|TensorFlow 1 version||View source on GitHub|
A distribution strategy for synchronous training on multiple workers.
tf.distribute.experimental.MultiWorkerMirroredStrategy( communication=tf.distribute.experimental.CollectiveCommunication.AUTO, cluster_resolver=None )
Used in the notebooks
|Used in the guide||Used in the tutorials|
This strategy implements synchronous distributed training across multiple
workers, each with potentially multiple GPUs. Similar to
tf.distribute.MirroredStrategy, it creates copies of all variables in the
model on each device across all workers.
It uses CollectiveOps's implementation of multi-worker all-reduce to to keep variables in sync. A collective op is a single op in the TensorFlow graph which can automatically choose an all-reduce algorithm in the TensorFlow runtime according to hardware, network topology and tensor sizes.
By default it uses all local GPUs or CPU for single-worker training.
When 'TF_CONFIG' environment variable is set, it parses cluster_spec, task_type and task_id from 'TF_CONFIG' and turns into a multi-worker strategy which mirrores models on GPUs of all machines in a cluster. In the current implementation, it uses all GPUs in a cluster and it assumes all workers have the same number of GPUs.
You can also pass a
when instantiating the strategy. The task_type, task_id etc. will be parsed
from the resolver instance instead of from the
TF_CONFIG env var.
It supports both eager mode and graph mode. However, for eager mode, it has to set up the eager context in its constructor and therefore all ops in eager mode have to run after the strategy object is created.
communication: optional Enum of type
distribute.experimental.CollectiveCommunication. This provides a way for the user to override the choice of collective op communication. Possible values include
distribute.cluster_resolver.ClusterResolverobject. The default ClusterResolver that is used is the TFConfigClusterResolver which is instantiated from the TF_CONFIG env var.
tf.distribute.StrategyExtendedwith additional methods.
num_replicas_in_sync: Returns number of replicas over which gradients are aggregated.
experimental_distribute_dataset( dataset )
Distributes a tf.data.Dataset instance provided via
The returned distributed dataset can be iterated over similar to how regular datasets can. NOTE: Currently, the user cannot add any more transformations to a distributed dataset.
The following is an example:
strategy = tf.distribute.MirroredStrategy() # Create a dataset dataset = dataset_ops.Dataset.TFRecordDataset([ "/a/1.tfr", "/a/2.tfr", "/a/3.tfr", "/a/4.tfr"]) # Distribute that dataset dist_dataset = strategy.experimental_distribute_dataset(dataset) # Iterate over the distributed dataset for x in dist_dataset: # process dataset elements strategy.experimental_run_v2(train_step, args=(x,))
We will assume that the input dataset is batched by the global batch size. With this assumption, we will make a best effort to divide each batch across all the replicas (one or more workers).
In a multi-worker setting, we will first attempt to distribute the dataset by attempting to detect whether the dataset is being created out of ReaderDatasets (e.g. TFRecordDataset, TextLineDataset, etc.) and if so, attempting to shard the input files. Note that there has to be at least one input file per worker. If you have less than one input file per worker, we suggest that you should disable distributing your dataset using the method below.
If that attempt is unsuccessful (e.g. the dataset is created from a
Dataset.range), we will shard the dataset evenly at the end by appending a
.shard operation to the end of the processing pipeline. This will cause
the entire preprocessing pipeline for all the data to be run on every
worker, and each worker will do redundant work. We will print a warning
if this method of sharding is selected. In this case, consider using
You can disable dataset sharding across workers using the
Within each worker, we will also split the data among all the worker devices (if more than one a present), and this will happen even if multi-worker sharding is disabled using the method above.
If the above batch splitting and dataset sharding logic is undesirable,
experimental_distribute_datasets_from_function instead, which
does not do any automatic splitting or sharding.
tf.data.Datasetthat will be sharded across all replicas using the rules stated above.
Dataset", which acts like a
it produces "per-replica" values.
experimental_distribute_datasets_from_function( dataset_fn )
tf.data.Dataset instances created by calls to
dataset_fn will be called once for each worker in the strategy. Each
replica on that worker will dequeue one batch of inputs from the local
Dataset (i.e. if a worker has two replicas, two batches will be dequeued
Dataset every step).
This method can be used for several purposes. For example, where
experimental_distribute_dataset is unable to shard the input files, this
method might be used to manually shard the dataset (avoiding the slow
fallback behavior in
experimental_distribute_dataset). In cases where the
dataset is infinite, this sharding can be done by creating dataset replicas
that differ only in their random seed.
experimental_distribute_dataset may also sometimes fail to split the
batch across replicas on a worker. In that case, this method can be used
where that limitation does not exist.
dataset_fn should take an
tf.distribute.InputContext instance where
information about batching and input replication can be accessed:
def dataset_fn(input_context): batch_size = input_context.get_per_replica_batch_size(global_batch_size) d = tf.data.Dataset.from_tensors([[1.]]).repeat().batch(batch_size) return d.shard( input_context.num_input_pipelines, input_context.input_pipeline_id) inputs = strategy.experimental_distribute_datasets_from_function(dataset_fn) for batch in inputs: replica_results = strategy.experimental_run_v2(replica_fn, args=(batch,))
tf.data.Dataset returned by
dataset_fn should have a
per-replica batch size, unlike
experimental_distribute_dataset, which uses
the global batch size. This may be computed using
dataset_fn: A function taking a
tf.distribute.InputContextinstance and returning a
Dataset", which acts like a
it produces "per-replica" values.
experimental_local_results( value )
Returns the list of all local per-replica values contained in
value: A value returned by
extended.call_for_each_replica(), or a variable created in
A tuple of values contained in
value represents a single
value, this returns
experimental_make_numpy_dataset( numpy_input )
tf.data.Dataset for input provided via a numpy array.
This avoids adding
numpy_input as a large constant in the graph,
and copies the data to the machine or machines that will be processing
Note that you will likely need to use
with the returned dataset to further distribute it with the strategy.
numpy_input = np.ones(, dtype=np.float32) dataset = strategy.experimental_make_numpy_dataset(numpy_input) dist_dataset = strategy.experimental_distribute_dataset(dataset)
numpy_input: A nest of NumPy input arrays that will be converted into a dataset. Note that lists of Numpy arrays are stacked, as that is normal
experimental_run_v2( fn, args=(), kwargs=None )
fn on each replica, with the given arguments.
Executes ops specified by
fn on each replica. If
"per-replica" values, such as those produced by a "distributed
fn is executed on a particular replica, it will be executed with the
component of those "per-replica" values that correspond to that replica.
fn may call
tf.distribute.get_replica_context() to access members such
All arguments in
kwargs should either be nest of tensors or
per-replica objects containing tensors or composite tensors.
IMPORTANT: Depending on the implementation of
whether eager execution is enabled,
fn may be called one or more times (
once for each replica).
fn: The function to run. The output must be a
args: (Optional) Positional arguments to
kwargs: (Optional) Keyword arguments to
Merged return value of
fn across replicas. The structure of the return
value is the same as the return value from
fn. Each element in the
structure can either be "per-replica"
Tensor objects or
(for example, if running on a single replica).
reduce( reduce_op, value, axis )
value across replicas.
Given a per-replica value returned by
experimental_run_v2, say a
per-example loss, the batch will be divided across all the replicas. This
function allows you to aggregate across replicas and optionally also across
batch elements. For example, if you have a global batch size of 8 and 2
replicas, values for examples
[0, 1, 2, 3] will be on replica 0 and
[4, 5, 6, 7] will be on replica 1. By default,
reduce will just
aggregate across replicas, returning
[0+4, 1+5, 2+6, 3+7]. This is useful
when each replica is computing a scalar or some other value that doesn't
have a "batch" dimension (like a gradient). More often you will want to
aggregate across the global batch, which you can get by specifying the batch
dimension as the
axis=0. In this case it would return a
If there is a last partial batch, you will need to specify an axis so
that the resulting shape is consistent across replicas. So if the last
batch has size 6 and it is divided into [0, 1, 2, 3] and [4, 5], you
would get a shape mismatch unless you specify
axis=0. If you specify
axis=0 will use the correct
denominator of 6. Contrast this with computing
reduce_mean to get a
scalar value on each replica and this function to average those means,
which will weigh some values
1/8 and others
tf.distribute.ReduceOpvalue specifying how values should be combined.
value: A "per replica" value, e.g. returned by
experimental_run_v2to be combined into a single tensor.
axis: Specifies the dimension to reduce along within each replica's tensor. Should typically be set to the batch dimension, or
Noneto only reduce across replicas (e.g. if the tensor has no batch dimension).
Returns a context manager selecting this Strategy as current.
with strategy.scope(): code block, this thread
will use a variable creator set by
strategy, and will
enter its "cross-replica context".
MultiWorkerMirroredStrategy, all variables created inside
`strategy.scope() will be mirrored on all replicas of each worker.
Moreover, it also sets a default device scope so that ops without
specified devices will end up on the correct worker.
A context manager to use for creating variables with this strategy.