INSTRUCTION stringlengths 1 8.43k | RESPONSE stringlengths 75 104k |
|---|---|
Get the error messages for all drivers or a specific driver. | def error_messages(self, driver_id=None):
"""Get the error messages for all drivers or a specific driver.
Args:
driver_id: The specific driver to get the errors for. If this is
None, then this method retrieves the errors for all drivers.
Returns:
A dicti... |
Get checkpoint info for the given actor id. Args: actor_id: Actor s ID. Returns: A dictionary with information about the actor s checkpoint IDs and their timestamps. | def actor_checkpoint_info(self, actor_id):
"""Get checkpoint info for the given actor id.
Args:
actor_id: Actor's ID.
Returns:
A dictionary with information about the actor's checkpoint IDs and
their timestamps.
"""
self._check_connected()
... |
Returns the total length of all of the flattened variables. | def get_flat_size(self):
"""Returns the total length of all of the flattened variables.
Returns:
The length of all flattened variables concatenated.
"""
return sum(
np.prod(v.get_shape().as_list()) for v in self.variables.values()) |
Gets the weights and returns them as a flat array. | def get_flat(self):
"""Gets the weights and returns them as a flat array.
Returns:
1D Array containing the flattened weights.
"""
self._check_sess()
return np.concatenate([
v.eval(session=self.sess).flatten()
for v in self.variables.values()
... |
Sets the weights to new_weights converting from a flat array. | def set_flat(self, new_weights):
"""Sets the weights to new_weights, converting from a flat array.
Note:
You can only set all weights in the network using this function,
i.e., the length of the array must match get_flat_size.
Args:
new_weights (np.ndarray): ... |
Returns a dictionary containing the weights of the network. | def get_weights(self):
"""Returns a dictionary containing the weights of the network.
Returns:
Dictionary mapping variable names to their weights.
"""
self._check_sess()
return {
k: v.eval(session=self.sess)
for k, v in self.variables.items()
... |
Sets the weights to new_weights. | def set_weights(self, new_weights):
"""Sets the weights to new_weights.
Note:
Can set subsets of variables as well, by only passing in the
variables you want to be set.
Args:
new_weights (Dict): Dictionary mapping variable names to their
weig... |
Construct a serialized ErrorTableData object. | def construct_error_message(driver_id, error_type, message, timestamp):
"""Construct a serialized ErrorTableData object.
Args:
driver_id: The ID of the driver that the error should go to. If this is
nil, then the error will go to all drivers.
error_type: The type of the error.
... |
Initialize synchronously. | def init():
"""
Initialize synchronously.
"""
loop = asyncio.get_event_loop()
if loop.is_running():
raise Exception("You must initialize the Ray async API by calling "
"async_api.init() or async_api.as_future(obj) before "
"the event loop start... |
Manually shutdown the async API. | def shutdown():
"""Manually shutdown the async API.
Cancels all related tasks and all the socket transportation.
"""
global handler, transport, protocol
if handler is not None:
handler.close()
transport.close()
handler = None
transport = None
protocol = None |
This removes some non - critical state from the primary Redis shard. | def flush_redis_unsafe(redis_client=None):
"""This removes some non-critical state from the primary Redis shard.
This removes the log files as well as the event log from Redis. This can
be used to try to address out-of-memory errors caused by the accumulation
of metadata in Redis. However, it will only... |
This removes some critical state from the Redis shards. | def flush_task_and_object_metadata_unsafe():
"""This removes some critical state from the Redis shards.
In a multitenant environment, this will flush metadata for all jobs, which
may be undesirable.
This removes all of the object and task metadata. This can be used to try
to address out-of-memory ... |
This removes some critical state from the Redis shards. | def flush_finished_tasks_unsafe():
"""This removes some critical state from the Redis shards.
In a multitenant environment, this will flush metadata for all jobs, which
may be undesirable.
This removes all of the metadata for finished tasks. This can be used to
try to address out-of-memory errors ... |
This removes some critical state from the Redis shards. | def flush_evicted_objects_unsafe():
"""This removes some critical state from the Redis shards.
In a multitenant environment, this will flush metadata for all jobs, which
may be undesirable.
This removes all of the metadata for objects that have been evicted. This
can be used to try to address out-... |
Creates a copy of self using existing input placeholders. | def copy(self, existing_inputs):
"""Creates a copy of self using existing input placeholders."""
return PPOPolicyGraph(
self.observation_space,
self.action_space,
self.config,
existing_inputs=existing_inputs) |
deepnn builds the graph for a deep net for classifying digits. | def deepnn(x):
"""deepnn builds the graph for a deep net for classifying digits.
Args:
x: an input tensor with the dimensions (N_examples, 784), where 784 is
the number of pixels in a standard MNIST image.
Returns:
A tuple (y, keep_prob). y is a tensor of shape (N_examples, 10)... |
Get signature parameters | def get_signature_params(func):
"""Get signature parameters
Support Cython functions by grabbing relevant attributes from the Cython
function and attaching to a no-op function. This is somewhat brittle, since
funcsigs may change, but given that funcsigs is written to a PEP, we hope
it is relatively... |
Check if we support the signature of this function. | def check_signature_supported(func, warn=False):
"""Check if we support the signature of this function.
We currently do not allow remote functions to have **kwargs. We also do not
support keyword arguments in conjunction with a *args argument.
Args:
func: The function whose signature should be... |
Extract the function signature from the function. | def extract_signature(func, ignore_first=False):
"""Extract the function signature from the function.
Args:
func: The function whose signature should be extracted.
ignore_first: True if the first argument should be ignored. This should
be used when func is a method of a class.
... |
Extend the arguments that were passed into a function. | def extend_args(function_signature, args, kwargs):
"""Extend the arguments that were passed into a function.
This extends the arguments that were passed into a function with the
default arguments provided in the function definition.
Args:
function_signature: The function signature of the funct... |
Poll for cloud resource manager operation until finished. | def wait_for_crm_operation(operation):
"""Poll for cloud resource manager operation until finished."""
logger.info("wait_for_crm_operation: "
"Waiting for operation {} to finish...".format(operation))
for _ in range(MAX_POLLS):
result = crm.operations().get(name=operation["name"]).e... |
Poll for global compute operation until finished. | def wait_for_compute_global_operation(project_name, operation):
"""Poll for global compute operation until finished."""
logger.info("wait_for_compute_global_operation: "
"Waiting for operation {} to finish...".format(
operation["name"]))
for _ in range(MAX_POLLS):
... |
Returns the ith default gcp_key_pair_name. | def key_pair_name(i, region, project_id, ssh_user):
"""Returns the ith default gcp_key_pair_name."""
key_name = "{}_gcp_{}_{}_{}".format(RAY, region, project_id, ssh_user, i)
return key_name |
Returns public and private key paths for a given key_name. | def key_pair_paths(key_name):
"""Returns public and private key paths for a given key_name."""
public_key_path = os.path.expanduser("~/.ssh/{}.pub".format(key_name))
private_key_path = os.path.expanduser("~/.ssh/{}.pem".format(key_name))
return public_key_path, private_key_path |
Create public and private ssh - keys. | def generate_rsa_key_pair():
"""Create public and private ssh-keys."""
key = rsa.generate_private_key(
backend=default_backend(), public_exponent=65537, key_size=2048)
public_key = key.public_key().public_bytes(
serialization.Encoding.OpenSSH,
serialization.PublicFormat.OpenSSH).de... |
Setup a Google Cloud Platform Project. | def _configure_project(config):
"""Setup a Google Cloud Platform Project.
Google Compute Platform organizes all the resources, such as storage
buckets, users, and instances under projects. This is different from
aws ec2 where everything is global.
"""
project_id = config["provider"].get("projec... |
Setup a gcp service account with IAM roles. | def _configure_iam_role(config):
"""Setup a gcp service account with IAM roles.
Creates a gcp service acconut and binds IAM roles which allow it to control
control storage/compute services. Specifically, the head node needs to have
an IAM role that allows it to create further gce instances and store it... |
Configure SSH access using an existing key pair if possible. | def _configure_key_pair(config):
"""Configure SSH access, using an existing key pair if possible.
Creates a project-wide ssh key that can be used to access all the instances
unless explicitly prohibited by instance config.
The ssh-keys created by ray are of format:
[USERNAME]:ssh-rsa [KEY_VALUE... |
Pick a reasonable subnet if not specified by the config. | def _configure_subnet(config):
"""Pick a reasonable subnet if not specified by the config."""
# Rationale: avoid subnet lookup if the network is already
# completely manually configured
if ("networkInterfaces" in config["head_node"]
and "networkInterfaces" in config["worker_nodes"]):
... |
Add new IAM roles for the service account. | def _add_iam_policy_binding(service_account, roles):
"""Add new IAM roles for the service account."""
project_id = service_account["projectId"]
email = service_account["email"]
member_id = "serviceAccount:" + email
policy = crm.projects().getIamPolicy(resource=project_id).execute()
already_con... |
Inserts an ssh - key into project commonInstanceMetadata | def _create_project_ssh_key_pair(project, public_key, ssh_user):
"""Inserts an ssh-key into project commonInstanceMetadata"""
key_parts = public_key.split(" ")
# Sanity checks to make sure that the generated key matches expectation
assert len(key_parts) == 2, key_parts
assert key_parts[0] == "ssh-... |
An experimental alternate way to submit remote functions. | def _remote(self,
args=None,
kwargs=None,
num_return_vals=None,
num_cpus=None,
num_gpus=None,
resources=None):
"""An experimental alternate way to submit remote functions."""
worker = ray.worker.get_global_wo... |
Append an object to the linked list. | def append(self, future):
"""Append an object to the linked list.
Args:
future (PlasmaObjectFuture): A PlasmaObjectFuture instance.
"""
future.prev = self.tail
if self.tail is None:
assert self.head is None
self.head = future
else:
... |
Remove an object from the linked list. | def remove(self, future):
"""Remove an object from the linked list.
Args:
future (PlasmaObjectFuture): A PlasmaObjectFuture instance.
"""
if self._loop.get_debug():
logger.debug("Removing %s from the linked list.", future)
if future.prev is None:
... |
Manually cancel all tasks assigned to this event loop. | def cancel(self, *args, **kwargs):
"""Manually cancel all tasks assigned to this event loop."""
# Because remove all futures will trigger `set_result`,
# we cancel itself first.
super().cancel()
for future in self.traverse():
# All cancelled futures should have callba... |
Complete all tasks. | def set_result(self, result):
"""Complete all tasks. """
for future in self.traverse():
# All cancelled futures should have callbacks to removed itself
# from this linked list. However, these callbacks are scheduled in
# an event loop, so we could still find them in o... |
Traverse this linked list. | def traverse(self):
"""Traverse this linked list.
Yields:
PlasmaObjectFuture: PlasmaObjectFuture instances.
"""
current = self.head
while current is not None:
yield current
current = current.next |
Process notifications. | def process_notifications(self, messages):
"""Process notifications."""
for object_id, object_size, metadata_size in messages:
if object_size > 0 and object_id in self._waiting_dict:
linked_list = self._waiting_dict[object_id]
self._complete_future(linked_list... |
Turn an object_id into a Future object. | def as_future(self, object_id, check_ready=True):
"""Turn an object_id into a Future object.
Args:
object_id: A Ray's object_id.
check_ready (bool): If true, check if the object_id is ready.
Returns:
PlasmaObjectFuture: A future object that waits the object_... |
Returns a list of all trials information. | def get_all_trials(self):
"""Returns a list of all trials' information."""
response = requests.get(urljoin(self._path, "trials"))
return self._deserialize(response) |
Returns trial information by trial_id. | def get_trial(self, trial_id):
"""Returns trial information by trial_id."""
response = requests.get(
urljoin(self._path, "trials/{}".format(trial_id)))
return self._deserialize(response) |
Adds a trial by name and specification ( dict ). | def add_trial(self, name, specification):
"""Adds a trial by name and specification (dict)."""
payload = {"name": name, "spec": specification}
response = requests.post(urljoin(self._path, "trials"), json=payload)
return self._deserialize(response) |
Requests to stop trial by trial_id. | def stop_trial(self, trial_id):
"""Requests to stop trial by trial_id."""
response = requests.put(
urljoin(self._path, "trials/{}".format(trial_id)))
return self._deserialize(response) |
Apply the given function to each remote worker. | def foreach_worker(self, fn):
"""Apply the given function to each remote worker.
Returns:
List of results from applying the function.
"""
results = ray.get([w.foreach_worker.remote(fn) for w in self.workers])
return results |
Apply the given function to each model replica in each worker. | def foreach_model(self, fn):
"""Apply the given function to each model replica in each worker.
Returns:
List of results from applying the function.
"""
results = ray.get([w.foreach_model.remote(fn) for w in self.workers])
out = []
for r in results:
... |
Apply the given function to a single model replica. | def for_model(self, fn):
"""Apply the given function to a single model replica.
Returns:
Result from applying the function.
"""
return ray.get(self.workers[0].for_model.remote(fn)) |
Run a single SGD step. | def step(self, fetch_stats=False):
"""Run a single SGD step.
Arguments:
fetch_stats (bool): Whether to return stats from the step. This can
slow down the computation by acting as a global barrier.
"""
if self.strategy == "ps":
return _distributed_... |
Wrapper for starting a router and register it. | def start_router(router_class, router_name):
"""Wrapper for starting a router and register it.
Args:
router_class: The router class to instantiate.
router_name: The name to give to the router.
Returns:
A handle to newly started router actor.
"""
handle = router_class.remote... |
Returns a list of one - hot encodings for all parameters. | def generate_random_one_hot_encoding(self):
"""Returns a list of one-hot encodings for all parameters.
1 one-hot np.array for 1 parameter,
and the 1's place is randomly chosen.
"""
encoding = []
for ps in self.param_list:
one_hot = np.zeros(ps.choices_count()... |
Apply one hot encoding to generate a specific config. | def apply_one_hot_encoding(self, one_hot_encoding):
"""Apply one hot encoding to generate a specific config.
Arguments:
one_hot_encoding (list): A list of one hot encodings,
1 for each parameter. The shape of each encoding
should match that ``ParameterSpace`... |
Pin an object in the object store. | def pin_in_object_store(obj):
"""Pin an object in the object store.
It will be available as long as the pinning process is alive. The pinned
object can be retrieved by calling get_pinned_object on the identifier
returned by this call.
"""
obj_id = ray.put(_to_pinnable(obj))
_pinned_objects... |
Retrieve a pinned object from the object store. | def get_pinned_object(pinned_id):
"""Retrieve a pinned object from the object store."""
from ray import ObjectID
return _from_pinnable(
ray.get(
ObjectID(base64.b64decode(pinned_id[len(PINNED_OBJECT_PREFIX):])))) |
Returns a new dict that is d1 and d2 deep merged. | def merge_dicts(d1, d2):
"""Returns a new dict that is d1 and d2 deep merged."""
merged = copy.deepcopy(d1)
deep_update(merged, d2, True, [])
return merged |
Updates original dict with values from new_dict recursively. If new key is introduced in new_dict then if new_keys_allowed is not True an error will be thrown. Further for sub - dicts if the key is in the whitelist then new subkeys can be introduced. | def deep_update(original, new_dict, new_keys_allowed, whitelist):
"""Updates original dict with values from new_dict recursively.
If new key is introduced in new_dict, then if new_keys_allowed is not
True, an error will be thrown. Further, for sub-dicts, if the key is
in the whitelist, then new subkeys ... |
Similar to completed but only returns once the object is local. | def completed_prefetch(self, blocking_wait=False, max_yield=999):
"""Similar to completed but only returns once the object is local.
Assumes obj_id only is one id."""
for worker, obj_id in self.completed(blocking_wait=blocking_wait):
plasma_id = ray.pyarrow.plasma.ObjectID(obj_id.b... |
Notify that some evaluators may be removed. | def reset_evaluators(self, evaluators):
"""Notify that some evaluators may be removed."""
for obj_id, ev in self._tasks.copy().items():
if ev not in evaluators:
del self._tasks[obj_id]
del self._objects[obj_id]
ok = []
for ev, obj_id in self._f... |
Iterate over train batches. | def iter_train_batches(self, max_yield=999):
"""Iterate over train batches.
Arguments:
max_yield (int): Max number of batches to iterate over in this
cycle. Setting this avoids iter_train_batches returning too
much data at once.
"""
for ev, s... |
Create or updates an autoscaling Ray cluster from a config json. | def create_or_update_cluster(config_file, override_min_workers,
override_max_workers, no_restart, restart_only,
yes, override_cluster_name):
"""Create or updates an autoscaling Ray cluster from a config json."""
config = yaml.load(open(config_file).read(... |
Destroys all nodes of a Ray cluster described by a config json. | def teardown_cluster(config_file, yes, workers_only, override_cluster_name):
"""Destroys all nodes of a Ray cluster described by a config json."""
config = yaml.load(open(config_file).read())
if override_cluster_name is not None:
config["cluster_name"] = override_cluster_name
validate_config(co... |
Kills a random Raylet worker. | def kill_node(config_file, yes, override_cluster_name):
"""Kills a random Raylet worker."""
config = yaml.load(open(config_file).read())
if override_cluster_name is not None:
config["cluster_name"] = override_cluster_name
config = _bootstrap_config(config)
confirm("This will kill a node in... |
Create the cluster head node which in turn creates the workers. | def get_or_create_head_node(config, config_file, no_restart, restart_only, yes,
override_cluster_name):
"""Create the cluster head node, which in turn creates the workers."""
provider = get_node_provider(config["provider"], config["cluster_name"])
try:
head_node_tags = {
... |
Attaches to a screen for the specified cluster. | def attach_cluster(config_file, start, use_tmux, override_cluster_name, new):
"""Attaches to a screen for the specified cluster.
Arguments:
config_file: path to the cluster yaml
start: whether to start the cluster if it isn't up
use_tmux: whether to use tmux as multiplexer
overr... |
Runs a command on the specified cluster. | def exec_cluster(config_file, cmd, docker, screen, tmux, stop, start,
override_cluster_name, port_forward):
"""Runs a command on the specified cluster.
Arguments:
config_file: path to the cluster yaml
cmd: command to run
docker: whether to run command in docker containe... |
Rsyncs files. | def rsync(config_file, source, target, override_cluster_name, down):
"""Rsyncs files.
Arguments:
config_file: path to the cluster yaml
source: source dir
target: target dir
override_cluster_name: set the name of the cluster
down: whether we're syncing remote -> local
... |
Returns head node IP for given configuration file if exists. | def get_head_node_ip(config_file, override_cluster_name):
"""Returns head node IP for given configuration file if exists."""
config = yaml.load(open(config_file).read())
if override_cluster_name is not None:
config["cluster_name"] = override_cluster_name
provider = get_node_provider(config["pr... |
Returns worker node IPs for given configuration file. | def get_worker_node_ips(config_file, override_cluster_name):
"""Returns worker node IPs for given configuration file."""
config = yaml.load(open(config_file).read())
if override_cluster_name is not None:
config["cluster_name"] = override_cluster_name
provider = get_node_provider(config["provid... |
Implements train () for a Function API. | def _train(self):
"""Implements train() for a Function API.
If the RunnerThread finishes without reporting "done",
Tune will automatically provide a magic keyword __duplicate__
along with a result with "done=True". The TrialRunner will handle the
result accordingly (see tune/tri... |
Returns logits and aux_logits from images. | def build_network(self,
images,
phase_train=True,
nclass=1001,
image_depth=3,
data_type=tf.float32,
data_format="NCHW",
use_tf_layers=True,
fp16... |
Helper class for renaming Agent = > Trainer with a warning. | def renamed_class(cls):
"""Helper class for renaming Agent => Trainer with a warning."""
class DeprecationWrapper(cls):
def __init__(self, config=None, env=None, logger_creator=None):
old_name = cls.__name__.replace("Trainer", "Agent")
new_name = cls.__name__
logger.... |
Profile a span of time so that it appears in the timeline visualization. | def profile(event_type, extra_data=None):
"""Profile a span of time so that it appears in the timeline visualization.
Note that this only works in the raylet code path.
This function can be used as follows (both on the driver or within a task).
.. code-block:: python
with ray.profile("custom... |
Drivers run this as a thread to flush profile data in the background. | def _periodically_flush_profile_events(self):
"""Drivers run this as a thread to flush profile data in the
background."""
# Note(rkn): This is run on a background thread in the driver. It uses
# the raylet client. This should be ok because it doesn't read
# from the raylet client... |
Push the logged profiling data to the global control store. | def flush_profile_data(self):
"""Push the logged profiling data to the global control store."""
with self.lock:
events = self.events
self.events = []
if self.worker.mode == ray.WORKER_MODE:
component_type = "worker"
else:
component_type = ... |
Add a key - value pair to the extra_data dict. | def set_attribute(self, key, value):
"""Add a key-value pair to the extra_data dict.
This can be used to add attributes that are not available when
ray.profile was called.
Args:
key: The attribute name.
value: The attribute value.
"""
if not isin... |
Syncs the local logdir on driver to worker if possible. | def sync_to_worker_if_possible(self):
"""Syncs the local logdir on driver to worker if possible.
Requires ray cluster to be started with the autoscaler. Also requires
rsync to be installed.
"""
if self.worker_ip == self.local_ip:
return
ssh_key = get_ssh_key(... |
Forward pass for the mixer. | def forward(self, agent_qs, states):
"""Forward pass for the mixer.
Arguments:
agent_qs: Tensor of shape [B, T, n_agents, n_actions]
states: Tensor of shape [B, T, state_dim]
"""
bs = agent_qs.size(0)
states = states.reshape(-1, self.state_dim)
ag... |
Passes the result to SigOpt unless early terminated or errored. | def on_trial_complete(self,
trial_id,
result=None,
error=False,
early_terminated=False):
"""Passes the result to SigOpt unless early terminated or errored.
If a trial fails, it will be reported as a ... |
Returns ranks in [ 0 len ( x )) | def compute_ranks(x):
"""Returns ranks in [0, len(x))
Note: This is different from scipy.stats.rankdata, which returns ranks in
[1, len(x)].
"""
assert x.ndim == 1
ranks = np.empty(len(x), dtype=int)
ranks[x.argsort()] = np.arange(len(x))
return ranks |
Bottleneck block with identity short - cut for ResNet v1. | def bottleneck_block_v1(cnn, depth, depth_bottleneck, stride):
"""Bottleneck block with identity short-cut for ResNet v1.
Args:
cnn: the network to append bottleneck blocks.
depth: the number of output filters for this bottleneck block.
depth_bottleneck: the number of bottleneck filters for this bloc... |
Bottleneck block with identity short - cut. | def bottleneck_block(cnn, depth, depth_bottleneck, stride, pre_activation):
"""Bottleneck block with identity short-cut.
Args:
cnn: the network to append bottleneck blocks.
depth: the number of output filters for this bottleneck block.
depth_bottleneck: the number of bottleneck filters for this block... |
Residual block with identity short - cut. | def residual_block(cnn, depth, stride, pre_activation):
"""Residual block with identity short-cut.
Args:
cnn: the network to append residual blocks.
depth: the number of output filters for this residual block.
stride: Stride used in the first layer of the residual block.
pre_activation: use pre_a... |
Applies updates from the buffer of another filter. | def apply_changes(self, other, with_buffer=False):
"""Applies updates from the buffer of another filter.
Params:
other (MeanStdFilter): Other filter to apply info from
with_buffer (bool): Flag for specifying if the buffer should be
copied from other.
Exa... |
Returns a copy of Filter. | def copy(self):
"""Returns a copy of Filter."""
other = MeanStdFilter(self.shape)
other.sync(self)
return other |
Syncs all fields together from other filter. | def sync(self, other):
"""Syncs all fields together from other filter.
Examples:
>>> a = MeanStdFilter(())
>>> a(1)
>>> a(2)
>>> print([a.rs.n, a.rs.mean, a.buffer.n])
[2, array(1.5), 2]
>>> b = MeanStdFilter(())
>>> b(... |
Returns non - concurrent version of current class | def as_serializable(self):
"""Returns non-concurrent version of current class"""
other = MeanStdFilter(self.shape)
other.sync(self)
return other |
Returns a copy of Filter. | def copy(self):
"""Returns a copy of Filter."""
other = ConcurrentMeanStdFilter(self.shape)
other.sync(self)
return other |
f ( x ) = - sum { sin ( xi ) * [ sin ( i * xi^2/ pi ) ] ^ ( 2m ) } | def michalewicz_function(config, reporter):
"""f(x) = -sum{sin(xi) * [sin(i*xi^2 / pi)]^(2m)}"""
import numpy as np
x = np.array(
[config["x1"], config["x2"], config["x3"], config["x4"], config["x5"]])
sin_x = np.sin(x)
z = (np.arange(1, 6) / np.pi * (x * x))
sin_z = np.power(np.sin(z), ... |
Parse integer with power - of - 2 suffix eg. 32k. | def parse_general_int(s):
"""Parse integer with power-of-2 suffix eg. 32k."""
mo = re.match(r"(\d+)([KkMGT]?)$", s)
if mo:
i, suffix = mo.group(1, 2)
v = int(i)
if suffix:
if suffix == "K" or suffix == "k":
v *= 1024
elif suffix == "M":
... |
Parse all_reduce_spec. | def parse_all_reduce_spec(all_reduce_spec):
"""Parse all_reduce_spec.
Args:
all_reduce_spec: a string specifying a combination of all-reduce
algorithms to apply for gradient reduction.
Returns:
a list of AllReduceSpecTuple.
Raises:
ValueError: all_reduce_spec is not well-formed.
An all... |
Build list of device prefix names for all_reduce. | def build_all_reduce_device_prefixes(job_name, num_tasks):
"""Build list of device prefix names for all_reduce.
Args:
job_name: "worker", "ps" or "localhost".
num_tasks: number of jobs across which device names should be generated.
Returns:
A list of device name prefix strings. Each element spell... |
Group device names into groups of group_size. | def group_device_names(devices, group_size):
"""Group device names into groups of group_size.
Args:
devices: list of strings naming devices.
group_size: int >= 1
Returns:
list of lists of devices, where each inner list is group_size long,
and each device appears at least once in an inner lis... |
Break gradients into two sets according to tensor size. | def split_grads_by_size(threshold_size, device_grads):
"""Break gradients into two sets according to tensor size.
Args:
threshold_size: int size cutoff for small vs large tensor.
device_grads: List of lists of (gradient, variable) tuples. The outer
list is over devices. The inner list is over in... |
Calculate the average gradient for a shared variable across all towers. | def aggregate_single_gradient(grad_and_vars, use_mean, check_inf_nan):
"""Calculate the average gradient for a shared variable across all towers.
Note that this function provides a synchronization point across all towers.
Args:
grad_and_vars: A list or tuple of (gradient, variable) tuples. Each
(gra... |
Aggregate gradients controlling device for the aggregation. | def aggregate_gradients_using_copy_with_device_selection(
tower_grads, avail_devices, use_mean=True, check_inf_nan=False):
"""Aggregate gradients, controlling device for the aggregation.
Args:
tower_grads: List of lists of (gradient, variable) tuples. The outer list
is over towers. The inner li... |
Apply all - reduce algorithm over specified gradient tensors. | def sum_grad_and_var_all_reduce(grad_and_vars,
num_workers,
alg,
gpu_indices,
aux_devices=None,
num_shards=1):
"""Apply all-reduce algorithm over specified ... |
Apply all - reduce algorithm over specified gradient tensors. | def sum_gradients_all_reduce(dev_prefixes,
tower_grads,
num_workers,
alg,
num_shards,
gpu_indices,
agg_small_grads_max_bytes=0):
"""Apply all-... |
Extract consecutive ranges and singles from index_list. | def extract_ranges(index_list, range_size_limit=32):
"""Extract consecutive ranges and singles from index_list.
Args:
index_list: List of monotone increasing non-negative integers.
range_size_limit: Largest size range to return. If a larger
consecutive range exists it will be returned as multiple
... |
Form the concatenation of a specified range of gradient tensors. | def pack_range(key, packing, grad_vars, rng):
"""Form the concatenation of a specified range of gradient tensors.
Args:
key: Value under which to store meta-data in packing that will be used
later to restore the grad_var list structure.
packing: Dict holding data describing packed ranges of small t... |
Unpack a previously packed collection of gradient tensors. | def unpack_grad_tuple(gv, gpt):
"""Unpack a previously packed collection of gradient tensors.
Args:
gv: A (grad, var) pair to be unpacked.
gpt: A GradPackTuple describing the packing operation that produced gv.
Returns:
A list of (grad, var) pairs corresponding to the values that were
origina... |
Concatenate gradients together more intelligently. | def pack_small_tensors(tower_grads, max_bytes=0):
"""Concatenate gradients together more intelligently.
Does binpacking
Args:
tower_grads: List of lists of (gradient, variable) tuples.
max_bytes: Int giving max number of bytes in a tensor that
may be considered small.
"""
assert max_bytes >... |
Undo the structure alterations to tower_grads done by pack_small_tensors. | def unpack_small_tensors(tower_grads, packing):
"""Undo the structure alterations to tower_grads done by pack_small_tensors.
Args:
tower_grads: List of List of (grad, var) tuples.
packing: A dict generated by pack_small_tensors describing the changes
it made to tower_grads.
Returns:
new_towe... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.