tf.distribute.InputReplicationMode

Replication mode for input function.

  • PER_WORKER: The input function will be called on each worker independently, creating as many input pipelines as number of workers. Replicas will dequeue from the local Dataset on their worker. tf.distribute.Strategy doesn't manage any state sharing between such separate input pipelines.
  • PER_REPLICA: The input function will be called on each replica seperately. tf.distribute.Strategy doesn't manage any state sharing between such separate input pipelines.

PER_REPLICA tf.distribute.InputReplicationMode
PER_WORKER tf.distribute.InputReplicationMode