tf.distribute.DistributedValues

Base class for representing distributed values.

A subclass instance of tf.distribute.DistributedValues is created when creating variables within a distribution strategy, iterating a tf.distribute.DistributedDataset or through tf.distribute.Strategy.run. This base class should never be instantiated directly. tf.distribute.DistributedValues contains a value per replica. Depending on the subclass, the values could either be synced on update, synced on demand, or never synced.

tf.distribute.DistributedValues can be reduced to obtain single value across replicas, as input into tf.distribute.Strategy.run or the per-replica values inspected using tf.distribute.Strategy.experimental_local_results.

Example usage:

  1. Created from a tf.distribute.DistributedDataset:
strategy = tf.distribute.MirroredStrategy()
dataset = tf.data.Dataset.from_tensor_slices([5., 6., 7., 8.]).batch(2)
dataset_iterator = iter(strategy.experimental_distribute_dataset(dataset))
distributed_values = next(dataset_iterator)
  1. Returned by run:
strategy = tf.distribute.MirroredStrategy()
@tf.function
def run():
  ctx = tf.distribute.get_replica_context()
  return ctx.replica_id_in_sync_group
distributed_values = strategy.run(run)
  1. As input into run:
strategy = tf.distribute.MirroredStrategy()
dataset = tf.data.Dataset.from_tensor_slices([5., 6., 7., 8.]).batch(2)
dataset_iterator = iter(strategy.experimental_distribute_dataset(dataset))
distributed_values = next(dataset_iterator)
@tf.function
def run(input):
  return input + 1.0
updated_value = strategy.run(run, args=(distributed_values,))
  1. Reduce value:
strategy = tf.distribute.MirroredStrategy(