def _experimental_distribute_dataset(self, dataset, options):
return input_util.get_distributed_dataset(
dataset,
self._input_workers_with_options(options),
self._container_strategy(),
num_replicas_in_sync=self._num_replicas_in_sync,
options=options)
def _experimental_distribute_dataset(self, dataset, options):
if (options and options.experimental_replication_mode
== distribute_lib.InputReplicationMode.PER_REPLICA):
raise NotImplementedError("InputReplicationMode.PER_REPLICA "
"is only supported in "
"`distribute_datasets_from_function`.")
return input_util.get_distributed_dataset(
dataset,
self._input_workers_with_options(options),
self._container_strategy(),
num_replicas_in_sync=self._num_replicas_in_sync,
options=options)
def _experimental_distribute_dataset(self, dataset, options):
input_workers_devices = self._input_workers_with_options()
# If this DistributedDataset is created outside ClusterCoordinator, i,e,
# outside a tf.function, we don't build its underlying datasets immediately
# until it is passed to ClusterCoordinator.create_per_worker_dataset.
return input_util.get_distributed_dataset(
dataset,
input_workers_devices,
self._container_strategy(),
num_replicas_in_sync=self._num_replicas_in_sync,
options=options,
build=ops.inside_function()) # will be built by ClusterCoordinator
def _experimental_distribute_dataset(self, dataset, options):
# Note that split_batch_by argument is not passed because it is always 1 in
# this strategy, and adding it adds unnecessary overhead to the dataset.
if (options and options.experimental_replication_mode
== distribute_lib.InputReplicationMode.PER_REPLICA):
raise NotImplementedError(
"InputReplicationMode.PER_REPLICA "
"is only supported in "
"`experimental_distribute_datasets_from_function`.")
return input_util.get_distributed_dataset(
dataset,
self._input_workers_with_options(options),
self._container_strategy(),
options=options)
