num_samples_trained_this_itr
self._timers["learner_dequeue"] = self.learner_thread.queue_timer
self._timers["learner_grad"] = self.learner_thread.grad_timer
self._timers["learner_overall"] = self.learner_thread.overall_timer
def update_target_networks(self, num_new_trained_samples) -> None:
"""Update the target networks."""
self._num_ts_trained_since_last_target_update += num_new_trained_samples
if (self._num_ts_trained_since_last_target_update >=
self.config["target_network_update_freq"]):
self._num_ts_trained_since_last_target_update = 0
with self._timers[TARGET_NET_UPDATE_TIMER]:
to_update = self.workers.local_worker().get_policies_to_train()
self.workers.local_worker().foreach_policy_to_train(
lambda p, pid: pid in to_update and p.update_target())
self._counters[NUM_TARGET_UPDATES] += 1
self._counters[LAST_TARGET_UPDATE_TS] = self._counters[
STEPS_TRAINED_COUNTER]
@override(Algorithm)
def on_worker_failures(self, removed_workers: List[ActorHandle],
new_workers: List[ActorHandle]):
"""Handle the failures of remote sampling workers
Args:
removed_workers: removed worker ids.
new_workers: ids of newly created workers.
"""
if self.config["_disable_execution_plan_api"]:
self._sampling_actor_manager.remove_workers(
removed_workers, remove_in_flight_requests=True)
self._sampling_actor_manager.add_workers(new_workers)
@override(Algorithm)
def _compile_iteration_results(self, *, step_ctx, iteration_results=None):
result = super()._compile_iteration_results(
step_ctx=step_ctx, iteration_results=iteration_results)
replay_stats = ray.get(self._replay_actors[0].stats.remote(
self.config["optimizer"].get("debug")))
exploration_infos_list = self.workers.foreach_policy_to_train(
lambda p, pid: {pid: p.get_exploration_state()})
exploration_infos = {}
for info in exploration_infos_list:
# we're guaranteed that each info has policy ids that are unique
exploration_infos.update(info)
other_results = {
"exploration_infos": exploration_infos,
"learner_queue": self.learner_thread.learner_queue_size.stats(),
"replay_shard_0": replay_stats,
}
result["info"].update(other_results)
return result
@classmethod
@override(Algorithm)
def default_resource_request(cls, config):
cf = dict(cls.get_default_config(), **config)
eval_config = cf["evaluation_config"]
# Return PlacementGroupFactory containing all needed resources
# (already properly defined as device bundles).
return PlacementGroupFactory(
bundles=[{
# Local worker + replay buffer actors.
# Force replay buffers to be on same node to maximize
# data bandwidth between buffers and the learner (driver).
# Replay buffer actors each contain one shard of the total
# replay buffer and use 1 CPU each.
"CPU":
cf["num_cpus_for_driver"] +
cf["optimizer"]["num_replay_buffer_shards"],
"GPU":
0 if cf["_fake_gpus"] else cf["num_gpus"],
}] + [
{
# RolloutWorkers.
"CPU": cf["num_cpus_per_worker"],
"GPU": cf["num_gpus_per_worker"],
**cf["custom_resources_per_worker"],
} for _ in range(cf["num_workers"])
] + ([
{
# Evaluation workers.
# Note: The local eval worker is located on the driver
# CPU.
"CPU":
eval_config.get("num_cpus_per_worker",
cf["num_cpus_per_worker"]),
"GPU":
eval_config.get("num_gpus_per_worker",
cf["num_gpus_per_worker"]),
**eval_config.get(
"custom_resources_per_worker",
cf["custom_resources_per_worker"],
),
} for _ in range(cf["evaluation_num_workers"])
] if cf["evaluation_interval"] else []),
strategy=config.get("placement_strategy", "PACK"),
)
def setup(self, config: PartialTrainerConfigDict):
super().setup(config)
self._worker_manager = AsyncRequestsManager(
self.workers.remote_workers(),
max_remote_requests_in_flight_per_worker=1)
class A3C(Trainer):
@classmethod
@override(Trainer)
def get_default_config(cls) -> TrainerConfigDict:
return A3CConfig().to_dict()
@override(Trainer)
def setup(self, config: PartialTrainerConfigDict):
super().setup(config)
self._worker_manager = AsyncRequestsManager(
self.workers.remote_workers(),
max_remote_requests_in_flight_per_worker=1)
@override(Trainer)
def validate_config(self, config: TrainerConfigDict) -> None:
# Call super's validation method.
super().validate_config(config)
if config["entropy_coeff"] < 0:
raise ValueError("`entropy_coeff` must be >= 0.0!")
if config["num_workers"] <= 0 and config["sample_async"]:
raise ValueError("`num_workers` for A3C must be >= 1!")
@override(Trainer)
def get_default_policy_class(self,
config: TrainerConfigDict) -> Type[Policy]:
if config["framework"] == "torch":
from ray.rllib.algorithms.a3c.a3c_torch_policy import A3CTorchPolicy
return A3CTorchPolicy
elif config["framework"] == "tf":
from ray.rllib.algorithms.a3c.a3c_tf_policy import A3CStaticGraphTFPolicy
return A3CStaticGraphTFPolicy
else:
from ray.rllib.algorithms.a3c.a3c_tf_policy import A3CEagerTFPolicy
return A3CEagerTFPolicy
def training_step(self) -> ResultDict:
# Shortcut.
local_worker = self.workers.local_worker()
# Define the function executed in parallel by all RolloutWorkers to collect
# samples + compute and return gradients (and other information).
def sample_and_compute_grads(worker: RolloutWorker) -> Dict[str, Any]:
"""Call sample() and compute_gradients() remotely on workers."""
samples = worker.sample()
grads, infos = worker.compute_gradients(samples)
return {
"grads": grads,
"infos": infos,
"agent_steps": samples.agent_steps(),
"env_steps": samples.env_steps(),
}
# Perform rollouts and gradient calculations asynchronously.
with self._timers[GRAD_WAIT_TIMER]:
# Results are a mapping from ActorHandle (RolloutWorker) to their
# returned gradient calculation results.
self._worker_manager.call_on_all_available(
sample_and_compute_grads)
async_results = self._worker_manager.get_ready()
# Loop through all fetched worker-computed gradients (if any)
# and apply them - one by one - to the local worker's model.
# After each apply step (one step per worker that returned some gradients),
# update that particular worker's weights.
global_vars = None
learner_info_builder = LearnerInfoBuilder(num_devices=1)
for worker, results in async_results.items():
for result in results:
# Apply gradients to local worker.
with self._timers[APPLY_GRADS_TIMER]:
local_worker.apply_gradients(result["grads"])
self._timers[APPLY_GRADS_TIMER].push_units_processed(
result["agent_steps"])
# Update all step counters.
self._counters[NUM_AGENT_STEPS_SAMPLED] += result[
"agent_steps"]
self._counters[NUM_ENV_STEPS_SAMPLED] += result["env_steps"]
self._counters[NUM_AGENT_STEPS_TRAINED] += result[
"agent_steps"]
self._counters[NUM_ENV_STEPS_TRAINED] += result["env_steps"]
learner_info_builder.add_learn_on_batch_results_multi_agent(
result["infos"])
# Create current global vars.
global_vars = {
"timestep": self._counters[NUM_AGENT_STEPS_SAMPLED],
}
# Synch updated weights back to the particular worker.
with self._timers[SYNCH_WORKER_WEIGHTS_TIMER]:
weights = local_worker.get_weights(
local_worker.get_policies_to_train())
worker.set_weights.remote(weights, global_vars)
# Update global vars of the local worker.
if global_vars:
local_worker.set_global_vars(global_vars)
return learner_info_builder.finalize()
@override(Trainer)
def on_worker_failures(self, removed_workers: List[ActorHandle],
new_workers: List[ActorHandle]):
"""Handle failures on remote A3C workers.
Args:
removed_workers: removed worker ids.
new_workers: ids of newly created workers.
"""
self._worker_manager.remove_workers(removed_workers)
self._worker_manager.add_workers(new_workers)
class DDPPOTrainer(PPOTrainer):
def __init__(
self,
config: Optional[PartialTrainerConfigDict] = None,
env: Optional[Union[str, EnvType]] = None,
logger_creator: Optional[Callable[[], Logger]] = None,
remote_checkpoint_dir: Optional[str] = None,
sync_function_tpl: Optional[str] = None,
):
"""Initializes a DDPPOTrainer instance.
Args:
config: Algorithm-specific configuration dict.
env: Name of the environment to use (e.g. a gym-registered str),
a full class path (e.g.
"ray.rllib.examples.env.random_env.RandomEnv"), or an Env
class directly. Note that this arg can also be specified via
the "env" key in `config`.
logger_creator: Callable that creates a ray.tune.Logger
object. If unspecified, a default logger is created.
"""
super().__init__(config, env, logger_creator, remote_checkpoint_dir,
sync_function_tpl)
if "train_batch_size" in config.keys(
) and config["train_batch_size"] != -1:
# Users should not define `train_batch_size` directly (always -1).
raise ValueError(
"Set rollout_fragment_length instead of train_batch_size for DDPPO."
)
# Auto-train_batch_size: Calculate from rollout len and
# envs-per-worker.
config["train_batch_size"] = config.get(
"rollout_fragment_length",
DEFAULT_CONFIG["rollout_fragment_length"]) * config.get(
"num_envs_per_worker", DEFAULT_CONFIG["num_envs_per_worker"])
@classmethod
@override(PPOTrainer)
def get_default_config(cls) -> TrainerConfigDict:
return DDPPOConfig().to_dict()
@override(PPOTrainer)
def validate_config(self, config):
"""Validates the Trainer's config dict.
Args:
config (TrainerConfigDict): The Trainer's config to check.
Raises:
ValueError: In case something is wrong with the config.
"""
# Call (base) PPO's config validation function first.
# Note that this will not touch or check on the train_batch_size=-1
# setting.
super().validate_config(config)
# Must have `num_workers` >= 1.
if config["num_workers"] < 1:
raise ValueError("Due to its distributed, decentralized nature, "
"DD-PPO requires `num_workers` to be >= 1!")
# Only supported for PyTorch so far.
if config["framework"] != "torch":
raise ValueError(
"Distributed data parallel is only supported for PyTorch")
if config["torch_distributed_backend"] not in ("gloo", "mpi", "nccl"):
raise ValueError("Only gloo, mpi, or nccl is supported for "
"the backend of PyTorch distributed.")
# `num_gpus` must be 0/None, since all optimization happens on Workers.
if config["num_gpus"]:
raise ValueError(
"When using distributed data parallel, you should set "
"num_gpus=0 since all optimization "
"is happening on workers. Enable GPUs for workers by setting "
"num_gpus_per_worker=1.")
# `batch_mode` must be "truncate_episodes".
if config["batch_mode"] != "truncate_episodes":
raise ValueError(
"Distributed data parallel requires truncate_episodes batch mode."
)
# DDPPO doesn't support KL penalties like PPO-1.
# In order to support KL penalties, DDPPO would need to become
# undecentralized, which defeats the purpose of the algorithm.
# Users can still tune the entropy coefficient to control the
# policy entropy (similar to controlling the KL penalty).
if config["kl_coeff"] != 0.0 or config["kl_target"] != 0.0:
raise ValueError("DDPPO doesn't support KL penalties like PPO-1")
@override(PPOTrainer)
def setup(self, config: PartialTrainerConfigDict):
super().setup(config)
# Initialize torch process group for
if self.config["_disable_execution_plan_api"] is True:
self._curr_learner_info = {}
ip = ray.get(self.workers.remote_workers()[0].get_node_ip.remote())
port = ray.get(
self.workers.remote_workers()[0].find_free_port.remote())
address = "tcp://{ip}:{port}".format(ip=ip, port=port)
logger.info(
"Creating torch process group with leader {}".format(address))
# Get setup tasks in order to throw errors on failure.
ray.get([
worker.setup_torch_data_parallel.remote(
url=address,
world_rank=i,
world_size=len(self.workers.remote_workers()),
backend=self.config["torch_distributed_backend"],
) for i, worker in enumerate(self.workers.remote_workers())
])
logger.info("Torch process group init completed")
self._ddppo_worker_manager = AsyncRequestsManager(
self.workers.remote_workers(),
max_remote_requests_in_flight_per_worker=1,
ray_wait_timeout_s=0.03,
)
@override(PPOTrainer)
def training_iteration(self) -> ResultDict:
# Shortcut.
first_worker = self.workers.remote_workers()[0]
self._ddppo_worker_manager.call_on_all_available(
self._sample_and_train_torch_distributed)
sample_and_update_results = self._ddppo_worker_manager.get_ready()
# For all results collected:
# - Update our counters and timers.
# - Update the worker's global_vars.
# - Build info dict using a LearnerInfoBuilder object.
learner_info_builder = LearnerInfoBuilder(num_devices=1)
steps_this_iter = 0
for worker, results in sample_and_update_results.items():
for result in results:
steps_this_iter += result["env_steps"]
self._counters[NUM_AGENT_STEPS_SAMPLED] += result[
"agent_steps"]
self._counters[NUM_AGENT_STEPS_TRAINED] += result[
"agent_steps"]
self._counters[NUM_ENV_STEPS_SAMPLED] += result["env_steps"]
self._counters[NUM_ENV_STEPS_TRAINED] += result["env_steps"]
self._timers[LEARN_ON_BATCH_TIMER].push(
result["learn_on_batch_time"])
self._timers[SAMPLE_TIMER].push(result["sample_time"])
# Add partial learner info to builder object.
learner_info_builder.add_learn_on_batch_results_multi_agent(
result["info"])
# Broadcast the local set of global vars.
global_vars = {"timestep": self._counters[NUM_AGENT_STEPS_SAMPLED]}
for worker in self.workers.remote_workers():
worker.set_global_vars.remote(global_vars)
self._counters[STEPS_TRAINED_THIS_ITER_COUNTER] = steps_this_iter
# Sync down the weights from 1st remote worker (only if we have received
# some results from it).
# As with the sync up, this is not really needed unless the user is
# reading the local weights.
if (self.config["keep_local_weights_in_sync"]
and first_worker in sample_and_update_results):
self.workers.local_worker().set_weights(
ray.get(first_worker.get_weights.remote()))
# Return merged laarner into results.
new_learner_info = learner_info_builder.finalize()
if new_learner_info:
self._curr_learner_info = new_learner_info
return self._curr_learner_info
@staticmethod
def _sample_and_train_torch_distributed(worker: RolloutWorker):
# This function is applied remotely on each rollout worker.
config = worker.policy_config
# Generate a sample.
start = time.perf_counter()
batch = worker.sample()
sample_time = time.perf_counter() - start
expected_batch_size = (config["rollout_fragment_length"] *
config["num_envs_per_worker"])
assert batch.count == expected_batch_size, (
"Batch size possibly out of sync between workers, expected:",
expected_batch_size,
"got:",
batch.count,
)
# Perform n minibatch SGD update(s) on the worker itself.
start = time.perf_counter()
info = do_minibatch_sgd(
batch,
worker.policy_map,
worker,
config["num_sgd_iter"],
config["sgd_minibatch_size"],
[Postprocessing.ADVANTAGES],
)
learn_on_batch_time = time.perf_counter() - start
return {
"info": info,
"env_steps": batch.env_steps(),
"agent_steps": batch.agent_steps(),
"sample_time": sample_time,
"learn_on_batch_time": learn_on_batch_time,
}
class AlphaStar(appo.APPO):
_allow_unknown_subkeys = appo.APPO._allow_unknown_subkeys + [
"league_builder_config",
]
_override_all_subkeys_if_type_changes = (
appo.APPO._override_all_subkeys_if_type_changes + [
"league_builder_config",
])
@classmethod
@override(Algorithm)
def default_resource_request(cls, config):
cf = dict(cls.get_default_config(), **config)
# Construct a dummy LeagueBuilder, such that it gets the opportunity to
# adjust the multiagent config, according to its setup, and we can then
# properly infer the resources to allocate.
from_config(cf["league_builder_config"],
trainer=None,
trainer_config=cf)
max_num_policies_to_train = cf["max_num_policies_to_train"] or len(
cf["multiagent"].get("policies_to_train")
or cf["multiagent"]["policies"])
num_learner_shards = min(cf["num_gpus"] or max_num_policies_to_train,
max_num_policies_to_train)
num_gpus_per_shard = cf["num_gpus"] / num_learner_shards
num_policies_per_shard = max_num_policies_to_train / num_learner_shards
fake_gpus = cf["_fake_gpus"]
eval_config = cf["evaluation_config"]
# Return PlacementGroupFactory containing all needed resources
# (already properly defined as device bundles).
return PlacementGroupFactory(
bundles=[{
# Driver (no GPUs).
"CPU": cf["num_cpus_for_driver"],
}] + [
{
# RolloutWorkers (no GPUs).
"CPU": cf["num_cpus_per_worker"],
} for _ in range(cf["num_workers"])
] + [
{
# Policy learners (and Replay buffer shards).
# 1 CPU for the replay buffer.
# 1 CPU (or fractional GPU) for each learning policy.
"CPU": 1 + (num_policies_per_shard if fake_gpus else 0),
"GPU": 0 if fake_gpus else num_gpus_per_shard,
} for _ in range(num_learner_shards)
] + ([
{
# Evaluation (remote) workers.
# Note: The local eval worker is located on the driver
# CPU or not even created iff >0 eval workers.
"CPU":
eval_config.get("num_cpus_per_worker",
cf["num_cpus_per_worker"]),
} for _ in range(cf["evaluation_num_workers"])
] if cf["evaluation_interval"] else []),
strategy=config.get("placement_strategy", "PACK"),
)
@classmethod
@override(appo.APPO)
def get_default_config(cls) -> AlgorithmConfigDict:
return AlphaStarConfig().to_dict()
@override(appo.APPO)
def validate_config(self, config: AlgorithmConfigDict):
# Create the LeagueBuilder object, allowing it to build the multiagent
# config as well.
self.league_builder = from_config(config["league_builder_config"],
trainer=self,
trainer_config=config)
super().validate_config(config)
@override(appo.APPO)
def setup(self, config: PartialAlgorithmConfigDict):
# Call super's setup to validate config, create RolloutWorkers
# (train and eval), etc..
num_gpus_saved = config["num_gpus"]
config["num_gpus"] = min(config["num_gpus"], 1)
super().setup(config)
self.config["num_gpus"] = num_gpus_saved
# - Create n policy learner actors (@ray.remote-converted Policies) on
# one or more GPU nodes.
# - On each such node, also locate one replay buffer shard.
ma_cfg = self.config["multiagent"]
# By default, set max_num_policies_to_train to the number of policy IDs
# provided in the multiagent config.
if self.config["max_num_policies_to_train"] is None:
self.config["max_num_policies_to_train"] = len(
self.workers.local_worker().get_policies_to_train())
# Single CPU replay shard (co-located with GPUs so we can place the
# policies on the same machine(s)).
num_gpus = (0.01 if (self.config["num_gpus"]
and not self.config["_fake_gpus"]) else 0)
ReplayActor = ray.remote(
num_cpus=1,
num_gpus=num_gpus,
)(MixInMultiAgentReplayBuffer)
# Setup remote replay buffer shards and policy learner actors
# (located on any GPU machine in the cluster):
replay_actor_args = [
self.config["replay_buffer_capacity"],
self.config["replay_buffer_replay_ratio"],
]
# Create a DistributedLearners utility object and set it up with
# the initial first n learnable policies (found in the config).
distributed_learners = DistributedLearners(
config=self.config,
max_num_policies_to_train=self.config["max_num_policies_to_train"],
replay_actor_class=ReplayActor,
replay_actor_args=replay_actor_args,
)
for pid, policy_spec in ma_cfg["policies"].items():
if pid in self.workers.local_worker().get_policies_to_train():
distributed_learners.add_policy(pid, policy_spec)
# Store distributed_learners on all RolloutWorkers
# so they know, to which replay shard to send samples to.
def _set_policy_learners(worker):
worker._distributed_learners = distributed_learners
ray.get([
w.apply.remote(_set_policy_learners)
for w in self.workers.remote_workers()
])
self.distributed_learners = distributed_learners
self._sampling_actor_manager = AsyncRequestsManager(
self.workers.remote_workers(),
max_remote_requests_in_flight_per_worker=self.
config["max_requests_in_flight_per_sampler_worker"],
ray_wait_timeout_s=self.config["timeout_s_sampler_manager"],
)
policy_actors = [
policy_actor for _, policy_actor, _ in distributed_learners
]
self._learner_worker_manager = AsyncRequestsManager(
workers=policy_actors,
max_remote_requests_in_flight_per_worker=self.
config["max_requests_in_flight_per_learner_worker"],
ray_wait_timeout_s=self.config["timeout_s_learner_manager"],
)
@override(Algorithm)
def step(self) -> ResultDict:
# Perform a full step (including evaluation).
result = super().step()
# Based on the (train + evaluate) results, perform a step of
# league building.
self.league_builder.build_league(result=result)
return result
@override(Algorithm)
def training_step(self) -> ResultDict:
# Trigger asynchronous rollouts on all RolloutWorkers.
# - Rollout results are sent directly to correct replay buffer
# shards, instead of here (to the driver).
with self._timers[SAMPLE_TIMER]:
# if there are no remote workers (e.g. num_workers=0)
if not self.workers.remote_workers():
worker = self.workers.local_worker()
statistics = worker.apply(self._sample_and_send_to_buffer)
sample_results = {worker: [statistics]}
else:
self._sampling_actor_manager.call_on_all_available(
self._sample_and_send_to_buffer)
sample_results = self._sampling_actor_manager.get_ready()
# Update sample counters.
for sample_result in sample_results.values():
for (env_steps, agent_steps) in sample_result:
self._counters[NUM_ENV_STEPS_SAMPLED] += env_steps
self._counters[NUM_AGENT_STEPS_SAMPLED] += agent_steps
# Trigger asynchronous training update requests on all learning
# policies.
with self._timers[LEARN_ON_BATCH_TIMER]:
for pid, pol_actor, repl_actor in self.distributed_learners:
if pol_actor not in self._learner_worker_manager.workers:
self._learner_worker_manager.add_workers(pol_actor)
self._learner_worker_manager.call(self._update_policy,
actor=pol_actor,
fn_args=[repl_actor, pid])
train_results = self._learner_worker_manager.get_ready()
# Update sample counters.
for train_result in train_results.values():
for result in train_result:
if NUM_AGENT_STEPS_TRAINED in result:
self._counters[NUM_AGENT_STEPS_TRAINED] += result[
NUM_AGENT_STEPS_TRAINED]
# For those policies that have been updated in this iteration
# (not all policies may have undergone an updated as we are
# requesting updates asynchronously):
# - Gather train infos.
# - Update weights to those remote rollout workers that contain
# the respective policy.
with self._timers[SYNCH_WORKER_WEIGHTS_TIMER]:
train_infos = {}
policy_weights = {}
for pol_actor, policy_results in train_results.items():
results_have_same_structure = True
for result1, result2 in zip(policy_results,
policy_results[1:]):
try:
tree.assert_same_structure(result1, result2)
except (ValueError, TypeError):
results_have_same_structure = False
break
if len(policy_results) > 1 and results_have_same_structure:
policy_result = tree.map_structure(
lambda *_args: sum(_args) / len(policy_results),
*policy_results)
else:
policy_result = policy_results[-1]
if policy_result:
pid = self.distributed_learners.get_policy_id(pol_actor)
train_infos[pid] = policy_result
policy_weights[pid] = pol_actor.get_weights.remote()
policy_weights_ref = ray.put(policy_weights)
global_vars = {
"timestep": self._counters[NUM_ENV_STEPS_SAMPLED],
"league_builder": self.league_builder.__getstate__(),
}
for worker in self.workers.remote_workers():
worker.set_weights.remote(policy_weights_ref, global_vars)
return train_infos
@override(Algorithm)
def add_policy(
self,
policy_id: PolicyID,
policy_cls: Type[Policy],
*,
observation_space: Optional[gym.spaces.Space] = None,
action_space: Optional[gym.spaces.Space] = None,
config: Optional[PartialAlgorithmConfigDict] = None,
policy_state: Optional[PolicyState] = None,
**kwargs,
) -> Policy:
# Add the new policy to all our train- and eval RolloutWorkers
# (including the local worker).
new_policy = super().add_policy(
policy_id,
policy_cls,
observation_space=observation_space,
action_space=action_space,
config=config,
policy_state=policy_state,
**kwargs,
)
# Do we have to create a policy-learner actor from it as well?
if policy_id in kwargs.get("policies_to_train", []):
new_policy_actor = self.distributed_learners.add_policy(
policy_id,
PolicySpec(
policy_cls,
new_policy.observation_space,
new_policy.action_space,
self.config,
),
)
# Set state of new policy actor, if provided.
if policy_state is not None:
ray.get(new_policy_actor.set_state.remote(policy_state))
return new_policy
@override(Algorithm)
def cleanup(self) -> None:
super().cleanup()
# Stop all policy- and replay actors.
self.distributed_learners.stop()
@staticmethod
def _sample_and_send_to_buffer(worker: RolloutWorker):
# Generate a sample.
sample = worker.sample()
# Send the per-agent SampleBatches to the correct buffer(s),
# depending on which policies participated in the episode.
assert isinstance(sample, MultiAgentBatch)
for pid, batch in sample.policy_batches.items():
# Don't send data, if policy is not trainable.
replay_actor, _ = worker._distributed_learners.get_replay_and_policy_actors(
pid)
if replay_actor is not None:
ma_batch = MultiAgentBatch({pid: batch}, batch.count)
replay_actor.add_batch.remote(ma_batch)
# Return counts (env-steps, agent-steps).
return sample.count, sample.agent_steps()
@staticmethod
def _update_policy(policy: Policy, replay_actor: ActorHandle,
pid: PolicyID):
if not hasattr(policy, "_target_and_kl_stats"):
policy._target_and_kl_stats = {
LAST_TARGET_UPDATE_TS: 0,
NUM_TARGET_UPDATES: 0,
NUM_AGENT_STEPS_TRAINED: 0,
TARGET_NET_UPDATE_TIMER: _Timer(),
}
train_results = policy.learn_on_batch_from_replay_buffer(
replay_actor=replay_actor, policy_id=pid)
if not train_results:
return train_results
# Update target net and KL.
with policy._target_and_kl_stats[TARGET_NET_UPDATE_TIMER]:
policy._target_and_kl_stats[
NUM_AGENT_STEPS_TRAINED] += train_results[
NUM_AGENT_STEPS_TRAINED]
target_update_freq = (policy.config["num_sgd_iter"] *
policy.config["replay_buffer_capacity"] *
policy.config["train_batch_size"])
cur_ts = policy._target_and_kl_stats[NUM_AGENT_STEPS_TRAINED]
last_update = policy._target_and_kl_stats[LAST_TARGET_UPDATE_TS]
# Update target networks on all policy learners.
if cur_ts - last_update > target_update_freq:
policy._target_and_kl_stats[NUM_TARGET_UPDATES] += 1
policy._target_and_kl_stats[LAST_TARGET_UPDATE_TS] = cur_ts
policy.update_target()
# Also update Policy's current KL coeff.
if policy.config["use_kl_loss"]:
kl = train_results[LEARNER_STATS_KEY].get("kl")
assert kl is not None, train_results
# Make the actual `Policy.update_kl()` call.
policy.update_kl(kl)
return train_results
@override(appo.APPO)
def __getstate__(self) -> dict:
state = super().__getstate__()
state.update({
"league_builder": self.league_builder.__getstate__(),
})
return state
@override(appo.APPO)
def __setstate__(self, state: dict) -> None:
state_copy = state.copy()
self.league_builder.__setstate__(state.pop("league_builder", {}))
super().__setstate__(state_copy)
def test_add_remove_actors(self):
"""Tests that the async manager can properly add and remove actors"""
workers = []
manager = AsyncRequestsManager(
workers, max_remote_requests_in_flight_per_worker=2
)
if not (
(
len(manager._all_workers)
== len(manager._remote_requests_in_flight)
== len(manager._pending_to_actor)
== len(manager._pending_remotes)
== 0
)
):
raise ValueError("We should have no workers in this case.")
assert not manager.call(lambda w: w.task()), (
"Task shouldn't have been "
"launched since there are no "
"workers in the manager."
)
worker = RemoteRLlibActor.remote(sleep_time=0.1)
manager.add_workers(worker)
manager.call(lambda w: w.task())
if not (
len(manager._remote_requests_in_flight[worker])
== len(manager._pending_to_actor)
== len(manager._all_workers)
== len(manager._pending_remotes)
== 1
):
raise ValueError("We should have 1 worker and 1 pending request")
time.sleep(3)
manager.get_ready()
# test worker removal
for i in range(2):
manager.call(lambda w: w.task())
assert len(manager._pending_remotes) == i + 1
manager.remove_workers(worker)
if not ((len(manager._all_workers) == 0)):
raise ValueError("We should have no workers that we can schedule tasks to")
if not (
(len(manager._pending_remotes) == 2 and len(manager._pending_to_actor) == 2)
):
raise ValueError(
"We should still have 2 pending requests in flight from the worker"
)
time.sleep(3)
result = manager.get_ready()
if not (
len(result) == 1
and len(result[worker]) == 2
and len(manager._pending_remotes) == 0
and len(manager._pending_to_actor) == 0
):
raise ValueError(
"We should have 2 ready results from the worker and no pending requests"
)
class Impala(Algorithm):
"""Importance weighted actor/learner architecture (IMPALA) Algorithm
== Overview of data flow in IMPALA ==
1. Policy evaluation in parallel across `num_workers` actors produces
batches of size `rollout_fragment_length * num_envs_per_worker`.
2. If enabled, the replay buffer stores and produces batches of size
`rollout_fragment_length * num_envs_per_worker`.
3. If enabled, the minibatch ring buffer stores and replays batches of
size `train_batch_size` up to `num_sgd_iter` times per batch.
4. The learner thread executes data parallel SGD across `num_gpus` GPUs
on batches of size `train_batch_size`.
"""
@classmethod
@override(Algorithm)
def get_default_config(cls) -> AlgorithmConfigDict:
return ImpalaConfig().to_dict()
@override(Algorithm)
def get_default_policy_class(
self,
config: PartialAlgorithmConfigDict) -> Optional[Type[Policy]]:
if config["framework"] == "torch":
if config["vtrace"]:
from ray.rllib.algorithms.impala.impala_torch_policy import (
ImpalaTorchPolicy, )
return ImpalaTorchPolicy
else:
from ray.rllib.algorithms.a3c.a3c_torch_policy import A3CTorchPolicy
return A3CTorchPolicy
elif config["framework"] == "tf":
if config["vtrace"]:
from ray.rllib.algorithms.impala.impala_tf_policy import ImpalaTF1Policy
return ImpalaTF1Policy
else:
from ray.rllib.algorithms.a3c.a3c_tf_policy import A3CTFPolicy
return A3CTFPolicy
else:
if config["vtrace"]:
from ray.rllib.algorithms.impala.impala_tf_policy import ImpalaTF2Policy
return ImpalaTF2Policy
else:
from ray.rllib.algorithms.a3c.a3c_tf_policy import A3CTFPolicy
return A3CTFPolicy
@override(Algorithm)
def validate_config(self, config):
# Call the super class' validation method first.
super().validate_config(config)
# Check the IMPALA specific config.
if config["num_data_loader_buffers"] != DEPRECATED_VALUE:
deprecation_warning("num_data_loader_buffers",
"num_multi_gpu_tower_stacks",
error=False)
config["num_multi_gpu_tower_stacks"] = config[
"num_data_loader_buffers"]
if config["entropy_coeff"] < 0.0:
raise ValueError("`entropy_coeff` must be >= 0.0!")
# Check whether worker to aggregation-worker ratio makes sense.
if config["num_aggregation_workers"] > config["num_workers"]:
raise ValueError(
"`num_aggregation_workers` must be smaller than or equal "
"`num_workers`! Aggregation makes no sense otherwise.")
elif config["num_aggregation_workers"] > config["num_workers"] / 2:
logger.warning(
"`num_aggregation_workers` should be significantly smaller "
"than `num_workers`! Try setting it to 0.5*`num_workers` or "
"less.")
# If two separate optimizers/loss terms used for tf, must also set
# `_tf_policy_handles_more_than_one_loss` to True.
if config["_separate_vf_optimizer"] is True:
# Only supported to tf so far.
# TODO(sven): Need to change APPO|IMPALATorchPolicies (and the
# models to return separate sets of weights in order to create
# the different torch optimizers).
if config["framework"] not in ["tf", "tf2", "tfe"]:
raise ValueError(
"`_separate_vf_optimizer` only supported to tf so far!")
if config["_tf_policy_handles_more_than_one_loss"] is False:
logger.warning(
"`_tf_policy_handles_more_than_one_loss` must be set to "
"True, for TFPolicy to support more than one loss "
"term/optimizer! Auto-setting it to True.")
config["_tf_policy_handles_more_than_one_loss"] = True
@override(Algorithm)
def setup(self, config: PartialAlgorithmConfigDict):
super().setup(config)
if self.config["_disable_execution_plan_api"]:
# Create extra aggregation workers and assign each rollout worker to
# one of them.
self.batches_to_place_on_learner = []
self.batch_being_built = []
if self.config["num_aggregation_workers"] > 0:
# This spawns `num_aggregation_workers` actors that aggregate
# experiences coming from RolloutWorkers in parallel. We force
# colocation on the same node (localhost) to maximize data bandwidth
# between them and the learner.
localhost = platform.node()
assert localhost != "", (
"ERROR: Cannot determine local node name! "
"`platform.node()` returned empty string.")
all_co_located = create_colocated_actors(
actor_specs=[
# (class, args, kwargs={}, count=1)
(
AggregatorWorker,
[
self.config,
],
{},
self.config["num_aggregation_workers"],
)
],
node=localhost,
)
self._aggregator_workers = [
actor for actor_groups in all_co_located
for actor in actor_groups
]
self._aggregator_actor_manager = AsyncRequestsManager(
self._aggregator_workers,
max_remote_requests_in_flight_per_worker=self.
config["max_requests_in_flight_per_aggregator_worker"],
ray_wait_timeout_s=self.
config["timeout_s_aggregator_manager"],
)
else:
# Create our local mixin buffer if the num of aggregation workers is 0.
self.local_mixin_buffer = MixInMultiAgentReplayBuffer(
capacity=(self.config["replay_buffer_num_slots"]
if self.config["replay_buffer_num_slots"] > 0
else 1),
replay_ratio=self.config["replay_ratio"],
replay_mode=ReplayMode.LOCKSTEP,
)
self._sampling_actor_manager = AsyncRequestsManager(
self.workers.remote_workers(),
max_remote_requests_in_flight_per_worker=self.
config["max_requests_in_flight_per_sampler_worker"],
return_object_refs=True,
ray_wait_timeout_s=self.config["timeout_s_sampler_manager"],
)
# Create and start the learner thread.
self._learner_thread = make_learner_thread(
self.workers.local_worker(), self.config)
self._learner_thread.start()
self.workers_that_need_updates = set()
@override(Algorithm)
def training_step(self) -> ResultDict:
unprocessed_sample_batches = self.get_samples_from_workers()
self.workers_that_need_updates |= unprocessed_sample_batches.keys()
if self.config["num_aggregation_workers"] > 0:
batch = self.process_experiences_tree_aggregation(
unprocessed_sample_batches)
else:
batch = self.process_experiences_directly(
unprocessed_sample_batches)
self.concatenate_batches_and_pre_queue(batch)
self.place_processed_samples_on_learner_queue()
train_results = self.process_trained_results()
self.update_workers_if_necessary()
return train_results
@staticmethod
@override(Algorithm)
def execution_plan(workers, config, **kwargs):
assert (
len(kwargs) == 0
), "IMPALA execution_plan does NOT take any additional parameters"
if config["num_aggregation_workers"] > 0:
train_batches = gather_experiences_tree_aggregation(
workers, config)
else:
train_batches = gather_experiences_directly(workers, config)
# Start the learner thread.
learner_thread = make_learner_thread(workers.local_worker(), config)
learner_thread.start()
# This sub-flow sends experiences to the learner.
enqueue_op = train_batches.for_each(Enqueue(learner_thread.inqueue))
# Only need to update workers if there are remote workers.
if workers.remote_workers():
enqueue_op = enqueue_op.zip_with_source_actor().for_each(
BroadcastUpdateLearnerWeights(
learner_thread,
workers,
broadcast_interval=config["broadcast_interval"],
))
def record_steps_trained(item):
count, fetches, _ = item
metrics = _get_shared_metrics()
# Manually update the steps trained counter since the learner
# thread is executing outside the pipeline.
metrics.counters[STEPS_TRAINED_THIS_ITER_COUNTER] = count
metrics.counters[STEPS_TRAINED_COUNTER] += count
return item
# This sub-flow updates the steps trained counter based on learner
# output.
dequeue_op = Dequeue(
learner_thread.outqueue,
check=learner_thread.is_alive).for_each(record_steps_trained)
merged_op = Concurrently([enqueue_op, dequeue_op],
mode="async",
output_indexes=[1])
# Callback for APPO to use to update KL, target network periodically.
# The input to the callback is the learner fetches dict.
if config["after_train_step"]:
merged_op = merged_op.for_each(lambda t: t[1]).for_each(
config["after_train_step"](workers, config))
return StandardMetricsReporting(merged_op, workers, config).for_each(
learner_thread.add_learner_metrics)
@classmethod
@override(Algorithm)
def default_resource_request(cls, config):
cf = dict(cls.get_default_config(), **config)
eval_config = cf["evaluation_config"]
# Return PlacementGroupFactory containing all needed resources
# (already properly defined as device bundles).
return PlacementGroupFactory(
bundles=[{
# Driver + Aggregation Workers:
# Force to be on same node to maximize data bandwidth
# between aggregation workers and the learner (driver).
# Aggregation workers tree-aggregate experiences collected
# from RolloutWorkers (n rollout workers map to m
# aggregation workers, where m < n) and always use 1 CPU
# each.
"CPU":
cf["num_cpus_for_driver"] + cf["num_aggregation_workers"],
"GPU":
0 if cf["_fake_gpus"] else cf["num_gpus"],
}] + [
{
# RolloutWorkers.
"CPU": cf["num_cpus_per_worker"],
"GPU": cf["num_gpus_per_worker"],
**cf["custom_resources_per_worker"],
} for _ in range(cf["num_workers"])
] + ([
{
# Evaluation (remote) workers.
# Note: The local eval worker is located on the driver
# CPU or not even created iff >0 eval workers.
"CPU":
eval_config.get("num_cpus_per_worker",
cf["num_cpus_per_worker"]),
"GPU":
eval_config.get("num_gpus_per_worker",
cf["num_gpus_per_worker"]),
**eval_config.get(
"custom_resources_per_worker",
cf["custom_resources_per_worker"],
),
} for _ in range(cf["evaluation_num_workers"])
] if cf["evaluation_interval"] else []),
strategy=config.get("placement_strategy", "PACK"),
)
def concatenate_batches_and_pre_queue(self, batches: List[SampleBatch]):
"""Concatenate batches that are being returned from rollout workers
Args:
batches: batches of experiences from rollout workers
"""
def aggregate_into_larger_batch():
if (sum(b.count for b in self.batch_being_built) >=
self.config["train_batch_size"]):
batch_to_add = SampleBatch.concat_samples(
self.batch_being_built)
self.batches_to_place_on_learner.append(batch_to_add)
self.batch_being_built = []
for batch in batches:
self.batch_being_built.append(batch)
aggregate_into_larger_batch()
def get_samples_from_workers(self) -> Dict[ActorHandle, List[SampleBatch]]:
# Perform asynchronous sampling on all (remote) rollout workers.
if self.workers.remote_workers():
self._sampling_actor_manager.call_on_all_available(
lambda worker: worker.sample())
sample_batches: Dict[
ActorHandle,
List[ObjectRef]] = self._sampling_actor_manager.get_ready()
else:
# only sampling on the local worker
sample_batches = {
self.workers.local_worker():
[self.workers.local_worker().sample()]
}
return sample_batches
def place_processed_samples_on_learner_queue(self) -> None:
self._counters["num_samples_added_to_queue"] = 0
while self.batches_to_place_on_learner:
batch = self.batches_to_place_on_learner[0]
try:
self._learner_thread.inqueue.put(batch, block=False)
self.batches_to_place_on_learner.pop(0)
self._counters[NUM_ENV_STEPS_SAMPLED] += batch.count
self._counters[NUM_AGENT_STEPS_SAMPLED] += batch.agent_steps()
self._counters["num_samples_added_to_queue"] = batch.count
except queue.Full:
self._counters["num_times_learner_queue_full"] += 1
def process_trained_results(self) -> ResultDict:
# Get learner outputs/stats from output queue.
final_learner_info = {}
learner_infos = []
num_env_steps_trained = 0
num_agent_steps_trained = 0
for _ in range(self._learner_thread.outqueue.qsize()):
if self._learner_thread.is_alive():
(
env_steps,
agent_steps,
learner_results,
) = self._learner_thread.outqueue.get(timeout=0.001)
num_env_steps_trained += env_steps
num_agent_steps_trained += agent_steps
if learner_results:
learner_infos.append(learner_results)
else:
raise RuntimeError("The learner thread died in while training")
if not learner_infos:
final_learner_info = copy.deepcopy(
self._learner_thread.learner_info)
else:
builder = LearnerInfoBuilder()
for info in learner_infos:
builder.add_learn_on_batch_results_multi_agent(info)
final_learner_info = builder.finalize()
# Update the steps trained counters.
self._counters[
STEPS_TRAINED_THIS_ITER_COUNTER] = num_agent_steps_trained
self._counters[NUM_ENV_STEPS_TRAINED] += num_env_steps_trained
self._counters[NUM_AGENT_STEPS_TRAINED] += num_agent_steps_trained
return final_learner_info
def process_experiences_directly(
self, actor_to_sample_batches_refs: Dict[ActorHandle, List[ObjectRef]]
) -> Union[SampleBatchType, None]:
processed_batches = []
batches = [
sample_batch_ref
for refs_batch in actor_to_sample_batches_refs.values()
for sample_batch_ref in refs_batch
]
if not batches:
return processed_batches
if batches and isinstance(batches[0], ray.ObjectRef):
batches = ray.get(batches)
for batch in batches:
batch = batch.decompress_if_needed()
self.local_mixin_buffer.add_batch(batch)
batch = self.local_mixin_buffer.replay(_ALL_POLICIES)
if batch:
processed_batches.append(batch)
return processed_batches
def process_experiences_tree_aggregation(
self, actor_to_sample_batches_refs: Dict[ActorHandle, List[ObjectRef]]
) -> Union[SampleBatchType, None]:
batches = [
sample_batch_ref
for refs_batch in actor_to_sample_batches_refs.values()
for sample_batch_ref in refs_batch
]
ready_processed_batches = []
for batch in batches:
self._aggregator_actor_manager.call(
lambda actor, b: actor.process_episodes(b),
fn_kwargs={"b": batch})
waiting_processed_sample_batches: Dict[
ActorHandle,
List[ObjectRef]] = self._aggregator_actor_manager.get_ready()
for ready_sub_batches in waiting_processed_sample_batches.values():
ready_processed_batches.extend(ready_sub_batches)
return ready_processed_batches
def update_workers_if_necessary(self) -> None:
# Only need to update workers if there are remote workers.
global_vars = {"timestep": self._counters[NUM_AGENT_STEPS_TRAINED]}
self._counters["steps_since_broadcast"] += 1
if (self.workers.remote_workers()
and self._counters["steps_since_broadcast"] >=
self.config["broadcast_interval"]
and self.workers_that_need_updates):
weights = ray.put(self.workers.local_worker().get_weights())
self._counters["steps_since_broadcast"] = 0
self._learner_thread.weights_updated = False
self._counters["num_weight_broadcasts"] += 1
for worker in self.workers_that_need_updates:
worker.set_weights.remote(weights, global_vars)
self.workers_that_need_updates = set()
# Update global vars of the local worker.
self.workers.local_worker().set_global_vars(global_vars)
@override(Algorithm)
def on_worker_failures(self, removed_workers: List[ActorHandle],
new_workers: List[ActorHandle]):
"""Handle the failures of remote sampling workers
Args:
removed_workers: removed worker ids.
new_workers: ids of newly created workers.
"""
if self.config["_disable_execution_plan_api"]:
self._sampling_actor_manager.remove_workers(
removed_workers, remove_in_flight_requests=True)
self._sampling_actor_manager.add_workers(new_workers)
@override(Algorithm)
def _compile_iteration_results(self, *, step_ctx, iteration_results=None):
result = super()._compile_iteration_results(
step_ctx=step_ctx, iteration_results=iteration_results)
result = self._learner_thread.add_learner_metrics(
result, overwrite_learner_info=False)
return result
def setup(self, config: PartialAlgorithmConfigDict):
super().setup(config)
if self.config["_disable_execution_plan_api"]:
# Create extra aggregation workers and assign each rollout worker to
# one of them.
self.batches_to_place_on_learner = []
self.batch_being_built = []
if self.config["num_aggregation_workers"] > 0:
# This spawns `num_aggregation_workers` actors that aggregate
# experiences coming from RolloutWorkers in parallel. We force
# colocation on the same node (localhost) to maximize data bandwidth
# between them and the learner.
localhost = platform.node()
assert localhost != "", (
"ERROR: Cannot determine local node name! "
"`platform.node()` returned empty string.")
all_co_located = create_colocated_actors(
actor_specs=[
# (class, args, kwargs={}, count=1)
(
AggregatorWorker,
[
self.config,
],
{},
self.config["num_aggregation_workers"],
)
],
node=localhost,
)
self._aggregator_workers = [
actor for actor_groups in all_co_located
for actor in actor_groups
]
self._aggregator_actor_manager = AsyncRequestsManager(
self._aggregator_workers,
max_remote_requests_in_flight_per_worker=self.
config["max_requests_in_flight_per_aggregator_worker"],
ray_wait_timeout_s=self.
config["timeout_s_aggregator_manager"],
)
else:
# Create our local mixin buffer if the num of aggregation workers is 0.
self.local_mixin_buffer = MixInMultiAgentReplayBuffer(
capacity=(self.config["replay_buffer_num_slots"]
if self.config["replay_buffer_num_slots"] > 0
else 1),
replay_ratio=self.config["replay_ratio"],
replay_mode=ReplayMode.LOCKSTEP,
)
self._sampling_actor_manager = AsyncRequestsManager(
self.workers.remote_workers(),
max_remote_requests_in_flight_per_worker=self.
config["max_requests_in_flight_per_sampler_worker"],
return_object_refs=True,
ray_wait_timeout_s=self.config["timeout_s_sampler_manager"],
)
# Create and start the learner thread.
self._learner_thread = make_learner_thread(
self.workers.local_worker(), self.config)
self._learner_thread.start()
self.workers_that_need_updates = set()
def setup(self, config: PartialAlgorithmConfigDict):
# Call super's setup to validate config, create RolloutWorkers
# (train and eval), etc..
num_gpus_saved = config["num_gpus"]
config["num_gpus"] = min(config["num_gpus"], 1)
super().setup(config)
self.config["num_gpus"] = num_gpus_saved
# - Create n policy learner actors (@ray.remote-converted Policies) on
# one or more GPU nodes.
# - On each such node, also locate one replay buffer shard.
ma_cfg = self.config["multiagent"]
# By default, set max_num_policies_to_train to the number of policy IDs
# provided in the multiagent config.
if self.config["max_num_policies_to_train"] is None:
self.config["max_num_policies_to_train"] = len(
self.workers.local_worker().get_policies_to_train())
# Single CPU replay shard (co-located with GPUs so we can place the
# policies on the same machine(s)).
num_gpus = (0.01 if (self.config["num_gpus"]
and not self.config["_fake_gpus"]) else 0)
ReplayActor = ray.remote(
num_cpus=1,
num_gpus=num_gpus,
)(MixInMultiAgentReplayBuffer)
# Setup remote replay buffer shards and policy learner actors
# (located on any GPU machine in the cluster):
replay_actor_args = [
self.config["replay_buffer_capacity"],
self.config["replay_buffer_replay_ratio"],
]
# Create a DistributedLearners utility object and set it up with
# the initial first n learnable policies (found in the config).
distributed_learners = DistributedLearners(
config=self.config,
max_num_policies_to_train=self.config["max_num_policies_to_train"],
replay_actor_class=ReplayActor,
replay_actor_args=replay_actor_args,
)
for pid, policy_spec in ma_cfg["policies"].items():
if pid in self.workers.local_worker().get_policies_to_train():
distributed_learners.add_policy(pid, policy_spec)
# Store distributed_learners on all RolloutWorkers
# so they know, to which replay shard to send samples to.
def _set_policy_learners(worker):
worker._distributed_learners = distributed_learners
ray.get([
w.apply.remote(_set_policy_learners)
for w in self.workers.remote_workers()
])
self.distributed_learners = distributed_learners
self._sampling_actor_manager = AsyncRequestsManager(
self.workers.remote_workers(),
max_remote_requests_in_flight_per_worker=self.
config["max_requests_in_flight_per_sampler_worker"],
ray_wait_timeout_s=self.config["timeout_s_sampler_manager"],
)
policy_actors = [
policy_actor for _, policy_actor, _ in distributed_learners
]
self._learner_worker_manager = AsyncRequestsManager(
workers=policy_actors,
max_remote_requests_in_flight_per_worker=self.
config["max_requests_in_flight_per_learner_worker"],
ray_wait_timeout_s=self.config["timeout_s_learner_manager"],
)