A parameter server DistributionStrategy.
This strategy class works for both local training and between-graph replicated
training for multiple workers. If
cluster_spec is specified, either passed
in to init() method or parsed from the
variables and updates to those variables are assigned to parameter servers and
other operations are assigned to workers. If
cluster_spec is not set, it
becomes local training where variables are assigned to local CPU or the only
GPU. When each worker has more than one GPU, operations will be replicated on
these GPUs. In both cases, operations are replicated but variables are not and
these workers share a common view for which paramater server a variable is
This class assumes between-graph replication will be used and works on a graph for a particular worker. Note that each graph and worker is independent. This means that while each worker will synchronously compute a single gradient update across all GPUs, updates between workers proceed asynchronously. Operations that occur only on the first replica (such as incrementing the global step), will occur on the first replica of every worker.
It is expected to call
call_for_each_replica(fn, ...) for any
operations which potentially can be replicated across replicas (i.e. multiple
GPUs) even if there is only CPU or one GPU. When defining the
caution needs to be taken:
2) It is generally not recommended to open a device scope under the strategy's
scope. A device scope (i.e. calling
tf.device) will be merged with or
override the device for operations but will not change the device for
3) It is also not recommended to open a colocation scope (i.e. calling
tf.colocate_with) under the strategy's scope. For colocating variables,
distribution.colocate_vars_with instead. Colocation of ops will possibly
create conflicts of device assignment.
Initializes this strategy.
num_gpus_per_worker: number of local GPUs or GPUs per worker, the default is 0 meaning CPU only.
cluster_specis given but
tf.distribute.StrategyExtended with additional methods.
Returns number of replicas over which gradients are aggregated.
Any final actions to be done at the end of all computations.
In eager mode, it executes any finalize actions as a side effect. In graph mode, it creates the finalize ops and returns them.
For example, TPU shutdown ops.
A list of ops to execute.
Any initialization to be done before running any computations.
In eager mode, it executes any initialization as a side effect. In graph mode, it creates the initialization ops and returns them.
For example, TPU initialize_system ops.
A list of ops to execute.
experimental_run( fn, input_iterator=None )
Runs ops in
fn on each replica, with inputs from
When eager execution is enabled, executes ops specified by
fn on each
replica. Otherwise, builds a graph to execute the ops on each replica.
Each replica will take a single, different input from the inputs provided by
get_next call on the input iterator.
fn may call
tf.distribute.get_replica_context() to access members such
IMPORTANT: Depending on the
DistributionStrategy being used, and whether
eager execution is enabled,
fn may be called one or more times (once for
fn: function to run. The inputs to the function must match the outputs of
input_iterator.get_next(). The output must be a
input_iterator: (Optional) input iterator from which the inputs are taken.
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
PerReplica (if the values are unsynchronized),
Mirrored (if the values are kept in sync), or
Tensor (if running on a
Makes an iterator for input provided via input_dataset.
Data from the given dataset will be distributed evenly across all the
compute replicas. 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).
If this effort fails, an error will be thrown, and the user should instead
make_input_fn_iterator which provides more control to the user, and
does not try to divide a batch across replicas.
The user could also use
make_input_fn_iterator if they want to
customize which input is fed to which replica/worker etc.
tf.data.Datasetthat will be distributed evenly across all replicas.
tf.distribute.InputIterator which returns inputs for each step of the
computation. User should call
initialize on the returned iterator.
make_input_fn_iterator( input_fn, replication_mode=tf.distribute.InputReplicationMode.PER_WORKER )
Returns an iterator split across replicas created from an input function.
input_fn should take an
tf.distribute.InputContext object where
information about input sharding can be accessed:
def input_fn(input_context): d = tf.data.Dataset.from_tensors([[1.]]).repeat() return d.shard(input_context.num_input_pipelines, input_context.input_pipeline_id) with strategy.scope(): iterator = strategy.make_input_fn_iterator( input_fn) replica_results = strategy.extended.call_for_each_replica( replica_fn, iterator.get_next())
input_fn: A function that returns a
tf.data.Dataset. This function is expected to take an
replication_mode: an enum value of
PER_WORKERis supported currently.
An iterator object that can be initialized and fetched next element.
reduce( reduce_op, value )
value across replicas.
tf.distribute.ReduceOpvalue specifying how values should be combined.
value: A "per replica" value to be combined into a single tensor.
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".
A context manager.
Returns a copy of
config_proto modified for use with this strategy.
The updated config has something needed to run a strategy, e.g. configuration to run collective ops, or device filters to improve distributed training performance.
The updated copy of the