Google I/O is a wrap! Catch up on TensorFlow sessions View sessions


View source on GitHub

A class wrapping information needed by a distribute function.

This is a context class that is passed to the value_fn in strategy.experimental_distribute_values_from_function and contains information about the compute replicas. The num_replicas_in_sync and replica_id can be used to customize the value on each replica.

Example usage:

  1. Directly constructed.
def value_fn(context):
  return context.replica_id_in_sync_group/context.num_replicas_in_sync
context = tf.distribute.experimental.ValueContext(
  replica_id_in_sync_group=2, num_replicas_in_sync=4)
per_replica_value = value_fn(context)
  1. Passed in by experimental_distribute_values_from_function.
strategy = tf.distribute.MirroredStrategy()
def value_fn(value_context):
  return value_context.num_replicas_in_sync
distributed_values = (
local_result = strategy.experimental_local_results(distributed_values)

replica_id_in_sync_group the current replica_id, should be an int in [0,num_replicas_in_sync).
num_replicas_in_sync the number of replicas that are in sync.

num_replicas_in_sync Returns the number of compute replicas in sync.
replica_id_in_sync_group Returns the replica ID.