xref: /external/tensorflow/tensorflow/python/distribute/input_lib.py

# Copyright 2018 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""Various classes representing distributed inputs."""

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import functools
import sys

import six

from tensorflow.python import tf2
from tensorflow.python.data.experimental.ops import batching
from tensorflow.python.data.experimental.ops import cardinality
from tensorflow.python.data.experimental.ops import distribute
from tensorflow.python.data.ops import dataset_ops
from tensorflow.python.data.ops import iterator_ops
from tensorflow.python.data.ops import multi_device_iterator_ops
from tensorflow.python.data.ops import optional_ops
from tensorflow.python.distribute import device_util
from tensorflow.python.distribute import distribute_utils
from tensorflow.python.distribute import distribution_strategy_context
from tensorflow.python.distribute import input_ops
from tensorflow.python.distribute import reduce_util
from tensorflow.python.distribute import values
from tensorflow.python.distribute.distribute_lib import InputReplicationMode
from tensorflow.python.eager import context
from tensorflow.python.framework import composite_tensor
from tensorflow.python.framework import constant_op
from tensorflow.python.framework import device as tf_device
from tensorflow.python.framework import dtypes
from tensorflow.python.framework import errors
from tensorflow.python.framework import ops
from tensorflow.python.framework import sparse_tensor
from tensorflow.python.framework import tensor_shape
from tensorflow.python.framework import tensor_util
from tensorflow.python.framework import type_spec
from tensorflow.python.ops import array_ops
from tensorflow.python.ops import control_flow_ops
from tensorflow.python.ops import math_ops
from tensorflow.python.ops.ragged import ragged_tensor
from tensorflow.python.types import distribute as distribute_types
from tensorflow.python.util import nest
from tensorflow.python.util.compat import collections_abc
from tensorflow.python.util.deprecation import deprecated
from tensorflow.python.util.tf_export import tf_export
from tensorflow.tools.docs import doc_controls


def get_distributed_dataset(dataset,
                            input_workers,
                            strategy,
                            num_replicas_in_sync=None,
                            input_context=None):
  """Returns a distributed dataset from the given tf.data.Dataset instance.

  This is a common function that is used by all strategies to return a
  distributed dataset. The distributed dataset instance returned is different
  depending on if we are in a TF 1 or TF 2 context. The distributed dataset
  instances returned differ from each other in the APIs supported by each of
  them.

  Args:
    dataset: a tf.data.Dataset instance.
    input_workers: an InputWorkers object which specifies devices on which
        iterators should be created.
    strategy: a `tf.distribute.Strategy` object, used to run all-reduce to
        handle last partial batch.
    num_replicas_in_sync: Optional integer. If this is not None, the value is
        used to decide how to rebatch datasets into smaller batches so that
        the total batch size for each step (across all workers and replicas)
        adds up to `dataset`'s batch size.
    input_context: `InputContext` for sharding. Only pass this in for between
        graph multi-worker cases where there is only one `input_worker`. In
        these cases, we will shard based on the `input_pipeline_id` and
        `num_input_pipelines` in the `InputContext`.

  Returns:
    A distributed dataset instance.
  """
  if tf2.enabled():
    return DistributedDataset(
        dataset,
        input_workers,
        strategy,
        num_replicas_in_sync=num_replicas_in_sync,
        input_context=input_context)
  else:
    return DistributedDatasetV1(
        dataset,
        input_workers,
        strategy,
        num_replicas_in_sync=num_replicas_in_sync,
        input_context=input_context)


def get_distributed_datasets_from_function(dataset_fn,
                                           input_workers,
                                           input_contexts,
                                           strategy,
                                           options=None):
  """Returns a distributed dataset from the given input function.

  This is a common function that is used by all strategies to return a
  distributed dataset. The distributed dataset instance returned is different
  depending on if we are in a TF 1 or TF 2 context. The distributed dataset
  instances returned differ from each other in the APIs supported by each of
  them.

  Args:
    dataset_fn: a function that returns a tf.data.Dataset instance.
    input_workers: an InputWorkers object which specifies devices on which
        iterators should be created.
    input_contexts: A list of `InputContext` instances to be passed to call(s)
        to `dataset_fn`. Length and order should match worker order in
        `worker_device_pairs`.
    strategy: a `tf.distribute.Strategy` object, used to run all-reduce to
        handle last partial batch.
    options: Default is None. `tf.distribute.InputOptions` used to control
        options on how this dataset is distributed.

  Returns:
    A distributed dataset instance.

  Raises:
    ValueError: if `options.experimental_replication_mode` and
    `options.experimental_place_dataset_on_device` are not consistent
  """
  if (options is not None and
      options.experimental_replication_mode != InputReplicationMode.PER_REPLICA
      and options.experimental_place_dataset_on_device):
    raise ValueError(
        "When `experimental_place_dataset_on_device` is set for dataset "
        "placement, you must also specify `PER_REPLICA` for the "
        "replication mode")

  if (options is not None and
      options.experimental_replication_mode == InputReplicationMode.PER_REPLICA
      and options.experimental_prefetch_to_device and
      options.experimental_place_dataset_on_device):
    raise ValueError(
        "`experimental_place_dataset_on_device` can not be set to True "
        "when experimental_prefetch_to_device is True and "
        "replication mode is set to `PER_REPLICA`")

  if tf2.enabled():
    return DistributedDatasetsFromFunction(dataset_fn, input_workers,
                                           input_contexts, strategy, options)
  else:
    return DistributedDatasetsFromFunctionV1(
        dataset_fn,
        input_workers,
        input_contexts,
        strategy,
        options)


@tf_export("distribute.DistributedIterator", v1=[])
class DistributedIteratorInterface(collections_abc.Iterator,
                                   distribute_types.Iterator):
  """An iterator over `tf.distribute.DistributedDataset`.

  `tf.distribute.DistributedIterator` is the primary mechanism for enumerating
  elements of a `tf.distribute.DistributedDataset`. It supports the Python
  Iterator protocol, which means it can be iterated over using a for-loop or by
  fetching individual elements explicitly via `get_next()`.

  You can create a `tf.distribute.DistributedIterator` by calling `iter` on
  a `tf.distribute.DistributedDataset` or creating a python loop over a
  `tf.distribute.DistributedDataset`.

  Visit the [tutorial](https://www.tensorflow.org/tutorials/distribute/input)
  on distributed input for more examples and caveats.
  """

  def get_next(self):
    """Returns the next input from the iterator for all replicas.

    Example use:

    >>> strategy = tf.distribute.MirroredStrategy(["GPU:0", "GPU:1"])
    >>> dataset = tf.data.Dataset.range(100).batch(2)
    >>> dist_dataset = strategy.experimental_distribute_dataset(dataset)
    >>> dist_dataset_iterator = iter(dist_dataset)
    >>> @tf.function
    ... def one_step(input):
    ...   return input
    >>> step_num = 5
    >>> for _ in range(step_num):
    ...   strategy.run(one_step, args=(dist_dataset_iterator.get_next(),))
    >>> strategy.experimental_local_results(dist_dataset_iterator.get_next())
    (<tf.Tensor: shape=(1,), dtype=int64, numpy=array([10])>,
     <tf.Tensor: shape=(1,), dtype=int64, numpy=array([11])>)

    Returns:
      A single `tf.Tensor` or a `tf.distribute.DistributedValues` which contains
      the next input for all replicas.

    Raises:
      `tf.errors.OutOfRangeError`: If the end of the iterator has been reached.
    """
    raise NotImplementedError(
        "DistributedIterator.get_next() must be implemented in descendants.")

  @property
  def element_spec(self):
    # pylint: disable=line-too-long
    """The type specification of an element of `tf.distribute.DistributedIterator`.

    Example usage:

    >>> global_batch_size = 16
    >>> strategy = tf.distribute.MirroredStrategy(["GPU:0", "GPU:1"])
    >>> dataset = tf.data.Dataset.from_tensors(([1.],[2])).repeat(100).batch(global_batch_size)
    >>> distributed_iterator = iter(strategy.experimental_distribute_dataset(dataset))
    >>> distributed_iterator.element_spec
    (PerReplicaSpec(TensorSpec(shape=(None, 1), dtype=tf.float32, name=None),
                    TensorSpec(shape=(None, 1), dtype=tf.float32, name=None)),
     PerReplicaSpec(TensorSpec(shape=(None, 1), dtype=tf.int32, name=None),
                    TensorSpec(shape=(None, 1), dtype=tf.int32, name=None)))

    Returns:
      A nested structure of `tf.TypeSpec` objects matching the structure of an
      element of this `tf.distribute.DistributedIterator`. This returned value
      is typically a `tf.distribute.DistributedValues` object and specifies the
      `tf.TensorSpec` of individual components.
    """
    raise NotImplementedError(
        "DistributedIterator.element_spec() must be implemented in descendants")

  def get_next_as_optional(self):
    # pylint: disable=line-too-long
    """Returns a `tf.experimental.Optional` that contains the next value for all replicas.

    If the `tf.distribute.DistributedIterator` has reached the end of the
    sequence, the returned `tf.experimental.Optional` will have no value.

    Example usage:

    >>> strategy = tf.distribute.MirroredStrategy(["GPU:0", "GPU:1"])
    >>> global_batch_size = 2
    >>> steps_per_loop = 2
    >>> dataset = tf.data.Dataset.range(10).batch(global_batch_size)
    >>> distributed_iterator = iter(
    ...     strategy.experimental_distribute_dataset(dataset))
    >>> def step_fn(x):
    ...   # train the model with inputs
    ...   return x
    >>> @tf.function
    ... def train_fn(distributed_iterator):
    ...   for _ in tf.range(steps_per_loop):
    ...     optional_data = distributed_iterator.get_next_as_optional()
    ...     if not optional_data.has_value():
    ...       break
    ...     per_replica_results = strategy.run(step_fn, args=(optional_data.get_value(),))
    ...     tf.print(strategy.experimental_local_results(per_replica_results))
    >>> train_fn(distributed_iterator)
    ... # ([0 1], [2 3])
    ... # ([4], [])

    Returns:
      An `tf.experimental.Optional` object representing the next value from the
      `tf.distribute.DistributedIterator` (if it has one) or no value.
    """
    # pylint: enable=line-too-long
    raise NotImplementedError(
        "get_next_as_optional() not implemented in descendants")


@tf_export("distribute.DistributedDataset", v1=[])
class DistributedDatasetInterface(collections_abc.Iterable,
                                  distribute_types.Iterable):
  # pylint: disable=line-too-long
  """Represents a dataset distributed among devices and machines.

  A `tf.distribute.DistributedDataset` could be thought of as a "distributed"
  dataset. When you use `tf.distribute` API to scale training to multiple
  devices or machines, you also need to distribute the input data, which leads
  to a `tf.distribute.DistributedDataset` instance, instead of a
  `tf.data.Dataset` instance in the non-distributed case. In TF 2.x,
  `tf.distribute.DistributedDataset` objects are Python iterables.

  Note: `tf.distribute.DistributedDataset` instances are *not* of type
  `tf.data.Dataset`. It only supports two usages we will mention below:
  iteration and `element_spec`. We don't support any other APIs to transform or
  inspect the dataset.

  There are two APIs to create a `tf.distribute.DistributedDataset` object:
  `tf.distribute.Strategy.experimental_distribute_dataset(dataset)`and
  `tf.distribute.Strategy.distribute_datasets_from_function(dataset_fn)`.
  *When to use which?* When you have a `tf.data.Dataset` instance, and the
  regular batch splitting (i.e. re-batch the input `tf.data.Dataset` instance
  with a new batch size that is equal to the global batch size divided by the
  number of replicas in sync) and autosharding (i.e. the
  `tf.data.experimental.AutoShardPolicy` options) work for you, use the former
  API. Otherwise, if you are *not* using a canonical `tf.data.Dataset` instance,
  or you would like to customize the batch splitting or sharding, you can wrap
  these logic in a `dataset_fn` and use the latter API. Both API handles
  prefetch to device for the user. For more details and examples, follow the
  links to the APIs.


  There are two main usages of a `DistributedDataset` object:

  1. Iterate over it to generate the input for a single device or multiple
  devices, which is a `tf.distribute.DistributedValues` instance. To do this,
  you can:

    * use a pythonic for-loop construct:

      >>> global_batch_size = 4
      >>> strategy = tf.distribute.MirroredStrategy(["GPU:0", "GPU:1"])
      >>> dataset = tf.data.Dataset.from_tensors(([1.],[1.])).repeat(4).batch(global_batch_size)
      >>> dist_dataset = strategy.experimental_distribute_dataset(dataset)
      >>> @tf.function
      ... def train_step(input):
      ...   features, labels = input
      ...   return labels - 0.3 * features
      >>> for x in dist_dataset:
      ...   # train_step trains the model using the dataset elements
      ...   loss = strategy.run(train_step, args=(x,))
      ...   print("Loss is", loss)
      Loss is PerReplica:{
        0: tf.Tensor(
      [[0.7]
       [0.7]], shape=(2, 1), dtype=float32),
        1: tf.Tensor(
      [[0.7]
       [0.7]], shape=(2, 1), dtype=float32)
      }

      Placing the loop inside a `tf.function` will give a performance boost.
      However `break` and `return` are currently not supported if the loop is
      placed inside a `tf.function`. We also don't support placing the loop
      inside a `tf.function` when using
      `tf.distribute.experimental.MultiWorkerMirroredStrategy` or
      `tf.distribute.experimental.TPUStrategy` with multiple workers.

    * use `__iter__` to create an explicit iterator, which is of type
      `tf.distribute.DistributedIterator`

      >>> global_batch_size = 4
      >>> strategy = tf.distribute.MirroredStrategy(["GPU:0", "GPU:1"])
      >>> train_dataset = tf.data.Dataset.from_tensors(([1.],[1.])).repeat(50).batch(global_batch_size)
      >>> train_dist_dataset = strategy.experimental_distribute_dataset(train_dataset)
      >>> @tf.function
      ... def distributed_train_step(dataset_inputs):
      ...   def train_step(input):
      ...     loss = tf.constant(0.1)
      ...     return loss
      ...   per_replica_losses = strategy.run(train_step, args=(dataset_inputs,))
      ...   return strategy.reduce(tf.distribute.ReduceOp.SUM, per_replica_losses,axis=None)
      >>> EPOCHS = 2
      >>> STEPS = 3
      >>> for epoch in range(EPOCHS):
      ...   total_loss = 0.0
      ...   num_batches = 0
      ...   dist_dataset_iterator = iter(train_dist_dataset)
      ...   for _ in range(STEPS):
      ...     total_loss += distributed_train_step(next(dist_dataset_iterator))
      ...     num_batches += 1
      ...   average_train_loss = total_loss / num_batches
      ...   template = ("Epoch {}, Loss: {:.4f}")
      ...   print (template.format(epoch+1, average_train_loss))
      Epoch 1, Loss: 0.2000
      Epoch 2, Loss: 0.2000


    To achieve a performance improvement, you can also wrap the `strategy.run`
    call with a `tf.range` inside a `tf.function`. This runs multiple steps in a
    `tf.function`. Autograph will convert it to a `tf.while_loop` on the worker.
    However, it is less flexible comparing with running a single step inside
    `tf.function`. For example, you cannot run things eagerly or arbitrary
    python code within the steps.


  2. Inspect the `tf.TypeSpec` of the data generated by `DistributedDataset`.

    `tf.distribute.DistributedDataset` generates
    `tf.distribute.DistributedValues` as input to the devices. If you pass the
    input to a `tf.function` and would like to specify the shape and type of
    each Tensor argument to the function, you can pass a `tf.TypeSpec` object to
    the `input_signature` argument of the `tf.function`. To get the
    `tf.TypeSpec` of the input, you can use the `element_spec` property of the
    `tf.distribute.DistributedDataset` or `tf.distribute.DistributedIterator`
    object.

    For example:

    >>> global_batch_size = 4
    >>> epochs = 1
    >>> steps_per_epoch = 1
    >>> mirrored_strategy = tf.distribute.MirroredStrategy(["GPU:0", "GPU:1"])
    >>> dataset = tf.data.Dataset.from_tensors(([2.])).repeat(100).batch(global_batch_size)
    >>> dist_dataset = mirrored_strategy.experimental_distribute_dataset(dataset)
    >>> @tf.function(input_signature=[dist_dataset.element_spec])
    ... def train_step(per_replica_inputs):
    ...   def step_fn(inputs):
    ...     return tf.square(inputs)
    ...   return mirrored_strategy.run(step_fn, args=(per_replica_inputs,))
    >>> for _ in range(epochs):
    ...   iterator = iter(dist_dataset)
    ...   for _ in range(steps_per_epoch):
    ...     output = train_step(next(iterator))
    ...     print(output)
    PerReplica:{
      0: tf.Tensor(
    [[4.]
     [4.]], shape=(2, 1), dtype=float32),
      1: tf.Tensor(
    [[4.]
     [4.]], shape=(2, 1), dtype=float32)
    }


  Visit the [tutorial](https://www.tensorflow.org/tutorials/distribute/input)
  on distributed input for more examples and caveats.
  """

  def __iter__(self):
    """Creates an iterator for the `tf.distribute.DistributedDataset`.

    The returned iterator implements the Python Iterator protocol.

    Example usage:

    >>> global_batch_size = 4
    >>> strategy = tf.distribute.MirroredStrategy(["GPU:0", "GPU:1"])
    >>> dataset = tf.data.Dataset.from_tensor_slices([1, 2, 3, 4]).repeat().batch(global_batch_size)
    >>> distributed_iterator = iter(strategy.experimental_distribute_dataset(dataset))
    >>> print(next(distributed_iterator))
    PerReplica:{
      0: tf.Tensor([1 2], shape=(2,), dtype=int32),
      1: tf.Tensor([3 4], shape=(2,), dtype=int32)
    }

    Returns:
      An `tf.distribute.DistributedIterator` instance for the given
      `tf.distribute.DistributedDataset` object to enumerate over the
      distributed data.
    """
    raise NotImplementedError("Must be implemented in descendants")

  @property
  def element_spec(self):
    """The type specification of an element of this `tf.distribute.DistributedDataset`.

    Example usage:

    >>> global_batch_size = 16
    >>> strategy = tf.distribute.MirroredStrategy(["GPU:0", "GPU:1"])
    >>> dataset = tf.data.Dataset.from_tensors(([1.],[2])).repeat(100).batch(global_batch_size)
    >>> dist_dataset = strategy.experimental_distribute_dataset(dataset)
    >>> dist_dataset.element_spec
    (PerReplicaSpec(TensorSpec(shape=(None, 1), dtype=tf.float32, name=None),
                    TensorSpec(shape=(None, 1), dtype=tf.float32, name=None)),
     PerReplicaSpec(TensorSpec(shape=(None, 1), dtype=tf.int32, name=None),
                    TensorSpec(shape=(None, 1), dtype=tf.int32, name=None)))

    Returns:
      A nested structure of `tf.TypeSpec` objects matching the structure of an
      element of this `tf.distribute.DistributedDataset`. This returned value is
      typically a `tf.distribute.DistributedValues` object and specifies the
      `tf.TensorSpec` of individual components.
    """
    raise NotImplementedError(
        "DistributedDataset.element_spec must be implemented in descendants.")

  @doc_controls.do_not_generate_docs
  def reduce(self, initial_state, reduce_func):
    raise NotImplementedError(
        "DistributedDataset.reduce must be implemented in descendants.")


class InputWorkers(object):
  """A 1-to-many mapping from input worker devices to compute devices."""

  def __init__(self, worker_device_pairs):
    """Initialize an `InputWorkers` object.

    Args:
      worker_device_pairs: A sequence of pairs:
        `(input device, a tuple of compute devices fed by that input device)`.
    """
    self._worker_device_pairs = worker_device_pairs
    self._input_worker_devices = tuple(d for d, _ in self._worker_device_pairs)
    self._fed_devices = tuple(tuple(device_util.canonicalize(d) for d in f)
                              for _, f in self._worker_device_pairs)

  @property
  def num_workers(self):
    return len(self._input_worker_devices)

  @property
  def worker_devices(self):
    return self._input_worker_devices

  def compute_devices_for_worker(self, worker_index):
    return self._fed_devices[worker_index]

  def __repr__(self):
    devices = self.worker_devices
    debug_repr = ",\n".join("  %d %s: %s" %
                            (i, devices[i], self._fed_devices[i])
                            for i in range(len(devices)))
    return "%s:{\n%s}" % (self.__class__.__name__, debug_repr)

  def serialize(self):
    return self._worker_device_pairs

  def deserialize(self, worker_device_pairs):
    return InputWorkers(worker_device_pairs)


def _get_next_as_optional(iterator, strategy, return_per_replica=False):
  """Returns an empty dataset indicator and the next input from the iterator.

  Args:
    iterator: a DistributedIterator object.
    strategy: the `tf.distribute.Strategy` instance.
    return_per_replica: a boolean. If True, the returned data will be wrapped
      with `PerReplica` structure. Otherwise it is a 2D
      num_input_workers*num_replicas_per_worker list.

  Returns:
    A tuple (a boolean tensor indicating whether the next batch has value
    globally, data from all replicas).
  """
  replicas = []
  worker_has_values = []
  worker_devices = []
  for i, worker in enumerate(iterator._input_workers.worker_devices):  # pylint: disable=protected-access
    with ops.device(worker):
      worker_has_value, next_element = (
          iterator._iterators[i].get_next_as_list())  # pylint: disable=protected-access
      # Collective all-reduce requires explicit devices for inputs.
      with ops.device("/cpu:0"):
        # Converting to integers for all-reduce.
        worker_has_value = math_ops.cast(worker_has_value, dtypes.int64)
        worker_devices.append(worker_has_value.device)
        worker_has_values.append(worker_has_value)
      # Make `replicas` a flat list of values across all replicas.
      replicas.append(next_element)

  if return_per_replica:
    flattened_data = []
    for per_worker_data in replicas:
      flattened_data.extend(per_worker_data)
    replicas = _create_per_replica(flattened_data, strategy)

  # Run an all-reduce to see whether any worker has values.
  # TODO(b/131423105): we should be able to short-cut the all-reduce in some
  # cases.
  if getattr(strategy.extended, "_support_per_replica_values", True):
    # `reduce` expects a `PerReplica`, so we pass it one, even
    # though it doesn't actually have a value per replica
    worker_has_values = values.PerReplica(worker_has_values)
    global_has_value = strategy.reduce(
        reduce_util.ReduceOp.SUM, worker_has_values, axis=None)
  else:
    assert len(worker_has_values) == 1
    global_has_value = worker_has_values[0]
  global_has_value = array_ops.reshape(
      math_ops.cast(global_has_value, dtypes.bool), [])
  return global_has_value, replicas


def _is_statically_shaped(element_spec):
  """Test if an iterator output is statically shaped.

  For sparse and ragged tensors this only tests the batch dimension.

  Args:
    element_spec: a nest structure of `tf.TypeSpec`. The element spec of the
      dataset of the iterator.

  Returns:
    True if the shape is static, false otherwise.
  """

  for spec in nest.flatten(element_spec):
    if isinstance(
        spec, (sparse_tensor.SparseTensorSpec, ragged_tensor.RaggedTensorSpec)):
      # For sparse or ragged tensor, we should only check the first
      # dimension in order to get_next_as_optional. This is because
      # when these tensors get batched by dataset only the batch dimension
      # is set.
      if spec.shape.rank > 0 and spec.shape.as_list()[0] is None:
        return False
    else:
      for component in nest.flatten(spec._component_specs):  # pylint: disable=protected-access
        if not component.shape.is_fully_defined():
          return False
  return True


class DistributedIteratorBase(DistributedIteratorInterface):
  """Common implementation for all input iterators."""

  # pylint: disable=super-init-not-called
  def __init__(self, input_workers, iterators, strategy,
               enable_get_next_as_optional):
    assert isinstance(input_workers, InputWorkers)
    if not input_workers.worker_devices:
      raise ValueError("Should have at least one worker for input iterator.")

    self._iterators = iterators
    self._input_workers = input_workers
    self._strategy = strategy
    self._enable_get_next_as_optional = enable_get_next_as_optional

  def next(self):
    return self.__next__()

  def __next__(self):
    try:
      return self.get_next()
    except errors.OutOfRangeError:
      raise StopIteration

  def __iter__(self):
    return self

  def get_next_as_optional(self):
    global_has_value, replicas = _get_next_as_optional(
        self, self._strategy, return_per_replica=True)

    def return_none():
      return optional_ops.Optional.empty(self._element_spec)

    return control_flow_ops.cond(
        global_has_value, lambda: optional_ops.Optional.from_value(replicas),
        return_none)

  def get_next(self, name=None):
    """Returns the next input from the iterator for all replicas."""
    if not self._enable_get_next_as_optional:
      replicas = []
      for i, worker in enumerate(self._input_workers.worker_devices):
        if name is not None:
          d = tf_device.DeviceSpec.from_string(worker)
          new_name = "%s_%s_%d" % (name, d.job, d.task)
        else:
          new_name = None
        with ops.device(worker):
          # Make `replicas` a flat list of values across all replicas.
          replicas.extend(
              self._iterators[i].get_next_as_list_static_shapes(new_name))
      return _create_per_replica(replicas, self._strategy)

    out_of_range_replicas = []
    def out_of_range_fn(worker_index, device):
      """This function will throw an OutOfRange error."""
      # As this will be only called when there is no data left, so calling
      # get_next() will trigger an OutOfRange error.
      data = self._iterators[worker_index].get_next(device)
      out_of_range_replicas.append(data)
      return data

    global_has_value, replicas = _get_next_as_optional(
        self, self._strategy, return_per_replica=False)
    results = []
    for i, worker in enumerate(self._input_workers.worker_devices):
      with ops.device(worker):
        devices = self._input_workers.compute_devices_for_worker(i)
        for j, device in enumerate(devices):
          with ops.device(device):
            # pylint: disable=undefined-loop-variable
            # pylint: disable=cell-var-from-loop
            # It is fine for the lambda to capture variables from the loop as
            # the lambda is executed in the loop as well.
            result = control_flow_ops.cond(
                global_has_value,
                lambda: replicas[i][j],
                lambda: out_of_range_fn(i, device),
                strict=True,
            )
            # pylint: enable=cell-var-from-loop
            # pylint: enable=undefined-loop-variable
            results.append(result)
    replicas = results

    return _create_per_replica(replicas, self._strategy)


class DistributedIteratorV1(DistributedIteratorBase):
  """Input Iterator for a distributed dataset."""

  # We need a private initializer method for re-initializing multidevice
  # iterators when used with Keras training loops. If we don't reinitialize the
  # iterator we run into memory leak issues (b/123315763).
  @property
  def _initializer(self):
    init_ops = []
    for it in self._iterators:
      init_ops.extend(it.initialize())
    return control_flow_ops.group(init_ops)

  @deprecated(None, "Use the iterator's `initializer` property instead.")
  def initialize(self):
    """Initialize underlying iterators.

    Returns:
      A list of any initializer ops that should be run.
    """
    return self._initializer

  @property
  def initializer(self):
    """Returns a list of ops that initialize the iterator."""
    return self.initialize()

  # TODO(priyag): Remove when we switch to using `MultiDeviceIterator` for TPUs.
  @property
  def output_classes(self):
    return self._iterators[0].output_classes

  # TODO(priyag): Remove when we switch to using `MultiDeviceIterator` for TPUs.
  @property
  def output_shapes(self):
    return self._iterators[0].output_shapes

  # TODO(priyag): Remove when we switch to using `MultiDeviceIterator` for TPUs.
  @property
  def output_types(self):
    return self._iterators[0].output_types

  # TODO(priyag): Remove when we switch to using `MultiDeviceIterator` for TPUs.
  def get_iterator(self, worker):
    for i, w in enumerate(self._input_workers.worker_devices):
      if worker == w:
        return self._iterators[i]
    return None

  @property
  def element_spec(self):
    """The type specification of an element of this iterator."""
    return self._element_spec


class DistributedIteratorSpec(type_spec.TypeSpec):
  """Type specification for `DistributedIterator`."""

  __slots__ = [
      "_input_workers", "_element_spec", "_strategy",
      "_enable_get_next_as_optional", "_options"
  ]

  def __init__(self, input_workers, element_spec, strategy,
               enable_get_next_as_optional, options):
    # We don't want to allow deserialization of this class because we don't
    # serialize the strategy object. Currently the only places where
    # _deserialize is called is when we save/restore using SavedModels.
    if isinstance(input_workers, tuple):
      raise NotImplementedError("DistributedIteratorSpec does not have support "
                                "for deserialization.")
    else:
      self._input_workers = input_workers
      self._element_spec = element_spec
      self._strategy = strategy
      self._enable_get_next_as_optional = enable_get_next_as_optional
      self._options = options

  @property
  def value_type(self):
    return DistributedIterator

  def _serialize(self):
    # We cannot serialize the strategy object so we convert it to an id that we
    # can use for comparison.
    return (self._input_workers.serialize(), self._element_spec,
            id(self._strategy), id(self._options))

  def _deserialize(self):
    raise ValueError("Deserialization is currently unsupported for "
                     "DistributedIteratorSpec.")

  # Overriding this method so that we can merge and reconstruct the spec object
  def most_specific_compatible_type(self, other):
    """Returns the most specific TypeSpec compatible with `self` and `other`.

    Args:
      other: A `TypeSpec`.

    Raises:
      ValueError: If there is no TypeSpec that is compatible with both `self`
        and `other`.
    """
    # pylint: disable=protected-access
    if type(self) is not type(other):
      raise ValueError("No TypeSpec is compatible with both %s and %s" %
                       (self, other))
    if self._input_workers.serialize() != other._input_workers.serialize():
      raise ValueError("_input_workers is not compatible with both %s "
                       "and %s" % (self, other))
    if self._strategy is not other._strategy:
      raise ValueError("tf.distribute strategy is not compatible with both %s "
                       "and %s" % (self, other))
    element_spec = nest.map_structure(
        lambda a, b: a.most_specific_compatible_type(b), self._element_spec,
        other._element_spec)
    return DistributedIteratorSpec(self._input_workers, element_spec,
                                   self._strategy,
                                   self._enable_get_next_as_optional,
                                   self._options)

  @property
  def _component_specs(self):
    specs = []
    worker_device_pairs = self._input_workers._worker_device_pairs  # pylint: disable=protected-access

    for i, (input_device, compute_devices) in enumerate(worker_device_pairs):
      element_spec = nest.map_structure(
          functools.partial(_replace_per_replica_spec, i=i), self._element_spec)
      specs.append(
          _SingleWorkerDatasetIteratorSpec(input_device, compute_devices,
                                           element_spec, self._options))
    return specs

  def _to_components(self, value):
    return value._iterators  # pylint: disable=protected-access

  def _from_components(self, components):
    return DistributedIterator(
        input_workers=self._input_workers,
        iterators=None,
        components=components,
        element_spec=self._element_spec,
        strategy=self._strategy,
        enable_get_next_as_optional=self._enable_get_next_as_optional,
        options=self._options)

  @staticmethod
  def from_value(value):
    # pylint: disable=protected-access
    return DistributedIteratorSpec(value._input_workers, value._element_spec,
                                   value._strategy,
                                   value._enable_get_next_as_optional,
                                   value._options)

  def _with_tensor_ranks_only(self):
    element_spec = nest.map_structure(
        lambda s: s._with_tensor_ranks_only(),  # pylint: disable=protected-access
        self._element_spec)
    return DistributedIteratorSpec(self._input_workers, element_spec,
                                   self._strategy,
                                   self._enable_get_next_as_optional,
                                   self._options)


class DistributedIterator(DistributedIteratorBase,
                          composite_tensor.CompositeTensor):
  """Input Iterator for a distributed dataset."""

  def __init__(self,
               input_workers=None,
               iterators=None,
               strategy=None,
               components=None,
               element_spec=None,
               enable_get_next_as_optional=False,
               options=None):
    if input_workers is None:
      raise ValueError("`input_workers` should be "
                       "provided.")

    error_message = ("Either `input_workers` or "
                     "both `components` and `element_spec` need to be "
                     "provided.")
    self._options = options

    if iterators is None:
      if (components is None or element_spec is None):
        raise ValueError(error_message)
      self._element_spec = element_spec
      self._input_workers = input_workers
      self._iterators = components
      self._strategy = strategy
      self._enable_get_next_as_optional = enable_get_next_as_optional
    else:
      if (components is not None and element_spec is not None):
        raise ValueError(error_message)

      super(DistributedIterator,
            self).__init__(input_workers, iterators, strategy,
                           enable_get_next_as_optional)

  @property
  def element_spec(self):
    # When partial batch handling is enabled, always set the batch dimension to
    # None, otherwise we just follow element_spec of the underlying dataset
    # (whose batch dimension may also be None). This is because with partial
    # batching handling we could always produce empty batches.
    if (self._enable_get_next_as_optional and
        self._strategy.extended._in_multi_worker_mode()):  # pylint: disable=protected-access
      return nest.map_structure(
          _rebatch_as_dynamic, self._element_spec, expand_composites=False)
    return self._element_spec

  @property
  def _type_spec(self):
    # Note that we use actual element_spec instead of the rebatched-as-dynamic
    # one to create DistributedIteratorSpec, to be consistent with the
    # underlying iterators' specs.
    return DistributedIteratorSpec(self._input_workers, self._element_spec,
                                   self._strategy,
                                   self._enable_get_next_as_optional,
                                   self._options)


class _IterableInput(DistributedDatasetInterface):
  """Base class for iterable inputs for distribution strategies."""

  # pylint: disable=super-init-not-called
  def __init__(self, input_workers):
    assert isinstance(input_workers, InputWorkers)
    self._input_workers = input_workers

  def __iter__(self):
    raise NotImplementedError("must be implemented in descendants")

  def reduce(self, initial_state, reduce_fn):
    """Execute a `reduce_fn` over all the elements of the input."""
    iterator = iter(self)
    has_data, data = _get_next_as_optional(
        iterator, self._strategy, return_per_replica=True)

    def cond(has_data, data, state):
      del data, state  # Unused.
      return has_data

    def loop_body(has_data, data, state):
      """Executes `reduce_fn` in a loop till the dataset is empty."""
      del has_data  # Unused.
      state = reduce_fn(state, data)
      has_data, data = _get_next_as_optional(
          iterator, self._strategy, return_per_replica=True)
      return has_data, data, state

    has_data, data, final_state = control_flow_ops.while_loop(
        cond, loop_body, [has_data, data, initial_state], parallel_iterations=1)
    return final_state


class DistributedDataset(_IterableInput):
  """Distributed dataset that supports prefetching to multiple devices."""

  def __init__(self,
               dataset,
               input_workers,
               strategy,
               num_replicas_in_sync=None,
               input_context=None):
    """Distribute the dataset on all workers.

    If `num_replicas_in_sync` is not None, we split each batch of the dataset
    into `num_replicas_in_sync` smaller batches, to be distributed among that
    worker's replicas, so that the batch size for a global step (across all
    workers and replicas) is as expected.

    Args:
      dataset: `tf.data.Dataset` that will be used as the input source.
      input_workers: an `InputWorkers` object.
      strategy: a `tf.distribute.Strategy` object, used to run all-reduce to
        handle last partial batch.
      num_replicas_in_sync: Optional integer. If this is not None, the value
        is used to decide how to rebatch datasets into smaller batches so that
        the total batch size for each step (across all workers and replicas)
        adds up to `dataset`'s batch size.
      input_context: `InputContext` for sharding. Only pass this in for between
        graph multi-worker cases where there is only one `input_worker`. In
        these cases, we will shard based on the `input_pipeline_id` and
        `num_input_pipelines` in the `InputContext`.
    """
    super(DistributedDataset, self).__init__(input_workers=input_workers)
    # We clone and shard the dataset on each worker. The current setup tries to
    # shard the dataset by files if possible so that each worker sees a
    # different subset of files. If that is not possible, will attempt to shard
    # the final input such that each worker will run the entire preprocessing
    # pipeline and only receive its own shard of the dataset.

    # Additionally, we rebatch the dataset on each worker into
    # `num_replicas_in_sync` smaller batches to be distributed among that
    # worker's replicas, so that the batch size for a global step (across all
    # workers and replicas) adds up to the original dataset's batch size.
    if num_replicas_in_sync is not None:
      num_workers = input_context.num_input_pipelines if input_context else len(
          input_workers.worker_devices)
      rebatch_fn = self._make_rebatch_fn(dataset, num_workers,
                                         num_replicas_in_sync)
    else:
      rebatch_fn = None

    self._cloned_datasets = []
    if input_context:
      # Between-graph where we rely on the input_context for sharding
      assert input_workers.num_workers == 1
      if rebatch_fn is not None:
        dataset = rebatch_fn(dataset, input_context.input_pipeline_id)
      dataset = input_ops.auto_shard_dataset(dataset,
                                             input_context.num_input_pipelines,
                                             input_context.input_pipeline_id,
                                             num_replicas_in_sync)
      self._cloned_datasets.append(dataset)
    else:
      replicated_ds = distribute.replicate(dataset,
                                           input_workers.worker_devices)
      for i, worker in enumerate(input_workers.worker_devices):
        with ops.device(worker):
          cloned_dataset = replicated_ds[worker]
          cloned_dataset = cloned_dataset.with_options(dataset.options())
          if rebatch_fn is not None:
            cloned_dataset = rebatch_fn(cloned_dataset, i)
          cloned_dataset = input_ops.auto_shard_dataset(
              cloned_dataset, len(input_workers.worker_devices), i,
              num_replicas_in_sync)
          self._cloned_datasets.append(cloned_dataset)

    self._input_workers = input_workers
    self._strategy = strategy
    self._enable_get_next_as_optional = _enable_get_next_as_optional(
        self._strategy, dataset)
    self._element_spec = _create_distributed_tensor_spec(
        self._strategy, self._cloned_datasets[0].element_spec)

  def _make_rebatch_fn(self, dataset, num_workers, num_replicas_in_sync):
    """Returns a callable that rebatches the input dataset.

    Args:
      dataset: A `tf.data.Dataset` representing the dataset to be distributed.
      num_workers: An integer representing the number of workers to distribute
        `dataset` among.
      num_replicas_in_sync: An integer representing the number of replicas in
        sync across all workers.
    """
    if num_replicas_in_sync % num_workers:
      raise ValueError(
          "tf.distribute expects every worker to have the same number of "
          "replicas. However, encountered `num_replicas_in_sync` ({}) that "
          "cannot be divided by `num_workers` ({})".format(
              num_replicas_in_sync, num_workers))

    num_replicas_per_worker = num_replicas_in_sync // num_workers
    with ops.colocate_with(dataset._variant_tensor):  # pylint: disable=protected-access
      batch_size = distribute.compute_batch_size(dataset)

    def rebatch_fn(dataset, worker_index):
      try:
        # pylint: disable=protected-access
        def apply_rebatch():
          batch_sizes = distribute.batch_sizes_for_worker(
              batch_size, num_workers, num_replicas_per_worker, worker_index)
          return distribute._RebatchDataset(
              dataset, batch_sizes).prefetch(num_replicas_per_worker)

        def apply_legacy_rebatch():
          return distribute._LegacyRebatchDataset(
              dataset, num_replicas_in_sync).prefetch(num_replicas_per_worker)

        with ops.colocate_with(dataset._variant_tensor):
          return control_flow_ops.cond(
              math_ops.not_equal(batch_size, -1),
              true_fn=apply_rebatch,
              false_fn=apply_legacy_rebatch)
      except errors.InvalidArgumentError as e:
        if "without encountering a batch" in str(e):
          six.reraise(
              ValueError,
              ValueError(
                  "Call the `batch` method on the input Dataset in order to be "
                  "able to split your input across {} replicas.\n Please see "
                  "the tf.distribute.Strategy guide. {}".format(
                      num_replicas_in_sync, e)),
              sys.exc_info()[2])
        else:
          raise

    return rebatch_fn

  def __iter__(self):
    if not (context.executing_eagerly() or
            ops.get_default_graph().building_function):
      raise RuntimeError("__iter__() is only supported inside of tf.function "
                         "or when eager execution is enabled.")

    # This is an optional flag that can be used to turn off using
    # OwnedMultiDeviceIterators and instead use the legacy MultiDeviceIterators
    # as a stop gap solution that will allow us to roll out this change.
    enable_legacy_iterators = getattr(self._strategy,
                                      "_enable_legacy_iterators", False)
    worker_iterators = _create_iterators_per_worker(self._cloned_datasets,
                                                    self._input_workers,
                                                    enable_legacy_iterators)
    if enable_legacy_iterators:
      iterator = DistributedIteratorV1(
          self._input_workers,
          worker_iterators,
          self._strategy,
          enable_get_next_as_optional=self._enable_get_next_as_optional)
    else:
      iterator = DistributedIterator(
          self._input_workers,
          worker_iterators,
          self._strategy,
          enable_get_next_as_optional=self._enable_get_next_as_optional)
    iterator._element_spec = self._element_spec  # pylint: disable=protected-access

    # When async eager is enabled, sometimes the iterator may not finish
    # initialization before passing to a multi device function, add a sync point
    # here to make sure all underlying iterators are initialized.
    if context.executing_eagerly():
      context.async_wait()

    return iterator

  @property
  def element_spec(self):
    """The type specification of an element of this dataset."""
    # When partial batch handling is enabled, always set the batch dimension to
    # None, otherwise we just follow element_spec of the underlying dataset
    # (whose batch dimension may also be None). This is because with partial
    # batching handling we could always produce empty batches.
    if (self._enable_get_next_as_optional and
        self._strategy.extended._in_multi_worker_mode()):  # pylint: disable=protected-access
      return nest.map_structure(
          _rebatch_as_dynamic, self._element_spec, expand_composites=False)
    return self._element_spec


class DistributedDatasetV1(DistributedDataset):
  """Distributed dataset that supports prefetching to multiple devices."""

  def __init__(self,
               dataset,
               input_workers,
               strategy,
               num_replicas_in_sync=None,
               input_context=None):
    self._input_workers = input_workers
    super(DistributedDatasetV1, self).__init__(
        dataset,
        input_workers,
        strategy,
        num_replicas_in_sync=num_replicas_in_sync,
        input_context=input_context)

  def make_one_shot_iterator(self):
    """Get a one time use iterator for DistributedDatasetV1.

    Note: This API is deprecated. Please use `for ... in dataset:` to iterate
    over the dataset or `iter` to create an iterator.

    Returns:
      A DistributedIteratorV1 instance.
    """
    return self._make_one_shot_iterator()

  def _make_one_shot_iterator(self):
    """Get an iterator for DistributedDatasetV1."""
    # Graph mode with one shot iterator is disabled because we have to call
    # `initialize` on the iterator which is only required if we are using a
    # tf.distribute strategy.
    if not context.executing_eagerly():
      raise ValueError("Cannot create a one shot iterator. Please use "
                       "`make_initializable_iterator()` instead.")
    return self._get_iterator()

  def make_initializable_iterator(self):
    """Get an initializable iterator for DistributedDatasetV1.

    Note: This API is deprecated. Please use
    `tf.compat.v1.data.make_initializable_iterator(dataset)` to create an
    initializable iterator.

    Returns:
      A DistributedIteratorV1 instance.
    """
    return self._make_initializable_iterator()

  def _make_initializable_iterator(self, shared_name=None):  # pylint: disable=unused-argument
    """Get an initializable iterator for DistributedDatasetV1."""
    # Eager mode generates already initialized iterators. Hence we cannot create
    # an initializable iterator.
    if context.executing_eagerly():
      raise ValueError("Cannot create initializable iterator in Eager mode. "
                       "Please use `iter()` instead.")
    return self._get_iterator()

  def _get_iterator(self):
    worker_iterators = _create_iterators_per_worker(self._cloned_datasets,
                                                    self._input_workers,
                                                    True)
    iterator = DistributedIteratorV1(self._input_workers, worker_iterators,
                                     self._strategy,
                                     self._enable_get_next_as_optional)
    iterator._element_spec = self.element_spec  # pylint: disable=protected-access

    # When async eager is enabled, sometimes the iterator may not finish
    # initialization before passing to a multi device function, add a sync point
    # here to make sure all underlying iterators are initialized.
    if context.executing_eagerly():
      context.async_wait()

    return iterator

  def __iter__(self):
    if (ops.executing_eagerly_outside_functions() or
        ops.get_default_graph().building_function):
      return self._get_iterator()

    raise RuntimeError("__iter__() is only supported inside of tf.function "
                       "or when eager execution is enabled.")


# TODO(priyag): Add other replication modes.
class DistributedDatasetsFromFunction(_IterableInput):
  """Inputs created from dataset function."""

  def __init__(self, dataset_fn, input_workers, input_contexts, strategy,
               options):
    """Makes an iterable from datasets created by the given function.

    Args:
      dataset_fn: A function that returns a `Dataset` given an `InputContext`.
      input_workers: an `InputWorkers` object.
      input_contexts: A list of `InputContext` instances to be passed to call(s)
        to `dataset_fn`. Length and order should match worker order in
        `worker_device_pairs`.
      strategy: a `tf.distribute.Strategy` object, used to run all-reduce to
        handle last partial batch.
      options: `tf.distribute.InputOptions` used to control options on how this
        dataset is distributed.
    """
    super(DistributedDatasetsFromFunction, self).__init__(
        input_workers=input_workers)

    if input_workers.num_workers != len(input_contexts):
      raise ValueError(
          "Number of input workers (%d) is not same as number of "
          "input_contexts (%d)" %
          (input_workers.num_workers, len(input_contexts)))

    self._input_workers = input_workers
    self._input_contexts = input_contexts
    self._strategy = strategy
    self._options = options
    self._datasets, element_spec = (
        _create_datasets_from_function_with_input_context(
            self._input_contexts, self._input_workers, dataset_fn))
    self._enable_get_next_as_optional = _enable_get_next_as_optional(
        self._strategy, self._datasets[0])
    self._element_spec = _create_distributed_tensor_spec(
        self._strategy, element_spec)

  def __iter__(self):
    if (ops.executing_eagerly_outside_functions() or
        ops.get_default_graph().building_function):
      # This is an optional flag that can be used to turn off using
      # OwnedMultiDeviceIterators and instead use the legacy
      # MultiDeviceIterators as a stop gap solution that will allow us to roll
      # out this change.
      enable_legacy_iterators = getattr(self._strategy,
                                        "_enable_legacy_iterators", False)
      iterators = _create_iterators_per_worker(self._datasets,
                                               self._input_workers,
                                               enable_legacy_iterators,
                                               self._options)
      if enable_legacy_iterators:
        iterator = DistributedIteratorV1(
            self._input_workers,
            iterators,
            self._strategy,
            enable_get_next_as_optional=self._enable_get_next_as_optional)
      else:
        iterator = DistributedIterator(
            input_workers=self._input_workers,
            iterators=iterators,
            strategy=self._strategy,
            enable_get_next_as_optional=self._enable_get_next_as_optional,
            options=self._options)
      iterator._element_spec = self._element_spec  # pylint: disable=protected-access

      # When async eager is enabled, sometimes the iterator may not finish
      # initialization before passing to a multi device function, add a sync
      # point here to make sure all underlying iterators are initialized.
      if context.executing_eagerly():
        context.async_wait()

      return iterator

    raise RuntimeError("__iter__() is only supported inside of tf.function "
                       "or when eager execution is enabled.")

  @property
  def element_spec(self):
    """The type specification of an element of this dataset."""
    # When partial batch handling is enabled, always set the batch dimension to
    # None, otherwise we just follow element_spec of the underlying dataset
    # (whose batch dimension may also be None). This is because with partial
    # batching handling we could always produce empty batches.
    if (self._enable_get_next_as_optional and
        self._strategy.extended._in_multi_worker_mode()):  # pylint: disable=protected-access
      return nest.map_structure(
          _rebatch_as_dynamic, self._element_spec, expand_composites=False)
    return self._element_spec


class DistributedDatasetsFromFunctionV1(DistributedDatasetsFromFunction):
  """Inputs created from dataset function."""

  def _make_initializable_iterator(self, shared_name=None):
    """Get an initializable iterator for DistributedDatasetsFromFunctionV1."""
    del shared_name  # Unused
    # Eager mode generates already initialized iterators. Hence we cannot create
    # an initializable iterator.
    if context.executing_eagerly():
      raise ValueError("Cannot create initializable iterator in Eager mode. "
                       "Please use `iter()` instead.")
    return self._get_iterator()

  def _make_one_shot_iterator(self):
    """Get an iterator for iterating over DistributedDatasetsFromFunctionV1."""
    # Graph mode with one shot iterator is disabled because we have to call
    # `initialize` on the iterator which is only required if we are using a
    # tf.distribute strategy.
    if not context.executing_eagerly():
      raise ValueError("Cannot create a one shot iterator. Please use "
                       "`make_initializable_iterator()` instead.")
    return self._get_iterator()

  def _get_iterator(self):
    iterators = _create_iterators_per_worker(self._datasets,
                                             self._input_workers, True)
    iterator = DistributedIteratorV1(self._input_workers, iterators,
                                     self._strategy,
                                     self._enable_get_next_as_optional)
    iterator._element_spec = self._element_spec  # pylint: disable=protected-access

    # When async eager is enabled, sometimes the iterator may not finish
    # initialization before passing to a multi device function, add a sync point
    # here to make sure all underlying iterators are initialized.
    if context.executing_eagerly():
      context.async_wait()

    return iterator

  def __iter__(self):
    if (ops.executing_eagerly_outside_functions() or
        ops.get_default_graph().building_function):
      return self._get_iterator()

    raise RuntimeError("__iter__() is only supported inside of tf.function "
                       "or when eager execution is enabled.")


# TODO(anjalisridhar): This class will be soon removed in favor of newer
# APIs.
class InputFunctionIterator(DistributedIteratorV1):
  """Iterator created from input function."""

  def __init__(self, input_fn, input_workers, input_contexts, strategy):
    """Make an iterator for input provided via an input function.

    Currently implements PER_WORKER mode, in which the `input_fn` is called
    once on each worker.

    TODO(priyag): Add other replication modes.

    Args:
      input_fn: Input function that returns a `tf.data.Dataset` object.
      input_workers: an `InputWorkers` object.
      input_contexts: A list of `InputContext` instances to be passed to call(s)
        to `input_fn`. Length and order should match worker order in
        `worker_device_pairs`.
      strategy: a `tf.distribute.Strategy` object, used to run all-reduce to
        handle last partial batch.
    """
    assert isinstance(input_workers, InputWorkers)
    if input_workers.num_workers != len(input_contexts):
      raise ValueError(
          "Number of input workers (%d) is not same as number of "
          "input_contexts (%d)" %
          (input_workers.num_workers, len(input_contexts)))

    iterators = []
    for i, ctx in enumerate(input_contexts):
      worker = input_workers.worker_devices[i]
      with ops.device(worker):
        result = input_fn(ctx)
        devices = input_workers.compute_devices_for_worker(i)
        if isinstance(result, dataset_ops.DatasetV2):
          iterator = _SingleWorkerDatasetIterator(result, worker, devices)
        elif callable(result):
          iterator = _SingleWorkerCallableIterator(result, worker, devices)
        else:
          raise ValueError(
              "input_fn must return a tf.data.Dataset or a callable.")
        iterators.append(iterator)

    super(InputFunctionIterator, self).__init__(
        input_workers, iterators, strategy, enable_get_next_as_optional=False)
    self._enable_get_next_as_optional = False


# TODO(anjalisridhar): This class will soon be removed and users should move
# to using DistributedIterator.
class DatasetIterator(DistributedIteratorV1):
  """Iterator created from input dataset."""

  def __init__(self,
               dataset,
               input_workers,
               strategy,
               num_replicas_in_sync=None,
               input_context=None):
    """Make an iterator for the dataset on given devices.

    If `num_replicas_in_sync` is not None, we split each batch of the dataset
    into `num_replicas_in_sync` smaller batches, to be distributed among that
    worker's replicas, so that the batch size for a global step (across all
    workers and replicas) is as expected.

    Args:
      dataset: `tf.data.Dataset` that will be used as the input source.
      input_workers: an `InputWorkers` object.
      strategy: a `tf.distribute.Strategy` object, used to run all-reduce to
        handle last partial batch.
      num_replicas_in_sync: Optional integer. If this is not None, the value is
        used to decide how to rebatch datasets into smaller batches so that the
        total batch size for each step (across all workers and replicas) adds up
        to `dataset`'s batch size.
      input_context: `InputContext` for sharding. Only pass this in for between
        graph multi-worker cases where there is only one `input_worker`. In
        these cases, we will shard based on the `input_pipeline_id` and
        `num_input_pipelines` in the `InputContext`.
    """
    dist_dataset = DistributedDatasetV1(
        dataset,
        input_workers,
        strategy,
        num_replicas_in_sync=num_replicas_in_sync,
        input_context=input_context)
    worker_iterators = _create_iterators_per_worker(
        dist_dataset._cloned_datasets, input_workers, True)  # pylint: disable=protected-access
    super(DatasetIterator,
          self).__init__(input_workers, worker_iterators, strategy,
                         dist_dataset._enable_get_next_as_optional)  # pylint: disable=protected-access
    self._element_spec = dist_dataset.element_spec


def _dummy_tensor_fn(value_structure):
  """A function to create dummy tensors from `value_structure`."""

  def create_dummy_tensor(spec):
    """Create a dummy tensor with possible batch dimensions set to 0."""
    if isinstance(spec, ragged_tensor.RaggedTensorSpec):
      # Splice out the ragged dimensions.
      # pylint: disable=protected-access
      feature_shape = spec._shape[:1].concatenate(
          spec._shape[(1 + spec._ragged_rank):])
      feature_type = spec._dtype
      # pylint: enable=protected-access
    else:
      feature_shape = spec.shape
      feature_type = spec.dtype
    # Ideally we should set the batch dimension to 0, however as in
    # DistributionStrategy we don't know the batch dimension, we try to
    # guess it as much as possible. If the feature has unknown dimensions, we
    # will set them to 0. If the feature shape is already static, we guess the
    # first dimension as batch dimension and set it to 0.
    dims = ([dim if dim is not None else 0 for dim in feature_shape.as_list()]
            if feature_shape else [])
    if dims and (isinstance(spec, ragged_tensor.RaggedTensorSpec) or
                 feature_shape.is_fully_defined()):
      dims[0] = tensor_shape.Dimension(0)

    if isinstance(spec, sparse_tensor.SparseTensorSpec):
      return sparse_tensor.SparseTensor(
          values=array_ops.zeros(0, feature_type),
          indices=array_ops.zeros((0, len(dims)), dtypes.int64),
          dense_shape=dims)

    # Create the dummy tensor.
    dummy_tensor = array_ops.zeros(tensor_shape.TensorShape(dims), feature_type)
    if isinstance(spec, ragged_tensor.RaggedTensorSpec):
      # Reinsert the ragged dimensions with size 0.
      # pylint: disable=protected-access
      row_splits = array_ops.zeros(1, spec._row_splits_dtype)
      dummy_tensor = ragged_tensor.RaggedTensor.from_nested_row_splits(
          dummy_tensor, (row_splits,) * spec._ragged_rank, validate=False)
      # pylint: enable=protected-access
    return dummy_tensor

  return nest.map_structure(create_dummy_tensor, value_structure)


def _recover_shape_fn(data, value_structure):
  """Recover the shape of `data` the same as shape of `value_structure`."""

  flattened_data = nest.flatten(data)
  for i, spec in enumerate(nest.flatten(value_structure)):
    for target, source in zip(
        nest.flatten(flattened_data[i], expand_composites=True),
        nest.flatten(spec, expand_composites=True)):
      target.set_shape(source.shape)
    # `SparseTensor` shape is not determined by the shape of its component
    # tensors. Rather, its shape depends on a tensor's values.
    if isinstance(spec, sparse_tensor.SparseTensorSpec) and spec.shape:
      dense_shape = spec.shape
      with ops.device(flattened_data[i].op.device):
        # For partially defined shapes, fill in missing values from tensor.
        if not dense_shape.is_fully_defined():
          dense_shape = array_ops.stack([
              flattened_data[i].dense_shape[j] if dim is None else dim
              for j, dim in enumerate(dense_shape.as_list())
          ])
        flattened_data[i] = sparse_tensor.SparseTensor(
            indices=flattened_data[i].indices,
            values=flattened_data[i].values,
            dense_shape=dense_shape)
  data = nest.pack_sequence_as(data, flattened_data)
  return data


class _SingleWorkerDatasetIteratorBase(object):
  """Iterator for a single `tf.data.Dataset`."""

  def __init__(self, dataset, worker, devices, options=None):
    """Create iterator for the `dataset` to fetch data to worker's `devices` .

    A `MultiDeviceIterator`  or `OwnedMultiDeviceIterator` is used to prefetch
    input to the devices on the given worker.

    Args:
      dataset: A `tf.data.Dataset` instance.
      worker: Worker on which ops should be created.
      devices: Distribute data from `dataset` to these devices.
      options: options.
    """
    self._dataset = dataset
    self._worker = worker
    self._devices = devices
    self._element_spec = dataset.element_spec
    self._options = options
    self._make_iterator()

  def _make_iterator(self):
    raise NotImplementedError("must be implemented in descendants")

  def _format_data_list_with_options(self, data_list):
    """Change the data in to a list type if required.

    The OwnedMultiDeviceIterator returns the list data type,
    while the PER_REPLICA iterator (when used with prefetch disabled)
    returns without the enclosed list. This is to fix the inconsistency.
    Args:
      data_list: data_list
    Returns:
      list
    """
    if (self._options and self._options.experimental_replication_mode ==
        InputReplicationMode.PER_REPLICA and
        not self._options.experimental_prefetch_to_device):
      return [data_list]
    else:
      return data_list

  def get_next(self, device, name=None):
    """Get next element for the given device."""
    del name
    with ops.device(self._worker):
      if _should_use_multi_device_iterator(self._options):
        return self._iterator.get_next(device)
      else:
        return self._iterator.get_next()

  def get_next_as_list_static_shapes(self, name=None):
    """Get next element from the underlying iterator.

    Runs the iterator get_next() within a device scope. Since this doesn't use
    get_next_as_optional(), it is considerably faster than get_next_as_list()
    (but can only be used when the shapes are static).

    Args:
      name: not used.

    Returns:
      A list consisting of the next data from each device.
    """
    del name
    with ops.device(self._worker):
      return self._format_data_list_with_options(self._iterator.get_next())

  def get_next_as_list(self, name=None):
    """Get next element from underlying iterator.

    If there is no data left, a list of dummy tensors with possible batch
    dimensions set to 0 will be returned. Use of get_next_as_optional() and
    extra logic adds overhead compared to get_next_as_list_static_shapes(), but
    allows us to handle non-static shapes.

    Args:
      name: not used.

    Returns:
      A boolean tensor indicates whether there is any data in next element and
      the real data as the next element or a list of dummy tensors if no data
      left.
    """
    del name
    with ops.device(self._worker):
      data_list = self._format_data_list_with_options(
          self._iterator.get_next_as_optional())
      result = []
      for i, data in enumerate(data_list):
        # Place the condition op in the same device as the data so the data
        # doesn't need to be sent back to the worker.
        with ops.device(self._devices[i]):
          # Data will be fetched in order, so we only need to check if the first
          # replica has value to see whether there is data left for this single
          # worker.
          if i == 0:
            worker_has_value = data.has_value()

          # pylint: disable=unnecessary-lambda
          # pylint: disable=cell-var-from-loop
          real_data = control_flow_ops.cond(
              data.has_value(),
              lambda: data.get_value(),
              lambda: _dummy_tensor_fn(data.element_spec),
              strict=True,
          )
          # Some dimensions in `replicas` will become unknown after we
          # conditionally return the real tensors or the dummy tensors. Recover
          # the shapes from `data.element_spec`. We only need to do this in
          # non eager mode because we always know the runtime shape of the
          # tensors in eager mode.
          if not context.executing_eagerly():
            real_data = _recover_shape_fn(real_data, data.element_spec)
          result.append(real_data)
          # pylint: enable=cell-var-from-loop
          # pylint: enable=unnecessary-lambda

      return worker_has_value, result


class _SingleWorkerDatasetIteratorSpec(type_spec.TypeSpec):
  """Type specification for `_SingleWorkerOwnedDatasetIterator`."""

  __slots__ = ["_worker", "_devices", "_element_spec", "_options"]

  def __init__(self, worker, devices, element_spec, options):
    self._worker = worker
    self._devices = tuple(device_util.canonicalize(d) for d in devices)
    self._element_spec = element_spec
    self._options = options

  @property
  def value_type(self):
    return _SingleWorkerOwnedDatasetIterator

  def _serialize(self):
    return (self._worker, self._devices, self._element_spec, self._options)

  @property
  def _component_specs(self):
    specs = []
    if _should_use_multi_device_iterator(self._options):
      specs.append(
          multi_device_iterator_ops.MultiDeviceIteratorSpec(
              self._devices, self._worker, element_spec=self._element_spec))
    else:
      specs.append(iterator_ops.IteratorSpec(element_spec=self._element_spec))
    return specs

  def _to_components(self, value):
    return [value._iterator]  # pylint: disable=protected-access

  def _from_components(self, components):
    return _SingleWorkerOwnedDatasetIterator(
        dataset=None,
        worker=self._worker,
        devices=self._devices,
        components=components,
        element_spec=self._element_spec,
        options=self._options)

  @staticmethod
  def from_value(value):
    # pylint: disable=protected-access
    return _SingleWorkerDatasetIteratorSpec(value._worker, value._devices,
                                            value._element_spec, value._options)


class _SingleWorkerOwnedDatasetIterator(_SingleWorkerDatasetIteratorBase,
                                        composite_tensor.CompositeTensor):
  """Iterator for a DistributedDataset instance."""

  def __init__(self,
               dataset=None,
               worker=None,
               devices=None,
               components=None,
               element_spec=None,
               options=None):
    """Create iterator for the `dataset` to fetch data to worker's `devices` .

    `OwnedMultiDeviceIterator` is used to prefetch input to the devices on the
    given worker. The lifetime of this iterator is tied to the encompassing
    python object. Once we go out of scope of the python object or return from
    a tf.function the underlying iterator resource is deleted.

    Args:
      dataset: A `tf.data.Dataset` instance.
      worker: Worker on which ops should be created.
      devices: Distribute data from `dataset` to these devices.
      components: Tensor components to construct the
        _SingleWorkerOwnedDatasetIterator from.
      element_spec: A nested structure of `TypeSpec` objects that represents the
      type specification of elements of the iterator.
      options: `tf.distribute.InputOptions` used to control options on how this
      dataset is distributed.
    """
    if worker is None or devices is None:
      raise ValueError("Both `worker` and `devices` should be provided")

    error_message = ("Either `dataset` or both `components` and `element_spec` "
                     "need to be provided.")

    self._options = options
    if dataset is None:
      if (components is None or element_spec is None):
        raise ValueError(error_message)
      self._element_spec = element_spec
      self._worker = worker
      self._devices = devices
      self._iterator = components[0]
    else:
      if (components is not None or element_spec is not None):
        raise ValueError(error_message)
      super(_SingleWorkerOwnedDatasetIterator,
            self).__init__(dataset, worker, devices, options)

  def _make_iterator(self):
    """Make appropriate iterator on the dataset."""
    if not self._worker:
      raise ValueError("Worked device must be specified when creating an "
                       "owned iterator.")
    if _should_use_multi_device_iterator(self._options):
      host_device = device_util.get_host_for_device(self._worker)
      with ops.device(self._worker):
        self._iterator = multi_device_iterator_ops.OwnedMultiDeviceIterator(
            self._dataset, self._devices, source_device=host_device)
    else:
      with ops.device(self._worker):
        self._iterator = iter(self._dataset)

  @property
  def element_spec(self):
    return self._element_spec

  @property
  def _type_spec(self):
    return _SingleWorkerDatasetIteratorSpec(self._worker, self._devices,
                                            self._element_spec, self._options)

  @property
  def output_classes(self):
    """Returns the class of each component of an element of this iterator.

    The expected values are `tf.Tensor` and `tf.SparseTensor`.

    Returns:
      A nested structure of Python `type` objects corresponding to each
      component of an element of this dataset.
    """
    return nest.map_structure(
        lambda component_spec: component_spec._to_legacy_output_classes(),  # pylint: disable=protected-access
        self._element_spec)

  @property
  def output_shapes(self):
    """Returns the shape of each component of an element of this iterator.

    Returns:
      A nested structure of `tf.TensorShape` objects corresponding to each
      component of an element of this dataset.
    """
    return nest.map_structure(
        lambda component_spec: component_spec._to_legacy_output_shapes(),  # pylint: disable=protected-access
        self._element_spec)

  @property
  def output_types(self):
    """Returns the type of each component of an element of this iterator.

    Returns:
      A nested structure of `tf.DType` objects corresponding to each component
      of an element of this dataset.
    """
    return nest.map_structure(
        lambda component_spec: component_spec._to_legacy_output_types(),  # pylint: disable=protected-access
        self._element_spec)


class _SingleWorkerDatasetIterator(_SingleWorkerDatasetIteratorBase):
  """Iterator for a single DistributedDatasetV1 instance."""

  def _make_iterator(self):
    """Make appropriate iterator on the dataset."""
    with ops.device(self._worker):
      self._iterator = multi_device_iterator_ops.MultiDeviceIterator(
          self._dataset, self._devices)

  def initialize(self):
    """Initialize underlying iterator.

    In eager execution, this simply recreates the underlying iterator.
    In graph execution, it returns the initializer ops for the underlying
    iterator.

    Returns:
      A list of any initializer ops that should be run.
    """
    if ops.executing_eagerly_outside_functions():
      self._iterator._eager_reset()  # pylint: disable=protected-access
      return []
    else:
      return [self._iterator.initializer]

  @property
  def output_classes(self):
    return dataset_ops.get_legacy_output_classes(self._iterator)

  @property
  def output_shapes(self):
    return dataset_ops.get_legacy_output_shapes(self._iterator)

  @property
  def output_types(self):
    return dataset_ops.get_legacy_output_types(self._iterator)


class _SingleWorkerCallableIterator(object):
  """Iterator for a single tensor-returning callable."""

  def __init__(self, fn, worker, devices):
    self._fn = fn
    self._worker = worker
    self._devices = devices

  def get_next(self, device, name=None):
    """Get next element for the given device from the callable."""
    del device, name
    with ops.device(self._worker):
      return self._fn()

  def get_next_as_list_static_shapes(self, name=None):
    """Get next element from the callable."""
    del name
    with ops.device(self._worker):
      data_list = [self._fn() for _ in self._devices]
      return data_list

  def get_next_as_list(self, name=None):
    """Get next element from the callable."""
    del name
    with ops.device(self._worker):
      data_list = [self._fn() for _ in self._devices]
      return constant_op.constant(True), data_list

  def initialize(self):
    # TODO(petebu) Should this throw an exception instead?
    return []


def _create_iterators_per_worker(worker_datasets,
                                 input_workers,
                                 enable_legacy_iterators,
                                 options=None):
  """Create a multidevice iterator on each of the workers."""
  assert isinstance(input_workers, InputWorkers)
  assert len(worker_datasets) == len(input_workers.worker_devices)
  iterators = []
  for i, worker in enumerate(input_workers.worker_devices):
    with ops.device(worker):
      worker_devices = input_workers.compute_devices_for_worker(i)
      if tf2.enabled() and not enable_legacy_iterators:
        iterator = _SingleWorkerOwnedDatasetIterator(
            dataset=worker_datasets[i],
            worker=worker,
            devices=worker_devices,
            options=options)
      else:
        iterator = _SingleWorkerDatasetIterator(worker_datasets[i], worker,
                                                worker_devices, options)
      iterators.append(iterator)
  return iterators


def _create_datasets_from_function_with_input_context(input_contexts,
                                                      input_workers,
                                                      dataset_fn):
  """Create device datasets per worker given a dataset function."""
  datasets = []
  for i, ctx in enumerate(input_contexts):
    worker = input_workers.worker_devices[i]
    with ops.device(worker):
      dataset = dataset_fn(ctx)
      datasets.append(dataset)
  return datasets, dataset.element_spec


# TODO(sourabhbajaj): Remove this in lieu of distributed datasets
def _get_batched_dataset(d):
  """Get the batched dataset from `d`."""
  # pylint: disable=protected-access
  if isinstance(d, dataset_ops.DatasetV1Adapter):
    d = d._dataset

  if isinstance(d, (dataset_ops.BatchDataset, batching._MapAndBatchDataset)):
    return d
  elif isinstance(d, (dataset_ops.PrefetchDataset,
                      dataset_ops._OptionsDataset)):
    return _get_batched_dataset(d._input_dataset)

  raise ValueError(
      "Unable to get batched dataset from the input dataset. `batch` "
      "`map_and_batch` need to be the last operations on the dataset. "
      "The batch operations can be followed by a prefetch.")


def _get_batched_dataset_attributes(d):
  """Get `batch_size`, `drop_remainder` of dataset."""
  # pylint: disable=protected-access
  assert isinstance(d,
                    (dataset_ops.BatchDataset, batching._MapAndBatchDataset))
  if isinstance(d, dataset_ops.BatchDataset):
    batch_size = d._batch_size
    drop_remainder = d._drop_remainder
  elif isinstance(d, batching._MapAndBatchDataset):
    batch_size = d._batch_size_t
    drop_remainder = d._drop_remainder_t
  # pylint: enable=protected-access

  if tensor_util.is_tf_type(batch_size):
    batch_size = tensor_util.constant_value(batch_size)

  if tensor_util.is_tf_type(drop_remainder):
    drop_remainder = tensor_util.constant_value(drop_remainder)

  return batch_size, drop_remainder


# TODO(sourabhbajaj): Remove this in lieu of distributed datasets
def _get_dataset_attributes(dataset):
  """Get the underlying attributes from the dataset object."""
  # pylint: disable=protected-access

  # First, get batch_size and drop_remainder from the dataset. We need
  # to walk back the dataset creation process and find the batched version in
  # order to get the attributes.
  batched_dataset = _get_batched_dataset(dataset)
  batch_size, drop_remainder = _get_batched_dataset_attributes(batched_dataset)

  # Second, prefetch buffer should be get from the original dataset.
  prefetch_buffer = None
  if isinstance(dataset, dataset_ops.PrefetchDataset):
    prefetch_buffer = dataset._buffer_size
  elif (isinstance(dataset, dataset_ops.DatasetV1Adapter)
        and isinstance(dataset._dataset, dataset_ops.PrefetchDataset)):
    prefetch_buffer = dataset._dataset._buffer_size

  return batch_size, drop_remainder, prefetch_buffer


def _should_use_multi_device_iterator(options):
  """Determine whether to use multi_device_iterator_ops."""
  if (options is None or
      options.experimental_replication_mode == InputReplicationMode.PER_WORKER
      or
      (options.experimental_replication_mode == InputReplicationMode.PER_REPLICA
       and options.experimental_prefetch_to_device)):
    return True
  return False


class MultiStepContext(object):
  """A context object that can be used to capture things when running steps.

  This context object is useful when running multiple steps at a time using the
  `experimental_run_steps_on_iterator` API. For e.g. it allows the user's step
  function to specify which outputs to emit at what frequency. Currently it
  supports capturing output from the last step, as well as capturing non tensor
  outputs.  In the future it will be augmented to support other use cases such
  as output each N steps.
  """

  def __init__(self):
    """Initialize an output context.

    Returns:
      A context object.
    """
    self._last_step_outputs = {}
    self._last_step_outputs_reduce_ops = {}
    self._non_tensor_outputs = {}

  @property
  def last_step_outputs(self):
    """A dictionary consisting of outputs to be captured on last step.

    Keys in the dictionary are names of tensors to be captured, as specified
    when `set_last_step_output` is called.
    Values in the dictionary are the tensors themselves. If
    `set_last_step_output` was called with a `reduce_op` for this output,
    then the value is the reduced value.

    Returns:
      A dictionary with last step outputs.
    """
    return self._last_step_outputs

  def _set_last_step_outputs(self, outputs):
    """Replace the entire dictionary of last step outputs."""
    if not isinstance(outputs, dict):
      raise ValueError("Need a dictionary to set last_step_outputs.")
    self._last_step_outputs = outputs

  def set_last_step_output(self, name, output, reduce_op=None):
    """Set `output` with `name` to be outputted from the last step.

    Args:
      name: String, name to identify the output. Doesn't need to match tensor
        name.
      output: The tensors that should be outputted with `name`. See below for
        actual types supported.
      reduce_op: Reduction method to use to reduce outputs from multiple
        replicas. Required if `set_last_step_output` is called in a replica
        context. Optional in cross_replica_context.
        When present, the outputs from all the replicas are reduced using the
        current distribution strategy's `reduce` method. Hence, the type of
        `output` must be what's supported by the corresponding `reduce` method.
        For e.g. if using MirroredStrategy and reduction is set, output
        must be a `PerReplica` value.
        The reduce method is also recorded in a dictionary
        `_last_step_outputs_reduce_ops` for later interpreting of the
        outputs as already reduced or not.
    """
    if distribution_strategy_context.in_cross_replica_context():
      self._last_step_outputs_reduce_ops[name] = reduce_op
      if reduce_op is None:
        self._last_step_outputs[name] = output
      else:
        distribution = distribution_strategy_context.get_strategy()
        self._last_step_outputs[name] = distribution.reduce(reduce_op, output,
                                                            axis=None)
    else:
      assert reduce_op is not None
      def merge_fn(distribution, value):
        self._last_step_outputs[name] = distribution.reduce(reduce_op, value,
                                                            axis=None)
        # Setting this inside the `merge_fn` because all replicas share the same
        # context object, so it's more robust to set it only once (even if all
        # the replicas are trying to set the same value).
        self._last_step_outputs_reduce_ops[name] = reduce_op

      distribution_strategy_context.get_replica_context().merge_call(
          merge_fn, args=(output,))

  @property
  def non_tensor_outputs(self):
    """A dictionary consisting of any non tensor outputs to be captured."""
    return self._non_tensor_outputs

  def set_non_tensor_output(self, name, output):
    """Set `output` with `name` to be captured as a non tensor output."""
    if distribution_strategy_context.in_cross_replica_context():
      self._non_tensor_outputs[name] = output
    else:
      def merge_fn(distribution, value):
        # NOTE(priyag): For non tensor outputs, we simply return all the values
        # in a list as reduction doesn't make sense on non tensors.
        self._non_tensor_outputs[name] = (
            distribution.experimental_local_results(value))
      distribution_strategy_context.get_replica_context().merge_call(
          merge_fn, args=(output,))


def _create_distributed_tensor_spec(strategy, tensor_spec):
  """Create a `tf.TypeSpec` for a given strategy and input `tensor_spec`.

  Args:
    strategy: The given `tf.distribute` strategy.
    tensor_spec: `tf.TensorSpec` of a given value. The batch dimension of the
      shape should be None if you have partial batches.

  Returns:
    A `tf.TypeSpec` that matches the values produced by a given strategy. This
    can be a `tf.TensorSpec` or a `PerRelicaSpec`.
  """
  num_replicas = len(strategy.extended.worker_devices)

  # For one device strategy that is not MultiWorkerMirroredStrategy,  return the
  # tensor_spec as is, since we don't wrap the output with PerReplica in this
  # case.
  # TODO(b/166464552): remove after we always wrap for all strategies.
  if not _always_wrap(strategy):
    return tensor_spec

  # For other cases we assume the input to tf.function is a per replica type.
  def _get_value_per_replica(tensor_spec_per_input):
    value_specs = [tensor_spec_per_input for _ in range(num_replicas)]
    return values.PerReplicaSpec(*value_specs)

  return nest.map_structure(_get_value_per_replica, tensor_spec)


def _replace_per_replica_spec(spec, i):
  """If `spec` is a `PerReplicaSpec`, then return its `i`th value_spec."""
  if isinstance(spec, values.PerReplicaSpec):
    return spec._value_specs[i]  # pylint: disable=protected-access
  else:
    return spec


def _enable_get_next_as_optional(strategy, dataset):
  """Returns whether to enable using partial batch handling."""
  # TODO(b/133073708): we currently need a flag to control the usage because
  # there is a performance difference between get_next() and
  # get_next_as_optional(). And we only enable get_next_as_optional when the
  # output shapes are not static.
  #
  # TODO(rxsang): We want to always enable the get_next_as_optional behavior
  # when user passed input_fn instead of dataset.
  if not getattr(strategy.extended, "experimental_enable_get_next_as_optional",
                 False):
    return False

  if context.executing_eagerly():
    # If the dataset is infinite, we don't need to enable last partial batch
    # support. Currently the logic only applies to the case that distributed
    # dataset is created in eager mode, as we need to evaluate the dataset
    # cardinality.
    with ops.device(dataset._variant_tensor.device):  # pylint: disable=protected-access
      if dataset.cardinality().numpy() == cardinality.INFINITE:
        return False

  return not _is_statically_shaped(
      dataset.element_spec) or strategy.extended._in_multi_worker_mode()  # pylint: disable=protected-access


def _create_per_replica(value_list, strategy):
  """Creates a PerReplica.

  For strategies other than OneDeviceStrategy, it creates a PerReplica whose
  type spec is set to the element spec of the dataset. This helps avoid
  retracing for partial batches. Retracing is problematic for multi client when
  different client retraces different time, since retracing changes the
  collective keys in the tf.function, and causes mismatches among clients.

  For single client strategies, this simply calls distribute_utils.regroup().

  Args:
    value_list: a list of values, one for each replica.
    strategy: the `tf.distribute.Strategy`.

  Returns:
    a structure of PerReplica.

  """
  # TODO(b/166464552): always wrap for all one device strategies as well.
  always_wrap = _always_wrap(strategy)
  per_replicas = distribute_utils.regroup(value_list, always_wrap=always_wrap)
  return per_replicas


def _always_wrap(strategy):
  """Returns whether to always wrap the values in a DistributedValues."""
  return strategy.extended._in_multi_worker_mode() or len(  # pylint: disable=protected-access
      strategy.extended.worker_devices) > 1


def _rebatch_as_dynamic(per_replica_spec):
  """Rebatch the spec to have a dynamic batch dimension."""
  assert isinstance(per_replica_spec, values.PerReplicaSpec), per_replica_spec

  # pylint: disable=protected-access
  def _rebatch(spec):
    # Rebatch if possible.
    try:
      return spec._unbatch()._batch(None)
    except ValueError:
      pass
    return spec

  return values.PerReplicaSpec(
      *nest.map_structure(_rebatch, per_replica_spec._value_specs))
  # pylint: enable=protected-access