def step(self) -> Optional[_NextValueNotReady]:
with self.queue_timer:
try:
batch, _ = self.minibatch_buffer.get()
except queue.Empty:
return _NextValueNotReady()
with self.grad_timer:
# Use LearnerInfoBuilder as a unified way to build the final
# results dict from `learn_on_loaded_batch` call(s).
# This makes sure results dicts always have the same structure
# no matter the setup (multi-GPU, multi-agent, minibatch SGD,
# tf vs torch).
learner_info_builder = LearnerInfoBuilder(num_devices=1)
multi_agent_results = self.local_worker.learn_on_batch(batch)
for pid, results in multi_agent_results.items():
learner_info_builder.add_learn_on_batch_results(results, pid)
self.learner_info = learner_info_builder.finalize()
self.weights_updated = True
self.num_steps += 1
# Put tuple: env-steps, agent-steps, and learner info into the queue.
self.outqueue.put(
(batch.count, batch.agent_steps(), self.learner_info))
self.learner_queue_size.push(self.inqueue.qsize())
def step(self):
with self.overall_timer:
with self.queue_timer:
replay_actor, ma_batch = self.inqueue.get()
if ma_batch is not None:
prio_dict = {}
with self.grad_timer:
# Use LearnerInfoBuilder as a unified way to build the
# final results dict from `learn_on_loaded_batch` call(s).
# This makes sure results dicts always have the same
# structure no matter the setup (multi-GPU, multi-agent,
# minibatch SGD, tf vs torch).
learner_info_builder = LearnerInfoBuilder(num_devices=1)
multi_agent_results = self.local_worker.learn_on_batch(
ma_batch)
for pid, results in multi_agent_results.items():
learner_info_builder.add_learn_on_batch_results(
results, pid)
td_error = results["td_error"]
# Switch off auto-conversion from numpy to torch/tf
# tensors for the indices. This may lead to errors
# when sent to the buffer for processing
# (may get manipulated if they are part of a tensor).
ma_batch.policy_batches[pid].set_get_interceptor(None)
prio_dict[pid] = (
ma_batch.policy_batches[pid].get("batch_indexes"),
td_error,
)
self.learner_info = learner_info_builder.finalize()
self.grad_timer.push_units_processed(ma_batch.count)
self.outqueue.put((replay_actor, prio_dict, ma_batch.count,
ma_batch.agent_steps()))
self.learner_queue_size.push(self.inqueue.qsize())
self.weights_updated = True
self.overall_timer.push_units_processed(ma_batch and ma_batch.count
or 0)
def step(self) -> None:
assert self.loader_thread.is_alive()
with self.load_wait_timer:
buffer_idx, released = self.ready_tower_stacks_buffer.get()
get_num_samples_loaded_into_buffer = 0
with self.grad_timer:
# Use LearnerInfoBuilder as a unified way to build the final
# results dict from `learn_on_loaded_batch` call(s).
# This makes sure results dicts always have the same structure
# no matter the setup (multi-GPU, multi-agent, minibatch SGD,
# tf vs torch).
learner_info_builder = LearnerInfoBuilder(
num_devices=len(self.devices))
for pid in self.policy_map.keys():
# Not a policy-to-train.
if not self.local_worker.is_policy_to_train(pid):
continue
policy = self.policy_map[pid]
default_policy_results = policy.learn_on_loaded_batch(
offset=0, buffer_index=buffer_idx)
learner_info_builder.add_learn_on_batch_results(
default_policy_results)
self.weights_updated = True
get_num_samples_loaded_into_buffer += (
policy.get_num_samples_loaded_into_buffer(buffer_idx))
self.learner_info = learner_info_builder.finalize()
if released:
self.idle_tower_stacks.put(buffer_idx)
self.outqueue.put(
(get_num_samples_loaded_into_buffer, self.learner_info))
self.learner_queue_size.push(self.inqueue.qsize())
def multi_gpu_train_one_step(trainer, train_batch) -> Dict:
"""Multi-GPU version of train_one_step.
Uses the policies' `load_batch_into_buffer` and `learn_on_loaded_batch` methods
to be more efficient wrt CPU/GPU data transfers. For example, when doing multiple
passes through a train batch (e.g. for PPO) using `config.num_sgd_iter`, the
actual train batch is only split once and loaded once into the GPU(s).
Examples:
>>> from ray.rllib.execution.rollout_ops import synchronous_parallel_sample
>>> trainer = [...] # doctest: +SKIP
>>> train_batch = synchronous_parallel_sample(trainer.workers) # doctest: +SKIP
>>> # This trains the policy on one batch.
>>> results = multi_gpu_train_one_step(trainer, train_batch)) # doctest: +SKIP
{"default_policy": ...}
Updates the NUM_ENV_STEPS_TRAINED and NUM_AGENT_STEPS_TRAINED counters as well as
the LOAD_BATCH_TIMER and LEARN_ON_BATCH_TIMER timers of the `trainer` object.
"""
config = trainer.config
workers = trainer.workers
local_worker = workers.local_worker()
num_sgd_iter = config.get("num_sgd_iter", 1)
sgd_minibatch_size = config.get("sgd_minibatch_size",
config["train_batch_size"])
# Determine the number of devices (GPUs or 1 CPU) we use.
num_devices = int(math.ceil(config["num_gpus"] or 1))
# Make sure total batch size is dividable by the number of devices.
# Batch size per tower.
per_device_batch_size = sgd_minibatch_size // num_devices
# Total batch size.
batch_size = per_device_batch_size * num_devices
assert batch_size % num_devices == 0
assert batch_size >= num_devices, "Batch size too small!"
# Handle everything as if multi-agent.
train_batch = train_batch.as_multi_agent()
# Load data into GPUs.
load_timer = trainer._timers[LOAD_BATCH_TIMER]
with load_timer:
num_loaded_samples = {}
for policy_id, batch in train_batch.policy_batches.items():
# Not a policy-to-train.
if not local_worker.is_policy_to_train(policy_id, train_batch):
continue
# Decompress SampleBatch, in case some columns are compressed.
batch.decompress_if_needed()
# Load the entire train batch into the Policy's only buffer
# (idx=0). Policies only have >1 buffers, if we are training
# asynchronously.
num_loaded_samples[policy_id] = local_worker.policy_map[
policy_id].load_batch_into_buffer(batch, buffer_index=0)
# Execute minibatch SGD on loaded data.
learn_timer = trainer._timers[LEARN_ON_BATCH_TIMER]
with learn_timer:
# Use LearnerInfoBuilder as a unified way to build the final
# results dict from `learn_on_loaded_batch` call(s).
# This makes sure results dicts always have the same structure
# no matter the setup (multi-GPU, multi-agent, minibatch SGD,
# tf vs torch).
learner_info_builder = LearnerInfoBuilder(num_devices=num_devices)
for policy_id, samples_per_device in num_loaded_samples.items():
policy = local_worker.policy_map[policy_id]
num_batches = max(
1,
int(samples_per_device) // int(per_device_batch_size))
logger.debug("== sgd epochs for {} ==".format(policy_id))
for _ in range(num_sgd_iter):
permutation = np.random.permutation(num_batches)
for batch_index in range(num_batches):
# Learn on the pre-loaded data in the buffer.
# Note: For minibatch SGD, the data is an offset into
# the pre-loaded entire train batch.
results = policy.learn_on_loaded_batch(
permutation[batch_index] * per_device_batch_size,
buffer_index=0)
learner_info_builder.add_learn_on_batch_results(
results, policy_id)
# Tower reduce and finalize results.
learner_info = learner_info_builder.finalize()
load_timer.push_units_processed(train_batch.count)
learn_timer.push_units_processed(train_batch.count)
trainer._counters[NUM_ENV_STEPS_TRAINED] += train_batch.count
trainer._counters[NUM_AGENT_STEPS_TRAINED] += train_batch.agent_steps()
return learner_info
def __call__(self,
samples: SampleBatchType) -> (SampleBatchType, List[dict]):
_check_sample_batch_type(samples)
# Handle everything as if multi agent.
samples = samples.as_multi_agent()
metrics = _get_shared_metrics()
load_timer = metrics.timers[LOAD_BATCH_TIMER]
learn_timer = metrics.timers[LEARN_ON_BATCH_TIMER]
# Load data into GPUs.
with load_timer:
num_loaded_samples = {}
for policy_id, batch in samples.policy_batches.items():
# Not a policy-to-train.
if not self.local_worker.is_policy_to_train(
policy_id, samples):
continue
# Decompress SampleBatch, in case some columns are compressed.
batch.decompress_if_needed()
# Load the entire train batch into the Policy's only buffer
# (idx=0). Policies only have >1 buffers, if we are training
# asynchronously.
num_loaded_samples[policy_id] = self.local_worker.policy_map[
policy_id].load_batch_into_buffer(batch, buffer_index=0)
# Execute minibatch SGD on loaded data.
with learn_timer:
# Use LearnerInfoBuilder as a unified way to build the final
# results dict from `learn_on_loaded_batch` call(s).
# This makes sure results dicts always have the same structure
# no matter the setup (multi-GPU, multi-agent, minibatch SGD,
# tf vs torch).
learner_info_builder = LearnerInfoBuilder(
num_devices=len(self.devices))
for policy_id, samples_per_device in num_loaded_samples.items():
policy = self.local_worker.policy_map[policy_id]
num_batches = max(
1,
int(samples_per_device) // int(self.per_device_batch_size))
logger.debug("== sgd epochs for {} ==".format(policy_id))
for _ in range(self.num_sgd_iter):
permutation = np.random.permutation(num_batches)
for batch_index in range(num_batches):
# Learn on the pre-loaded data in the buffer.
# Note: For minibatch SGD, the data is an offset into
# the pre-loaded entire train batch.
results = policy.learn_on_loaded_batch(
permutation[batch_index] *
self.per_device_batch_size,
buffer_index=0,
)
learner_info_builder.add_learn_on_batch_results(
results, policy_id)
# Tower reduce and finalize results.
learner_info = learner_info_builder.finalize()
load_timer.push_units_processed(samples.count)
learn_timer.push_units_processed(samples.count)
metrics.counters[STEPS_TRAINED_COUNTER] += samples.count
metrics.counters[STEPS_TRAINED_THIS_ITER_COUNTER] = samples.count
metrics.counters[AGENT_STEPS_TRAINED_COUNTER] += samples.agent_steps()
metrics.info[LEARNER_INFO] = learner_info
if self.workers.remote_workers():
with metrics.timers[WORKER_UPDATE_TIMER]:
weights = ray.put(self.workers.local_worker().get_weights(
self.local_worker.get_policies_to_train()))
for e in self.workers.remote_workers():
e.set_weights.remote(weights, _get_global_vars())
# Also update global vars of the local worker.
self.workers.local_worker().set_global_vars(_get_global_vars())
return samples, learner_info
def training_iteration(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.
async_results: Dict[
ActorHandle, Dict] = asynchronous_parallel_requests(
remote_requests_in_flight=self.remote_requests_in_flight,
actors=self.workers.remote_workers(),
ray_wait_timeout_s=0.0,
max_remote_requests_in_flight_per_actor=1,
remote_fn=sample_and_compute_grads,
)
# 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, result in async_results.items():
# 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"]
# 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)
learner_info_builder.add_learn_on_batch_results_multi_agent(
result["infos"])
# Update global vars of the local worker.
if global_vars:
local_worker.set_global_vars(global_vars)
return learner_info_builder.finalize()
def multi_gpu_train_one_step(trainer, train_batch) -> Dict:
config = trainer.config
workers = trainer.workers
local_worker = workers.local_worker()
num_sgd_iter = config.get("sgd_num_iter", 1)
sgd_minibatch_size = config.get("sgd_minibatch_size", config["train_batch_size"])
# Determine the number of devices (GPUs or 1 CPU) we use.
num_devices = int(math.ceil(config["num_gpus"] or 1))
# Make sure total batch size is dividable by the number of devices.
# Batch size per tower.
per_device_batch_size = sgd_minibatch_size // num_devices
# Total batch size.
batch_size = per_device_batch_size * num_devices
assert batch_size % num_devices == 0
assert batch_size >= num_devices, "Batch size too small!"
# Handle everything as if multi-agent.
train_batch = train_batch.as_multi_agent()
# Load data into GPUs.
load_timer = trainer._timers[LOAD_BATCH_TIMER]
with load_timer:
num_loaded_samples = {}
for policy_id, batch in train_batch.policy_batches.items():
# Not a policy-to-train.
if not local_worker.is_policy_to_train(policy_id, train_batch):
continue
# Decompress SampleBatch, in case some columns are compressed.
batch.decompress_if_needed()
# Load the entire train batch into the Policy's only buffer
# (idx=0). Policies only have >1 buffers, if we are training
# asynchronously.
num_loaded_samples[policy_id] = local_worker.policy_map[
policy_id
].load_batch_into_buffer(batch, buffer_index=0)
# Execute minibatch SGD on loaded data.
learn_timer = trainer._timers[LEARN_ON_BATCH_TIMER]
with learn_timer:
# Use LearnerInfoBuilder as a unified way to build the final
# results dict from `learn_on_loaded_batch` call(s).
# This makes sure results dicts always have the same structure
# no matter the setup (multi-GPU, multi-agent, minibatch SGD,
# tf vs torch).
learner_info_builder = LearnerInfoBuilder(num_devices=num_devices)
for policy_id, samples_per_device in num_loaded_samples.items():
policy = local_worker.policy_map[policy_id]
num_batches = max(1, int(samples_per_device) // int(per_device_batch_size))
logger.debug("== sgd epochs for {} ==".format(policy_id))
for _ in range(num_sgd_iter):
permutation = np.random.permutation(num_batches)
for batch_index in range(num_batches):
# Learn on the pre-loaded data in the buffer.
# Note: For minibatch SGD, the data is an offset into
# the pre-loaded entire train batch.
results = policy.learn_on_loaded_batch(
permutation[batch_index] * per_device_batch_size, buffer_index=0
)
learner_info_builder.add_learn_on_batch_results(results, policy_id)
# Tower reduce and finalize results.
learner_info = learner_info_builder.finalize()
load_timer.push_units_processed(train_batch.count)
learn_timer.push_units_processed(train_batch.count)
trainer._counters[NUM_ENV_STEPS_TRAINED] += train_batch.count
trainer._counters[NUM_AGENT_STEPS_TRAINED] += train_batch.agent_steps()
# Update weights - after learning on the local worker - on all remote
# workers.
if workers.remote_workers():
with trainer._timers[WORKER_UPDATE_TIMER]:
weights = ray.put(
local_worker.get_weights(
local_worker.get_policies_to_train(train_batch)
)
)
for e in workers.remote_workers():
e.set_weights.remote(weights)
return learner_info
def do_minibatch_sgd(
samples,
policies,
local_worker,
num_sgd_iter,
sgd_minibatch_size,
standardize_fields,
):
"""Execute minibatch SGD.
Args:
samples: Batch of samples to optimize.
policies: Dictionary of policies to optimize.
local_worker: Master rollout worker instance.
num_sgd_iter: Number of epochs of optimization to take.
sgd_minibatch_size: Size of minibatches to use for optimization.
standardize_fields: List of sample field names that should be
normalized prior to optimization.
Returns:
averaged info fetches over the last SGD epoch taken.
"""
# Handle everything as if multi-agent.
samples = samples.as_multi_agent()
# Use LearnerInfoBuilder as a unified way to build the final
# results dict from `learn_on_loaded_batch` call(s).
# This makes sure results dicts always have the same structure
# no matter the setup (multi-GPU, multi-agent, minibatch SGD,
# tf vs torch).
learner_info_builder = LearnerInfoBuilder(num_devices=1)
for policy_id, policy in policies.items():
if policy_id not in samples.policy_batches:
continue
batch = samples.policy_batches[policy_id]
for field in standardize_fields:
batch[field] = standardized(batch[field])
# Check to make sure that the sgd_minibatch_size is not smaller
# than max_seq_len otherwise this will cause indexing errors while
# performing sgd when using a RNN or Attention model
if (
policy.is_recurrent()
and policy.config["model"]["max_seq_len"] > sgd_minibatch_size
):
raise ValueError(
"`sgd_minibatch_size` ({}) cannot be smaller than"
"`max_seq_len` ({}).".format(
sgd_minibatch_size, policy.config["model"]["max_seq_len"]
)
)
for i in range(num_sgd_iter):
for minibatch in minibatches(batch, sgd_minibatch_size):
results = (
local_worker.learn_on_batch(
MultiAgentBatch({policy_id: minibatch}, minibatch.count)
)
)[policy_id]
learner_info_builder.add_learn_on_batch_results(results, policy_id)
learner_info = learner_info_builder.finalize()
return learner_info