tensorflow/python/ops/array_ops.py

# Copyright 2015 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.
# ==============================================================================
# Tests for this file live in python/kernel_tests/array_ops_test.py
"""Support for manipulating tensors."""

import numbers
import numpy as np

from tensorflow.core.config import flags
from tensorflow.dtensor.python import api as d_api
from tensorflow.python.eager import context
from tensorflow.python.eager import record
from tensorflow.python.framework import common_shapes
from tensorflow.python.framework import composite_tensor
from tensorflow.python.framework import constant_op
from tensorflow.python.framework import constant_tensor_conversion  # pylint: disable=unused-import
from tensorflow.python.framework import dtypes
from tensorflow.python.framework import errors
from tensorflow.python.framework import indexed_slices
from tensorflow.python.framework import ops
from tensorflow.python.framework import sparse_tensor
from tensorflow.python.framework import tensor as tensor_lib
from tensorflow.python.framework import tensor_conversion_registry
from tensorflow.python.framework import tensor_shape
from tensorflow.python.framework import tensor_util
# 'Constant' gets imported in the module 'array_ops'.
from tensorflow.python.framework.constant_op import constant
from tensorflow.python.ops import array_ops_stack
from tensorflow.python.ops import gen_array_ops
from tensorflow.python.ops import gen_math_ops
from tensorflow.python.ops import shape_util
from tensorflow.python.ops import tensor_getitem_override  # pylint: disable=unused-import
# go/tf-wildcard-import
# pylint: disable=wildcard-import
from tensorflow.python.ops.gen_array_ops import *
from tensorflow.python.ops.gen_array_ops import reverse_v2 as reverse  # pylint: disable=unused-import
from tensorflow.python.types import core
from tensorflow.python.util import _pywrap_utils
from tensorflow.python.util import deprecation
from tensorflow.python.util import dispatch
from tensorflow.python.util import nest
from tensorflow.python.util import tf_decorator
from tensorflow.python.util.tf_export import tf_export
# pylint: enable=wildcard-import

# Used for slicing to specify a new 1 size dimension
newaxis = None
tf_export("newaxis").export_constant(__name__, "newaxis")


@tf_export("reshape", v1=["reshape", "manip.reshape"])
@dispatch.add_dispatch_support
def reshape(tensor, shape, name=None):  # pylint: disable=redefined-outer-name
  r"""Reshapes a tensor.

  Given `tensor`, this operation returns a new `tf.Tensor` that has the same
  values as `tensor` in the same order, except with a new shape given by
  `shape`.

  >>> t1 = [[1, 2, 3],
  ...       [4, 5, 6]]
  >>> print(tf.shape(t1).numpy())
  [2 3]
  >>> t2 = tf.reshape(t1, [6])
  >>> t2
  <tf.Tensor: shape=(6,), dtype=int32,
    numpy=array([1, 2, 3, 4, 5, 6], dtype=int32)>
  >>> tf.reshape(t2, [3, 2])
  <tf.Tensor: shape=(3, 2), dtype=int32, numpy=
    array([[1, 2],
           [3, 4],
           [5, 6]], dtype=int32)>

  The `tf.reshape` does not change the order of or the total number of elements
  in the tensor, and so it can reuse the underlying data buffer. This makes it
  a fast operation independent of how big of a tensor it is operating on.

  >>> tf.reshape([1, 2, 3], [2, 2])
  Traceback (most recent call last):
  ...
  InvalidArgumentError: Input to reshape is a tensor with 3 values, but the
  requested shape has 4

  To instead reorder the data to rearrange the dimensions of a tensor, see
  `tf.transpose`.

  >>> t = [[1, 2, 3],
  ...      [4, 5, 6]]
  >>> tf.reshape(t, [3, 2]).numpy()
  array([[1, 2],
         [3, 4],
         [5, 6]], dtype=int32)
  >>> tf.transpose(t, perm=[1, 0]).numpy()
  array([[1, 4],
         [2, 5],
         [3, 6]], dtype=int32)

  If one component of `shape` is the special value -1, the size of that
  dimension is computed so that the total size remains constant.  In particular,
  a `shape` of `[-1]` flattens into 1-D.  At most one component of `shape` can
  be -1.

  >>> t = [[1, 2, 3],
  ...      [4, 5, 6]]
  >>> tf.reshape(t, [-1])
  <tf.Tensor: shape=(6,), dtype=int32,
    numpy=array([1, 2, 3, 4, 5, 6], dtype=int32)>
  >>> tf.reshape(t, [3, -1])
  <tf.Tensor: shape=(3, 2), dtype=int32, numpy=
    array([[1, 2],
           [3, 4],
           [5, 6]], dtype=int32)>
  >>> tf.reshape(t, [-1, 2])
  <tf.Tensor: shape=(3, 2), dtype=int32, numpy=
    array([[1, 2],
           [3, 4],
           [5, 6]], dtype=int32)>

  `tf.reshape(t, [])` reshapes a tensor `t` with one element to a scalar.

  >>> tf.reshape([7], []).numpy().item()
  7

  More examples:

  >>> t = [1, 2, 3, 4, 5, 6, 7, 8, 9]
  >>> print(tf.shape(t).numpy())
  [9]
  >>> tf.reshape(t, [3, 3])
  <tf.Tensor: shape=(3, 3), dtype=int32, numpy=
    array([[1, 2, 3],
           [4, 5, 6],
           [7, 8, 9]], dtype=int32)>

  >>> t = [[[1, 1], [2, 2]],
  ...      [[3, 3], [4, 4]]]
  >>> print(tf.shape(t).numpy())
  [2 2 2]
  >>> tf.reshape(t, [2, 4])
  <tf.Tensor: shape=(2, 4), dtype=int32, numpy=
    array([[1, 1, 2, 2],
           [3, 3, 4, 4]], dtype=int32)>

  >>> t = [[[1, 1, 1],
  ...       [2, 2, 2]],
  ...      [[3, 3, 3],
  ...       [4, 4, 4]],
  ...      [[5, 5, 5],
  ...       [6, 6, 6]]]
  >>> print(tf.shape(t).numpy())
  [3 2 3]
  >>> # Pass '[-1]' to flatten 't'.
  >>> tf.reshape(t, [-1])
  <tf.Tensor: shape=(18,), dtype=int32,
    numpy=array([1, 1, 1, 2, 2, 2, 3, 3, 3, 4, 4, 4, 5, 5, 5, 6, 6, 6],
    dtype=int32)>
  >>> # -- Using -1 to infer the shape --
  >>> # Here -1 is inferred to be 9:
  >>> tf.reshape(t, [2, -1])
  <tf.Tensor: shape=(2, 9), dtype=int32, numpy=
    array([[1, 1, 1, 2, 2, 2, 3, 3, 3],
           [4, 4, 4, 5, 5, 5, 6, 6, 6]], dtype=int32)>
  >>> # -1 is inferred to be 2:
  >>> tf.reshape(t, [-1, 9])
  <tf.Tensor: shape=(2, 9), dtype=int32, numpy=
    array([[1, 1, 1, 2, 2, 2, 3, 3, 3],
           [4, 4, 4, 5, 5, 5, 6, 6, 6]], dtype=int32)>
  >>> # -1 is inferred to be 3:
  >>> tf.reshape(t, [ 2, -1, 3])
  <tf.Tensor: shape=(2, 3, 3), dtype=int32, numpy=
    array([[[1, 1, 1],
            [2, 2, 2],
            [3, 3, 3]],
           [[4, 4, 4],
            [5, 5, 5],
            [6, 6, 6]]], dtype=int32)>

  Args:
    tensor: A `Tensor`.
    shape: A `Tensor`. Must be one of the following types: `int32`, `int64`.
      Defines the shape of the output tensor.
    name: Optional string. A name for the operation.

  Returns:
    A `Tensor`. Has the same type as `tensor`.
  """
  result = gen_array_ops.reshape(tensor, shape, name)
  shape_util.maybe_set_static_shape(result, shape)
  return result


@tf_export("fill")
@dispatch.add_dispatch_support
def fill(dims, value, name=None, layout=None):
  r"""Creates a tensor filled with a scalar value.

  See also `tf.ones`, `tf.zeros`, `tf.one_hot`, `tf.eye`.

  This operation creates a tensor of shape `dims` and fills it with `value`.

  For example:

  >>> tf.fill([2, 3], 9)
  <tf.Tensor: shape=(2, 3), dtype=int32, numpy=
  array([[9, 9, 9],
         [9, 9, 9]], dtype=int32)>

  `tf.fill` evaluates at graph runtime and supports dynamic shapes based on
  other runtime `tf.Tensors`, unlike `tf.constant(value, shape=dims)`, which
  embeds the value as a `Const` node.

  Args:
    dims: A 1-D sequence of non-negative numbers. Represents the shape of the
      output `tf.Tensor`. Entries should be of type: `int32`, `int64`.
    value: A value to fill the returned `tf.Tensor`.
    name: Optional string. The name of the output `tf.Tensor`.
    layout: Optional, `tf.experimental.dtensor.Layout`. If provided, the result
      is a [DTensor](https://www.tensorflow.org/guide/dtensor_overview) with the
      provided layout.

  Returns:
    A `tf.Tensor` with shape `dims` and the same dtype as `value`.

  Raises:
    InvalidArgumentError: `dims` contains negative entries.
    NotFoundError: `dims` contains non-integer entries.

  @compatibility(numpy)
  Similar to `np.full`. In `numpy`, more parameters are supported. Passing a
  number argument as the shape (`np.full(5, value)`) is valid in `numpy` for
  specifying a 1-D shaped result, while TensorFlow does not support this syntax.
  @end_compatibility
  """
  result = d_api.call_with_layout(
      gen_array_ops.fill, layout=layout, dims=dims, value=value, name=name
  )
  shape_util.maybe_set_static_shape(result, dims)
  return result


@tf_export("identity")
@dispatch.add_dispatch_support
def identity(input, name=None):  # pylint: disable=redefined-builtin
  r"""Return a Tensor with the same shape and contents as input.

  The return value is not the same Tensor as the original, but contains the same
  values.  This operation is fast when used on the same device.

  For example:

  >>> a = tf.constant([0.78])
  >>> a_identity = tf.identity(a)
  >>> a.numpy()
  array([0.78], dtype=float32)
  >>> a_identity.numpy()
  array([0.78], dtype=float32)

  Calling `tf.identity` on a variable will make a Tensor that represents the
  value of that variable at the time it is called. This is equivalent to calling
  `<variable>.read_value()`.

  >>> a = tf.Variable(5)
  >>> a_identity = tf.identity(a)
  >>> a.assign_add(1)
  <tf.Variable ... shape=() dtype=int32, numpy=6>
  >>> print(a.numpy())
  6
  >>> print(a_identity.numpy())
  5

  This function can also be used to explicitly transfer tensors between devices.
  For example, to transfer a tensor in GPU memory back to host memory, one can
  use:

  >>> with tf.device("/gpu:0"):
  ...   x_on_gpu = tf.constant(1)
  >>> with tf.device("/cpu:0"):
  ...   x_on_cpu = tf.identity(x_on_gpu)
  >>> x_on_cpu.device
  '/job:localhost/replica:0/task:0/device:CPU:0'

  Args:
    input: A `Tensor`, a `Variable`, a `CompositeTensor` or anything that can be
    converted to a tensor using `tf.convert_to_tensor`.
    name: A name for the operation (optional).

  Returns:
    A `Tensor` or CompositeTensor. Has the same type and contents as `input`.
  """
  # Don't expand ResourceVariables, so identity(variable) will return a Tensor.
  if (isinstance(input, composite_tensor.CompositeTensor) and
      not _pywrap_utils.IsResourceVariable(input)):
    return nest.map_structure(identity, input, expand_composites=True)
  if context.executing_eagerly() and not hasattr(input, "graph"):
    # Make sure we get an input with handle data attached from the resource
    # variables. Variables have correct handle data when graph building.
    input = ops.convert_to_tensor(input)
  ret = gen_array_ops.identity(input, name=name)
  # Propagate handles data for happier shape inference for resource variables.
  if hasattr(input, "_handle_data"):
    ret._handle_data = input._handle_data  # pylint: disable=protected-access
  return ret


# pylint: disable=redefined-builtin,protected-access
@tf_export(v1=["expand_dims"])
@dispatch.add_dispatch_support
@deprecation.deprecated_args(None, "Use the `axis` argument instead", "dim")
def expand_dims(input, axis=None, name=None, dim=None):
  """Returns a tensor with a length 1 axis inserted at index `axis`.

  Given a tensor `input`, this operation inserts a dimension of length 1 at the
  dimension index `axis` of the `input`'s shape. The dimension index follows
  Python
  indexing rules: It's zero-based, a negative index that is counted backward
  from the end.

  This operation is useful to:

  * Add an outer "batch" dimension to a single element.
  * Align axes for broadcasting.
  * To add an inner vector length axis to a tensor of scalars.

  For example:

  If you have a single image of shape `[height, width, channels]`:

  >>> image = tf.zeros([10,10,3])

  You can add an outer `batch` axis by passing `axis=0`:

  >>> tf.expand_dims(image, axis=0).shape.as_list()
  [1, 10, 10, 3]

  The new axis location matches Python `list.insert(axis, 1)`:

  >>> tf.expand_dims(image, axis=1).shape.as_list()
  [10, 1, 10, 3]

  Following standard Python indexing rules, a negative `axis` counts from the
  end so `axis=-1` adds an innermost dimension:

  >>> tf.expand_dims(image, -1).shape.as_list()
  [10, 10, 3, 1]

  This operation requires that `axis` is a valid index for `input.shape`,
  following Python indexing rules:

  ```
  -1-tf.rank(input) <= axis <= tf.rank(input)
  ```

  This operation is related to:

  * `tf.squeeze`, which removes dimensions of size 1.
  * `tf.reshape`, which provides a more flexible reshaping capability.
  * `tf.sparse.expand_dims`, which provides this functionality for
    `tf.SparseTensor`

  Args:
    input: A `Tensor`.
    axis: 0-D (scalar). Specifies the dimension index at which to expand the
      shape of `input`. Must be in the range `[-rank(input) - 1, rank(input)]`.
    name: The name of the output is `Tensor` (optional).
    dim: 0-D (scalar). Equivalent to `axis`, to be deprecated.

  Returns:
    A `Tensor` with the same data as `input`, but its shape has an additional
    dimension of size 1 added.

  Raises:
    ValueError: if either both or neither of `dim` and `axis` are specified.
  """
  axis = deprecation.deprecated_argument_lookup("axis", axis, "dim", dim)
  if axis is None:
    raise ValueError("Must specify an axis argument to tf.expand_dims()")
  return expand_dims_v2(input, axis, name)


@tf_export("expand_dims", v1=[])
@dispatch.add_dispatch_support
def expand_dims_v2(input, axis, name=None):
  """Returns a tensor with a length 1 axis inserted at index `axis`.

  Given a tensor `input`, this operation inserts a dimension of length 1 at the
  dimension index `axis` of `input`'s shape. The dimension index follows Python
  indexing rules: It's zero-based, and a negative index is counted backward
  from the end.

  This operation is useful to:

  * Add an outer "batch" dimension to a single element.
  * Align axes for broadcasting.
  * To add an inner vector length axis to a tensor of scalars.

  For example:

  If you have a single image of shape `[height, width, channels]`:

  >>> image = tf.zeros([10,10,3])

  You can add an outer `batch` axis by passing `axis=0`:

  >>> tf.expand_dims(image, axis=0).shape.as_list()
  [1, 10, 10, 3]

  The new axis location matches Python `list.insert(axis, 1)`:

  >>> tf.expand_dims(image, axis=1).shape.as_list()
  [10, 1, 10, 3]

  Following standard Python indexing rules, a negative `axis` counts from the
  end so `axis=-1` adds an inner most dimension:

  >>> tf.expand_dims(image, -1).shape.as_list()
  [10, 10, 3, 1]

  This operation requires that `axis` is a valid index for `input.shape`,
  following Python indexing rules:

  ```
  -1-tf.rank(input) <= axis <= tf.rank(input)
  ```

  This operation is related to:

  * `tf.squeeze`, which removes dimensions of size 1.
  * `tf.reshape`, which provides more flexible reshaping capability.
  * `tf.sparse.expand_dims`, which provides this functionality for
    `tf.SparseTensor`

  Args:
    input: A `Tensor`.
    axis: Integer specifying the dimension index at which to expand the
      shape of `input`. Given an input of D dimensions, `axis` must be in range
      `[-(D+1), D]` (inclusive).
    name: Optional string. The name of the output `Tensor`.

  Returns:
    A tensor with the same data as `input`, with an additional dimension
    inserted at the index specified by `axis`.

  Raises:
    TypeError: If `axis` is not specified.
    InvalidArgumentError: If `axis` is out of range `[-(D+1), D]`.
  """
  return gen_array_ops.expand_dims(input, axis, name)


# pylint: enable=redefined-builtin,protected-access


# Aliases for some automatically generated names.
# pylint: disable=protected-access
@deprecation.deprecated("2016-11-30",
                        "This op will be removed after the deprecation date. "
                        "Please switch to tf.setdiff1d().")
def listdiff(x, y, out_idx=None, name=None):
  return gen_array_ops.list_diff(x, y, out_idx, name)


listdiff.__doc__ = gen_array_ops.list_diff.__doc__ + "\n" + listdiff.__doc__

# pylint: enable=protected-access


# pylint: disable=undefined-variable
@deprecation.deprecated("2018-11-30",
                        "This op will be removed after the deprecation date. "
                        "Please switch to tf.sets.difference().")
@tf_export(v1=["setdiff1d"])
@dispatch.add_dispatch_support
def setdiff1d(x, y, index_dtype=dtypes.int32, name=None):
  """Computes the difference between two lists of numbers or strings.

  Given a list x and a list y, this operation returns a list out that
  represents all values that are in x but not in y. The returned list
  out is sorted in the same order that the numbers appear in x
  (duplicates are preserved). This operation also returns a list idx
  that represents the position of each out element in x.

  In other words:

  ```python
  out[i] = x[idx[i]] for i in [0, 1, ..., len(out) - 1]
  ```

  Example usage:

  >>> x = [1, 2, 3, 4, 5, 6]
  >>> y = [1, 3, 5]
  >>> setdiff1d(x,y)
  ListDiff(out=<tf.Tensor: id=2, shape=(3,), dtype=int32,
  numpy=array([2, 4, 6], dtype=int32)>, idx=<tf.Tensor: id=3,
  shape=(3,), dtype=int32, numpy=array([1, 3, 5], dtype=int32)>)

  Args:
    x: A Tensor. 1-D. Values to keep.
    y: A Tensor. Must have the same type as x. 1-D. Values to remove.
    out_idx: An optional tf.DType from: tf.int32, tf.int64. Defaults to
      tf.int32.
    name: A name for the operation (optional).

  Returns:
    A tuple of Tensor objects (out, idx).
    out: A Tensor. Has the same type as x.
    idx: A Tensor of type out_idx.
  """
  return gen_array_ops.list_diff(x, y, index_dtype, name)


setdiff1d.__doc__ = gen_array_ops.list_diff.__doc__


@tf_export("broadcast_dynamic_shape")
@dispatch.add_dispatch_support
def broadcast_dynamic_shape(shape_x, shape_y):
  """Computes the shape of a broadcast given symbolic shapes.

  When `shape_x` and `shape_y` are Tensors representing shapes (i.e. the result
  of calling tf.shape on another Tensor) this computes a Tensor which is the
  shape of the result of a broadcasting op applied in tensors of shapes
  `shape_x` and `shape_y`.

  This is useful when validating the result of a broadcasting operation when the
  tensors do not have statically known shapes.

  Example:

  >>> shape_x = (1, 2, 3)
  >>> shape_y = (5, 1, 3)
  >>> tf.broadcast_dynamic_shape(shape_x, shape_y)
  <tf.Tensor: shape=(3,), dtype=int32, numpy=array([5, 2, 3], ...>

  Args:
    shape_x: A rank 1 integer `Tensor`, representing the shape of x.
    shape_y: A rank 1 integer `Tensor`, representing the shape of y.

  Returns:
    A rank 1 integer `Tensor` representing the broadcasted shape.

  Raises:
    InvalidArgumentError: If the two shapes are incompatible for
    broadcasting.
  """
  return gen_array_ops.broadcast_args(shape_x, shape_y)


@tf_export("broadcast_static_shape")
@dispatch.add_dispatch_support
def broadcast_static_shape(shape_x, shape_y):
  """Computes the shape of a broadcast given known shapes.

  When `shape_x` and `shape_y` are fully known `TensorShape`s this computes a
  `TensorShape` which is the shape of the result of a broadcasting op applied in
  tensors of shapes `shape_x` and `shape_y`.

  For example, if shape_x is `TensorShape([1, 2, 3])` and shape_y is
  `TensorShape([5, 1, 3])`, the result is a TensorShape whose value is
  `TensorShape([5, 2, 3])`.

  This is useful when validating the result of a broadcasting operation when the
  tensors have statically known shapes.

  Example:

  >>> shape_x = tf.TensorShape([1, 2, 3])
  >>> shape_y = tf.TensorShape([5, 1 ,3])
  >>> tf.broadcast_static_shape(shape_x, shape_y)
  TensorShape([5, 2, 3])

  Args:
    shape_x: A `TensorShape`
    shape_y: A `TensorShape`

  Returns:
    A `TensorShape` representing the broadcasted shape.

  Raises:
    ValueError: If the two shapes can not be broadcasted.
  """
  return common_shapes.broadcast_shape(shape_x, shape_y)


@tf_export("shape", v1=[])
@dispatch.add_dispatch_support
def shape_v2(input, out_type=None, name=None):
  # pylint: disable=redefined-builtin
  # TODO(b/274626120) Update `tf_shape_default_int64` comment when it is better
  # supported.
  """Returns a tensor containing the shape of the input tensor.

  See also `tf.size`, `tf.rank`.

  `tf.shape` returns a 1-D integer tensor representing the shape of `input`.
  For a scalar input, the tensor returned has a shape of (0,) and its value is
  the empty vector (i.e. []).

  For example:

  >>> tf.shape(1.)
  <tf.Tensor: shape=(0,), dtype=int32, numpy=array([], dtype=int32)>

  >>> t = tf.constant([[[1, 1, 1], [2, 2, 2]], [[3, 3, 3], [4, 4, 4]]])
  >>> tf.shape(t)
  <tf.Tensor: shape=(3,), dtype=int32, numpy=array([2, 2, 3], dtype=int32)>

  Note: When using symbolic tensors, such as when using the Keras API,
  tf.shape() will return the shape of the symbolic tensor.

  >>> a = tf.keras.layers.Input((None, 10))
  >>> tf.shape(a)
  <... shape=(3,) dtype=int32...>

  In these cases, using `tf.Tensor.shape` will return more informative results.

  >>> a.shape
  TensorShape([None, None, 10])

  (The first `None` represents the as yet unknown batch size.)

  `tf.shape` and `Tensor.shape` should be identical in eager mode.  Within
  `tf.function` or within a `compat.v1` context, not all dimensions may be
  known until execution time. Hence, when defining custom layers and models
  for graph mode, prefer the dynamic `tf.shape(x)` over the static `x.shape`.

  Args:
    input: A `Tensor` or `SparseTensor`.
    out_type: (Optional) The specified output type of the operation (`int32` or
      `int64`). Defaults to `tf.int32`. (Note: there is an experimental
      flag, `tf_shape_default_int64` that changes the default to `tf.int64`.
      This is an unsupported, experimental setting that causes known breakages.)
    name: A name for the operation (optional).

  Returns:
    A `Tensor` of type `out_type`.
  """
  if out_type is None:
    if flags.config().tf_shape_default_int64.value():
      out_type = dtypes.int64
    else:
      out_type = dtypes.int32
  return shape(input, name, out_type)


@tf_export(v1=["shape"])
@dispatch.add_dispatch_support
def shape(input, name=None, out_type=None):
  # pylint: disable=redefined-builtin
  """Returns the shape of a tensor.

  This operation returns a 1-D integer tensor representing the shape of `input`.

  For example:

  ```python
  t = tf.constant([[[1, 1, 1], [2, 2, 2]], [[3, 3, 3], [4, 4, 4]]])
  tf.shape(t)  # [2, 2, 3]
  ```

  Args:
    input: A `Tensor` or `SparseTensor`.
    name: A name for the operation (optional).
    out_type: (Optional) The specified output type of the operation (`int32`
    or `int64`). Defaults to `tf.int32`.

  Returns:
    A `Tensor` of type `out_type`.
  """
  if out_type is None:
    if flags.config().tf_shape_default_int64.value():
      out_type = dtypes.int64
    else:
      out_type = dtypes.int32
  return shape_internal(input, name, optimize=True, out_type=out_type)


def shape_internal(input, name=None, optimize=True, out_type=None):
  # pylint: disable=redefined-builtin
  """Returns the shape of a tensor.

  If `out_type` is not specified and the shape is fully known, then we look at
  the dimension values to determine whether to return an int32 or int64 tensor.
  If the shape is not fully known, we default to int32.

  Args:
    input: A `Tensor` or `SparseTensor`.
    name: A name for the operation (optional).
    optimize: if true, encode the shape as a constant when possible.
    out_type: (Optional) The specified output type of the operation (`int32` or
      `int64`). Defaults to tf.int32.

  Returns:
    A `Tensor` of type `out_type`.

  """
  with ops.name_scope(name, "Shape", [input]) as name:
    if isinstance(
        input, (sparse_tensor.SparseTensor, sparse_tensor.SparseTensorValue)):
      if not out_type:
        out_type = dtypes.int32
      return gen_math_ops.cast(input.dense_shape, out_type)
    else:
      if not context.executing_eagerly():
        input = ops.convert_to_tensor(input)
        input_shape = input.get_shape()
        if optimize and input_shape.is_fully_defined():
          # For fully defined shapes, if the out_type is not specified, we pick
          # int32 / int64 based on the actual values.
          if not out_type:
            return constant_op._tensor_shape_tensor_conversion_function(  # pylint: disable=protected-access
                input_shape)
          return constant(input_shape.as_list(), out_type, name=name)
      if not out_type:
        out_type = dtypes.int32
      return gen_array_ops.shape(input, name=name, out_type=out_type)


@tf_export("shape_n")
@dispatch.add_dispatch_support
def shape_n(input, out_type=dtypes.int32, name=None):
  # pylint: disable=redefined-builtin
  """Returns shape of a list of tensors.

  Given a list of tensors, `tf.shape_n` is much faster than applying `tf.shape`
  to each tensor individually.
  >>> a = tf.ones([1, 2])
  >>> b = tf.ones([2, 3])
  >>> c = tf.ones([3, 4])
  >>> tf.shape_n([a, b, c])
  [<tf.Tensor: shape=(2,), dtype=int32, numpy=array([1, 2], dtype=int32)>,
  <tf.Tensor: shape=(2,), dtype=int32, numpy=array([2, 3], dtype=int32)>,
  <tf.Tensor: shape=(2,), dtype=int32, numpy=array([3, 4], dtype=int32)>]

  Args:
    input: A list of at least 1 `Tensor` object with the same dtype.
    out_type: The specified output type of the operation (`int32` or `int64`).
      Defaults to `tf.int32`(optional).
    name: A name for the operation (optional).

  Returns:
    A list of `Tensor` specifying the shape of each input tensor with type of
    `out_type`.
  """

  return gen_array_ops.shape_n(input, out_type=out_type, name=name)


@tf_export("size", v1=[])
@dispatch.add_dispatch_support
def size_v2(input, out_type=None, name=None):
  # pylint: disable=redefined-builtin
  """Returns the size of a tensor.

  See also `tf.shape`.

  Returns a 0-D `Tensor` representing the number of elements in `input`
  of type `out_type`. Defaults to tf.int32.

  For example:

  >>> t = tf.constant([[[1, 1, 1], [2, 2, 2]], [[3, 3, 3], [4, 4, 4]]])
  >>> tf.size(t)
  <tf.Tensor: shape=(), dtype=int32, numpy=12>

  Args:
    input: A `Tensor` or `SparseTensor`.
    out_type: (Optional) The specified non-quantized numeric output type of the
      operation. Defaults to `tf.int32`. (Note: there is an experimental
      flag, `tf_shape_default_int64` that changes the default to `tf.int64`.
      This is an unsupported, experimental setting that causes known breakages.)
    name: A name for the operation (optional).

  Returns:
    A `Tensor` of type `out_type`. Defaults to `tf.int32`.

  @compatibility(numpy)
  Equivalent to np.size()
  @end_compatibility
  """
  if out_type is None:
    if flags.config().tf_shape_default_int64.value():
      out_type = dtypes.int64
    else:
      out_type = dtypes.int32
  return size(input, name, out_type)


@tf_export(v1=["size"])
@dispatch.add_dispatch_support
def size(input, name=None, out_type=None):
  # pylint: disable=redefined-builtin
  """Returns the size of a tensor.

  Returns a 0-D `Tensor` representing the number of elements in `input`
  of type `out_type`. Defaults to tf.int32.

  For example:

  ```python
  t = tf.constant([[[1, 1, 1], [2, 2, 2]], [[3, 3, 3], [4, 4, 4]]])
  tf.size(t)  # 12
  ```

  Args:
    input: A `Tensor` or `SparseTensor`.
    name: A name for the operation (optional).
    out_type: (Optional) The specified non-quantized numeric output type of the
      operation. Defaults to `tf.int32`. (Note: there is an experimental
      flag, `tf_shape_default_int64` that changes the default to `tf.int64`.
      This is an unsupported, experimental setting that causes known breakages.)

  Returns:
    A `Tensor` of type `out_type`. Defaults to `tf.int32`.

  @compatibility(numpy)
  Equivalent to np.size()
  @end_compatibility
  """
  if out_type is None:
    if flags.config().tf_shape_default_int64.value():
      out_type = dtypes.int64
    else:
      out_type = dtypes.int32
  return size_internal(input, name, optimize=True, out_type=out_type)


def size_internal(input, name=None, optimize=True, out_type=dtypes.int32):
  # pylint: disable=redefined-builtin,protected-access
  """Returns the size of a tensor.

  Args:
    input: A `Tensor` or `SparseTensor`.
    name: A name for the operation (optional).
    optimize: if true, encode the size as a constant when possible.
    out_type: (Optional) The specified non-quantized numeric output type of the
      operation. Defaults to `tf.int32`.

  Returns:
    A `Tensor` of type `out_type`. Defaults to `tf.int32`.
  """
  if (context.executing_eagerly() and not hasattr(input, "graph") and
      not isinstance(
          input,
          (sparse_tensor.SparseTensor, sparse_tensor.SparseTensorValue))):
    input = ops.convert_to_tensor(input)
    np_out_type = out_type.as_numpy_dtype
    num_elements = np.prod(input._shape_tuple(), dtype=np_out_type)  # pylint: disable=protected-access
    return ops.convert_to_tensor(num_elements, dtype=out_type)
  with ops.name_scope(name, "Size", [input]) as name:
    if isinstance(
        input, (sparse_tensor.SparseTensor, sparse_tensor.SparseTensorValue)):
      return gen_math_ops.prod(
          gen_math_ops.cast(input.dense_shape, out_type), 0, name=name)
    else:
      input = ops.convert_to_tensor(input)
      input_shape = input.get_shape()
      if optimize:
        if input_shape.is_fully_defined():
          return constant(input_shape.num_elements(), out_type, name=name)
        if input_shape.dims and any(dim == 0 for dim in input_shape.dims):
          return constant(0, out_type, name=name)
      return gen_array_ops.size(input, name=name, out_type=out_type)


@tf_export("rank")
@dispatch.add_dispatch_support
def rank(input, name=None):
  # pylint: disable=redefined-builtin
  """Returns the rank of a tensor.

  See also `tf.shape`.

  Returns a 0-D `int32` `Tensor` representing the rank of `input`.

  For example:

  ```python
  # shape of tensor 't' is [2, 2, 3]
  t = tf.constant([[[1, 1, 1], [2, 2, 2]], [[3, 3, 3], [4, 4, 4]]])
  tf.rank(t)  # 3
  ```

  **Note**: The rank of a tensor is not the same as the rank of a matrix. The
  rank of a tensor is the number of indices required to uniquely select each
  element of the tensor. Rank is also known as "order", "degree", or "ndims."

  Args:
    input: A `Tensor` or `SparseTensor`.
    name: A name for the operation (optional).

  Returns:
    A `Tensor` of type `int32`.

  @compatibility(numpy)
  Equivalent to np.ndim
  @end_compatibility
  """
  return rank_internal(input, name, optimize=True)


def rank_internal(input, name=None, optimize=True):
  # pylint: disable=redefined-builtin
  """Returns the rank of a tensor.

  Args:
    input: A `Tensor` or `SparseTensor`.
    name: A name for the operation (optional).
    optimize: if true, encode the rank as a constant when possible.

  Returns:
    A `Tensor` of type `int32`.
  """
  with ops.name_scope(name, "Rank", [input]) as name:
    if isinstance(
        input, (sparse_tensor.SparseTensor, sparse_tensor.SparseTensorValue)):
      return gen_array_ops.size(input.dense_shape, name=name)
    else:
      input = ops.convert_to_tensor(input)
      input_shape = input.get_shape()
      if optimize and input_shape.ndims is not None:
        return constant(input_shape.ndims, dtypes.int32, name=name)
      return gen_array_ops.rank(input, name=name)


# pylint: disable=undefined-variable,protected-access,redefined-outer-name
@tf_export("slice")
@dispatch.add_dispatch_support
def slice(input_, begin, size, name=None):
  # pylint: disable=redefined-builtin
  """Extracts a slice from a tensor.

  See also `tf.strided_slice`.

  This operation extracts a slice of size `size` from a tensor `input_` starting
  at the location specified by `begin`. The slice `size` is represented as a
  tensor shape, where `size[i]` is the number of elements of the 'i'th dimension
  of `input_` that you want to slice. The starting location (`begin`) for the
  slice is represented as an offset in each dimension of `input_`. In other
  words, `begin[i]` is the offset into the i'th dimension of `input_` that you
  want to slice from.

  Note that `tf.Tensor.__getitem__` is typically a more pythonic way to
  perform slices, as it allows you to write `foo[3:7, :-2]` instead of
  `tf.slice(foo, [3, 0], [4, foo.get_shape()[1]-2])`.

  `begin` is zero-based; `size` is one-based. If `size[i]` is -1,
  all remaining elements in dimension i are included in the
  slice. In other words, this is equivalent to setting:

  `size[i] = input_.dim_size(i) - begin[i]`

  This operation requires that:

  `0 <= begin[i] <= begin[i] + size[i] <= Di  for i in [0, n]`

  For example:

  ```python
  t = tf.constant([[[1, 1, 1], [2, 2, 2]],
                   [[3, 3, 3], [4, 4, 4]],
                   [[5, 5, 5], [6, 6, 6]]])
  tf.slice(t, [1, 0, 0], [1, 1, 3])  # [[[3, 3, 3]]]
  tf.slice(t, [1, 0, 0], [1, 2, 3])  # [[[3, 3, 3],
                                     #   [4, 4, 4]]]
  tf.slice(t, [1, 0, 0], [2, 1, 3])  # [[[3, 3, 3]],
                                     #  [[5, 5, 5]]]
  ```

  Args:
    input_: A `Tensor`.
    begin: An `int32` or `int64` `Tensor`.
    size: An `int32` or `int64` `Tensor`.
    name: A name for the operation (optional).

  Returns:
    A `Tensor` the same type as `input_`.
  """
  return gen_array_ops._slice(input_, begin, size, name=name)


# pylint: disable=invalid-name
@tf_export("strided_slice")
@dispatch.add_dispatch_support
def strided_slice(input_,
                  begin,
                  end,
                  strides=None,
                  begin_mask=0,
                  end_mask=0,
                  ellipsis_mask=0,
                  new_axis_mask=0,
                  shrink_axis_mask=0,
                  var=None,
                  name=None):
  """Extracts a strided slice of a tensor (generalized Python array indexing).

  See also `tf.slice`.

  **Instead of calling this op directly most users will want to use the
  NumPy-style slicing syntax (e.g. `tensor[..., 3:4:-1, tf.newaxis, 3]`), which
  is supported via `tf.Tensor.__getitem__` and `tf.Variable.__getitem__`.**
  The interface of this op is a low-level encoding of the slicing syntax.

  Roughly speaking, this op extracts a slice of size `(end-begin)/stride`
  from the given `input_` tensor. Starting at the location specified by `begin`
  the slice continues by adding `stride` to the index until all dimensions are
  not less than `end`.
  Note that a stride can be negative, which causes a reverse slice.

  Given a Python slice `input[spec0, spec1, ..., specn]`,
  this function will be called as follows.

  `begin`, `end`, and `strides` will be vectors of length n.
  n in general is not equal to the rank of the `input_` tensor.

  In each mask field (`begin_mask`, `end_mask`, `ellipsis_mask`,
  `new_axis_mask`, `shrink_axis_mask`) the ith bit will correspond to
  the ith spec.

  If the ith bit of `begin_mask` is set, `begin[i]` is ignored and
  the fullest possible range in that dimension is used instead.
  `end_mask` works analogously, except with the end range.

  `foo[5:,:,:3]` on a 7x8x9 tensor is equivalent to `foo[5:7,0:8,0:3]`.
  `foo[::-1]` reverses a tensor with shape 8.

  If the ith bit of `ellipsis_mask` is set, as many unspecified dimensions
  as needed will be inserted between other dimensions. Only one
  non-zero bit is allowed in `ellipsis_mask`.

  For example `foo[3:5,...,4:5]` on a shape 10x3x3x10 tensor is
  equivalent to `foo[3:5,:,:,4:5]` and
  `foo[3:5,...]` is equivalent to `foo[3:5,:,:,:]`.

  If the ith bit of `new_axis_mask` is set, then `begin`,
  `end`, and `stride` are ignored and a new length 1 dimension is
  added at this point in the output tensor.

  For example,
  `foo[:4, tf.newaxis, :2]` would produce a shape `(4, 1, 2)` tensor.

  If the ith bit of `shrink_axis_mask` is set, it implies that the ith
  specification shrinks the dimensionality by 1, taking on the value at index
  `begin[i]`. `end[i]` and `strides[i]` are ignored in this case. For example in
  Python one might do `foo[:, 3, :]` which would result in `shrink_axis_mask`
  equal to 2.


  NOTE: `begin` and `end` are zero-indexed.
  `strides` entries must be non-zero.


  ```python
  t = tf.constant([[[1, 1, 1], [2, 2, 2]],
                   [[3, 3, 3], [4, 4, 4]],
                   [[5, 5, 5], [6, 6, 6]]])
  tf.strided_slice(t, [1, 0, 0], [2, 1, 3], [1, 1, 1])  # [[[3, 3, 3]]]
  tf.strided_slice(t, [1, 0, 0], [2, 2, 3], [1, 1, 1])  # [[[3, 3, 3],
                                                        #   [4, 4, 4]]]
  tf.strided_slice(t, [1, -1, 0], [2, -3, 3], [1, -1, 1])  # [[[4, 4, 4],
                                                           #   [3, 3, 3]]]
  ```

  Args:
    input_: A `Tensor`.
    begin: An `int32` or `int64` `Tensor`.
    end: An `int32` or `int64` `Tensor`.
    strides: An `int32` or `int64` `Tensor`.
    begin_mask: An `int32` mask.
    end_mask: An `int32` mask.
    ellipsis_mask: An `int32` mask.
    new_axis_mask: An `int32` mask.
    shrink_axis_mask: An `int32` mask.
    var: The variable corresponding to `input_` or None
    name: A name for the operation (optional).

  Returns:
    A `Tensor` the same type as `input`.
  """

  if strides is None:
    strides = ones_like(begin)

  op = gen_array_ops.strided_slice(
      input=input_,
      begin=begin,
      end=end,
      strides=strides,
      name=name,
      begin_mask=begin_mask,
      end_mask=end_mask,
      ellipsis_mask=ellipsis_mask,
      new_axis_mask=new_axis_mask,
      shrink_axis_mask=shrink_axis_mask)

  parent_name = name

  if var is not None:
    def assign(val, name=None):
      """Closure that holds all the arguments to create an assignment."""

      if name is None:
        name = parent_name + "_assign"

      return var._strided_slice_assign(
          begin=begin,
          end=end,
          strides=strides,
          value=val,
          name=name,
          begin_mask=begin_mask,
          end_mask=end_mask,
          ellipsis_mask=ellipsis_mask,
          new_axis_mask=new_axis_mask,
          shrink_axis_mask=shrink_axis_mask)

    op.assign = assign

  return op


@tf_export("parallel_stack")
@dispatch.add_dispatch_support
def parallel_stack(values, name="parallel_stack"):
  """Stacks a list of rank-`R` tensors into one rank-`(R+1)` tensor in parallel.

  Requires that the shape of inputs be known at graph construction time.

  Packs the list of tensors in `values` into a tensor with rank one higher than
  each tensor in `values`, by packing them along the first dimension.
  Given a list of length `N` of tensors of shape `(A, B, C)`; the `output`
  tensor will have the shape `(N, A, B, C)`.

  For example:

  ```python
  x = tf.constant([1, 4])
  y = tf.constant([2, 5])
  z = tf.constant([3, 6])
  tf.parallel_stack([x, y, z])  # [[1, 4], [2, 5], [3, 6]]
  ```

  The difference between `stack` and `parallel_stack` is that `stack` requires
  all the inputs be computed before the operation will begin but doesn't require
  that the input shapes be known during graph construction.

  `parallel_stack` will copy pieces of the input into the output as they become
  available, in some situations this can provide a performance benefit.

  Unlike `stack`, `parallel_stack` does NOT support backpropagation.

  This is the opposite of unstack.  The numpy equivalent is

      tf.parallel_stack([x, y, z]) = np.asarray([x, y, z])

  @compatibility(eager)
  parallel_stack is not compatible with eager execution.
  @end_compatibility

  Args:
    values: A list of `Tensor` objects with the same shape and type.
    name: A name for this operation (optional).

  Returns:
    output: A stacked `Tensor` with the same type as `values`.

  Raises:
    RuntimeError: if executed in eager mode.
  """
  if context.executing_eagerly():
    raise RuntimeError("tf.parallel_stack() is not compatible with "
                       "eager execution.")
  with ops.name_scope(name):
    value_t = ops.convert_to_tensor(values[0])
    value_shape = ops.convert_to_tensor(value_t).get_shape()

    output_shape = tensor_shape.TensorShape([len(values)])
    output_shape = output_shape.concatenate(value_shape)
    # expand_dims converts concat to stack.
    return gen_array_ops.parallel_concat(
        [expand_dims(value, 0) for value in values], shape=output_shape)


# pylint: disable=invalid-name
def _autopacking_helper(list_or_tuple, dtype, name):
  """Converts the given list or tuple to a tensor by packing.

  Args:
    list_or_tuple: A (possibly nested) list or tuple containing a tensor.
    dtype: The element type of the returned tensor.
    name: A name for the returned tensor.

  Returns:
    A `tf.Tensor` with value equivalent to `list_or_tuple`.
  """
  if context.executing_eagerly():
    # NOTE: Fast path when all the items are tensors, this doesn't do any type
    # checking.
    if all(isinstance(elem, core.Tensor) for elem in list_or_tuple):
      return gen_array_ops.pack(list_or_tuple, name=name)
  must_pack = False
  converted_elems = []
  with ops.name_scope(name) as scope:
    for i, elem in enumerate(list_or_tuple):
      if isinstance(elem, core.Tensor):
        if dtype is not None and elem.dtype.base_dtype != dtype:
          raise TypeError(f"Cannot convert a list containing a tensor of dtype "
                          f"{elem.dtype} to {dtype} (Tensor is: {elem!r})")
        converted_elems.append(elem)
        must_pack = True
      elif isinstance(elem, (list, tuple)):
        converted_elem = _autopacking_helper(elem, dtype, str(i))
        if isinstance(converted_elem, core.Tensor):
          must_pack = True
        converted_elems.append(converted_elem)
      else:
        converted_elems.append(elem)
    if must_pack:
      elems_as_tensors = []
      for i, elem in enumerate(converted_elems):
        if isinstance(elem, core.Tensor):
          elems_as_tensors.append(elem)
        else:
          # NOTE(mrry): This is inefficient, but it enables us to
          # handle the case where the list arguments are other
          # convertible-to-tensor types, such as numpy arrays.
          elems_as_tensors.append(
              constant_op.constant(elem, dtype=dtype, name=str(i)))
      return gen_array_ops.pack(elems_as_tensors, name=scope)
    else:
      return converted_elems


def _get_dtype_from_nested_lists(list_or_tuple):
  """Returns the dtype of any tensor-like object in `list_or_tuple`, if found.

  Args:
    list_or_tuple: A list or tuple representing an object that can be converted
      to a `tf.Tensor`.

  Returns:
    The dtype of any tensor-like object in `list_or_tuple`, or `None` if no
    such object exists.
  """
  for elem in list_or_tuple:
    if isinstance(elem, core.Tensor):
      return elem.dtype.base_dtype
    elif isinstance(elem, (list, tuple)):
      maybe_dtype = _get_dtype_from_nested_lists(elem)
      if maybe_dtype is not None:
        return maybe_dtype
  return None


def _cast_nested_seqs_to_dtype(dtype):

  def _maybe_cast(elem):
    if isinstance(elem, core.Tensor):
      if dtype != elem.dtype.base_dtype:
        elem = gen_math_ops.cast(elem, dtype)
    return elem

  return _maybe_cast


_NON_AUTOPACKABLE_TYPES = set((
    int,
    float,
    complex,
    bool,
    bytes,
    str,
    memoryview,
    np.bool_,
    np.complex64,
    np.clongdouble,
    np.complex128,
    np.float16,
    np.float32,
    np.float64,
    np.longdouble,
    np.int8,
    np.int16,
    np.int32,
    np.int64,
    np.longlong,
    np.timedelta64,
    np.datetime64,
    np.object_,
    np.bytes_,
    np.str_,
    np.uint8,
    np.uint16,
    np.uint32,
    np.uint64,
    np.ulonglong,
    np.void,
))
_NON_AUTOPACKABLE_TYPES.add(np.ndarray)


def _should_not_autopack(v):
  # The condition we really want is
  #    any(isinstance(elem, core.Tensor))
  # but it is >5x slower due to abc.ABCMeta.__instancecheck__.
  # pylint: disable=unidiomatic-typecheck
  # TODO(slebedev): add nest.all?
  return all(type(elem) in _NON_AUTOPACKABLE_TYPES for elem in nest.flatten(v))
  # pylint: enable=unidiomatic-typecheck


def _autopacking_conversion_function(v, dtype=None, name=None, as_ref=False):
  """Tensor conversion function that automatically packs arguments."""
  if as_ref or _should_not_autopack(v):
    return NotImplemented
  inferred_dtype = _get_dtype_from_nested_lists(v)
  if inferred_dtype is None:
    # We did not find any tensor-like objects in the nested lists, so defer to
    # other conversion functions.
    return NotImplemented
  if dtype is None:
    dtype = inferred_dtype
  elif dtype != inferred_dtype:
    v = nest.map_structure(_cast_nested_seqs_to_dtype(dtype), v)
  return _autopacking_helper(v, dtype, name or "packed")


# pylint: enable=invalid-name

# NOTE: Register this conversion function to run *before* one that
# assumes every element is a value.
tensor_conversion_registry.register_tensor_conversion_function(
    (list, tuple), _autopacking_conversion_function, 99)


@tf_export("concat")
@dispatch.add_dispatch_support
def concat(values, axis, name="concat"):
  """Concatenates tensors along one dimension.

  See also `tf.tile`, `tf.stack`, `tf.repeat`.

  Concatenates the list of tensors `values` along dimension `axis`.  If
  `values[i].shape = [D0, D1, ... Daxis(i), ...Dn]`, the concatenated
  result has shape

      [D0, D1, ... Raxis, ...Dn]

  where

      Raxis = sum(Daxis(i))

  That is, the data from the input tensors is joined along the `axis`
  dimension.

  The number of dimensions of the input tensors must match, and all dimensions
  except the `axis` must be equal.

  For example:

  >>> t1 = [[1, 2, 3], [4, 5, 6]]
  >>> t2 = [[7, 8, 9], [10, 11, 12]]
  >>> tf.concat([t1, t2], 0)
  <tf.Tensor: shape=(4, 3), dtype=int32, numpy=
  array([[ 1,  2,  3],
         [ 4,  5,  6],
         [ 7,  8,  9],
         [10, 11, 12]], dtype=int32)>

  >>> tf.concat([t1, t2], 1)
  <tf.Tensor: shape=(2, 6), dtype=int32, numpy=
  array([[ 1,  2,  3,  7,  8,  9],
         [ 4,  5,  6, 10, 11, 12]], dtype=int32)>

  As in Python, the `axis` could also be negative numbers. Negative `axis`
  are interpreted as counting from the end of the rank, i.e.,
   `axis + rank(values)`-th dimension.

  For example:

  >>> t1 = [[[1, 2], [2, 3]], [[4, 4], [5, 3]]]
  >>> t2 = [[[7, 4], [8, 4]], [[2, 10], [15, 11]]]
  >>> tf.concat([t1, t2], -1)
  <tf.Tensor: shape=(2, 2, 4), dtype=int32, numpy=
    array([[[ 1,  2,  7,  4],
            [ 2,  3,  8,  4]],
           [[ 4,  4,  2, 10],
            [ 5,  3, 15, 11]]], dtype=int32)>

  Note: If you are concatenating along a new axis consider using stack.
  E.g.

  ```python
  tf.concat([tf.expand_dims(t, axis) for t in tensors], axis)
  ```

  can be rewritten as

  ```python
  tf.stack(tensors, axis=axis)
  ```

  Args:
    values: A list of `Tensor` objects or a single `Tensor`.
    axis: 0-D `int32` `Tensor`.  Dimension along which to concatenate. Must be
      in the range `[-rank(values), rank(values))`. As in Python, indexing for
      axis is 0-based. Positive axis in the rage of `[0, rank(values))` refers
      to `axis`-th dimension. And negative axis refers to `axis +
      rank(values)`-th dimension.
    name: A name for the operation (optional).

  Returns:
    A `Tensor` resulting from concatenation of the input tensors.
  """
  if not isinstance(values, (list, tuple)):
    values = [values]
  # TODO(mrry): Change to return values?
  if len(values) == 1:  # Degenerate case of one tensor.
    # Make a throwaway call to convert_to_tensor to make sure
    # that axis is of the correct type, and make sure that
    # the returned tensor is a scalar.
    # TODO(keveman): Implement a standalone type and shape checker.
    with ops.name_scope(name) as scope:
      ops.convert_to_tensor(
          axis, name="concat_dim",
          dtype=dtypes.int32).get_shape().assert_has_rank(0)
      return identity(values[0], name=name)
  return gen_array_ops.concat_v2(values=values, axis=axis, name=name)


@tf_export(v1=["boolean_mask"])
@dispatch.add_dispatch_support
def boolean_mask(tensor, mask, name="boolean_mask", axis=None):
  """Apply boolean mask to tensor.

  Numpy equivalent is `tensor[mask]`.

  In general, `0 < dim(mask) = K <= dim(tensor)`, and `mask`'s shape must match
  the first K dimensions of `tensor`'s shape.  We then have:
    `boolean_mask(tensor, mask)[i, j1,...,jd] = tensor[i1,...,iK,j1,...,jd]`
  where `(i1,...,iK)` is the ith `True` entry of `mask` (row-major order).
  The `axis` could be used with `mask` to indicate the axis to mask from.
  In that case, `axis + dim(mask) <= dim(tensor)` and `mask`'s shape must match
  the first `axis + dim(mask)` dimensions of `tensor`'s shape.

  See also: `tf.ragged.boolean_mask`, which can be applied to both dense and
  ragged tensors, and can be used if you need to preserve the masked dimensions
  of `tensor` (rather than flattening them, as `tf.boolean_mask` does).

  Examples:

  ```python
  # 1-D example
  tensor = [0, 1, 2, 3]
  mask = np.array([True, False, True, False])
  tf.boolean_mask(tensor, mask)  # [0, 2]

  # 2-D example
  tensor = [[1, 2], [3, 4], [5, 6]]
  mask = np.array([True, False, True])
  tf.boolean_mask(tensor, mask)  # [[1, 2], [5, 6]]
  ```

  Args:
    tensor:  N-D Tensor.
    mask:  K-D boolean Tensor, K <= N and K must be known statically.
    name:  A name for this operation (optional).
    axis:  A 0-D int Tensor representing the axis in `tensor` to mask from. By
      default, axis is 0 which will mask from the first dimension. Otherwise K +
      axis <= N.

  Returns:
    (N-K+1)-dimensional tensor populated by entries in `tensor` corresponding
    to `True` values in `mask`.

  Raises:
    ValueError:  If shapes do not conform.
  """

  def _apply_mask_1d(reshaped_tensor, mask, axis=None):
    """Mask tensor along dimension 0 with a 1-D mask."""
    indices = squeeze(where_v2(mask), axis=[1])
    return gather(reshaped_tensor, indices, axis=axis)

  with ops.name_scope(name, values=[tensor, mask]):
    tensor = ops.convert_to_tensor(tensor, name="tensor")
    mask = ops.convert_to_tensor(mask, name="mask")

    shape_mask = mask.get_shape()
    ndims_mask = shape_mask.ndims
    shape_tensor = tensor.get_shape()
    if ndims_mask == 0:
      raise ValueError("mask cannot be scalar.")
    if ndims_mask is None:
      raise ValueError(
          "Number of mask dimensions must be specified, even if some dimensions"
          " are None.  E.g. shape=[None] is ok, but shape=None is not.")
    axis = 0 if axis is None else axis
    axis_value = tensor_util.constant_value(axis)
    if axis_value is not None:
      axis = axis_value
      shape_tensor[axis:axis + ndims_mask].assert_is_compatible_with(shape_mask)

    leading_size = gen_math_ops.prod(shape(tensor)[axis:axis + ndims_mask], [0])
    tensor = reshape(
        tensor,
        concat([
            shape(tensor)[:axis], [leading_size],
            shape(tensor)[axis + ndims_mask:]
        ], 0))
    # TODO(yongtang): tf.reshape in C++ kernel might have set the shape
    # correctly, so the following may not be needed? It still might be possible
    # that there are some edge case where tensor_util.constant_value resolves
    # more cases than ShapeInference of tf.reshape in C++ kernel.
    if axis_value is not None:
      first_dim = shape_tensor[axis:axis + ndims_mask].num_elements()
      tensor.set_shape(
          tensor_shape.as_shape(shape_tensor[:axis]).concatenate(
              [first_dim]).concatenate(shape_tensor[axis + ndims_mask:]))

    mask = reshape(mask, [-1])
    return _apply_mask_1d(tensor, mask, axis)


@tf_export("boolean_mask", v1=[])
@dispatch.add_dispatch_support
def boolean_mask_v2(tensor, mask, axis=None, name="boolean_mask"):
  """Apply boolean mask to tensor.

  Numpy equivalent is `tensor[mask]`.

  In general, `0 < dim(mask) = K <= dim(tensor)`, and `mask`'s shape must match
  the first K dimensions of `tensor`'s shape.  We then have:
    `boolean_mask(tensor, mask)[i, j1,...,jd] = tensor[i1,...,iK,j1,...,jd]`
  where `(i1,...,iK)` is the ith `True` entry of `mask` (row-major order).
  The `axis` could be used with `mask` to indicate the axis to mask from.
  In that case, `axis + dim(mask) <= dim(tensor)` and `mask`'s shape must match
  the first `axis + dim(mask)` dimensions of `tensor`'s shape.

  See also: `tf.ragged.boolean_mask`, which can be applied to both dense and
  ragged tensors, and can be used if you need to preserve the masked dimensions
  of `tensor` (rather than flattening them, as `tf.boolean_mask` does).

  Examples:

  >>> tensor = [0, 1, 2, 3]  # 1-D example
  >>> mask = np.array([True, False, True, False])
  >>> tf.boolean_mask(tensor, mask)
  <tf.Tensor: shape=(2,), dtype=int32, numpy=array([0, 2], dtype=int32)>

  >>> tensor = [[1, 2], [3, 4], [5, 6]] # 2-D example
  >>> mask = np.array([True, False, True])
  >>> tf.boolean_mask(tensor, mask)
  <tf.Tensor: shape=(2, 2), dtype=int32, numpy=
  array([[1, 2],
         [5, 6]], dtype=int32)>

  Args:
    tensor:  N-D Tensor.
    mask:  K-D boolean Tensor, K <= N and K must be known statically.
    axis:  A 0-D int Tensor representing the axis in `tensor` to mask from. By
      default, axis is 0 which will mask from the first dimension. Otherwise K +
      axis <= N.
    name:  A name for this operation (optional).

  Returns:
    (N-K+1)-dimensional tensor populated by entries in `tensor` corresponding
    to `True` values in `mask`.

  Raises:
    ValueError:  If shapes do not conform.

  Examples:

  ```python
  # 2-D example
  tensor = [[1, 2], [3, 4], [5, 6]]
  mask = np.array([True, False, True])
  boolean_mask(tensor, mask)  # [[1, 2], [5, 6]]
  ```
  """
  return boolean_mask(tensor, mask, name, axis)


@tf_export("sparse.mask", v1=["sparse.mask", "sparse_mask"])
@deprecation.deprecated_endpoints("sparse_mask")
def sparse_mask(a, mask_indices, name=None):
  """Masks elements of `IndexedSlices`.

  Given an `IndexedSlices` instance `a`, returns another `IndexedSlices` that
  contains a subset of the slices of `a`. Only the slices at indices not
  specified in `mask_indices` are returned.

  This is useful when you need to extract a subset of slices in an
  `IndexedSlices` object.

  For example:

  ```python
  # `a` contains slices at indices [12, 26, 37, 45] from a large tensor
  # with shape [1000, 10]
  a.indices  # [12, 26, 37, 45]
  tf.shape(a.values)  # [4, 10]

  # `b` will be the subset of `a` slices at its second and third indices, so
  # we want to mask its first and last indices (which are at absolute
  # indices 12, 45)
  b = tf.sparse.mask(a, [12, 45])

  b.indices  # [26, 37]
  tf.shape(b.values)  # [2, 10]
  ```

  Args:
    a: An `IndexedSlices` instance.
    mask_indices: Indices of elements to mask.
    name: A name for the operation (optional).

  Returns:
    The masked `IndexedSlices` instance.
  """
  with ops.name_scope(name, "sparse_mask", [a, mask_indices]) as name:
    indices = a.indices
    out_indices, to_gather = gen_array_ops.list_diff(indices, mask_indices)
    out_values = gather(a.values, to_gather, name=name)
    return indexed_slices.IndexedSlices(out_values, out_indices, a.dense_shape)


@tf_export("unique")
@dispatch.add_dispatch_support
def unique(x, out_idx=dtypes.int32, name=None):
  """Finds unique elements in a 1-D tensor.

  See also `tf.unique_with_counts`.

  This operation returns a tensor `y` containing all the unique elements
  of `x` sorted in the same order that they occur in `x`. This operation
  also returns a tensor `idx` the same size as `x` that contains the index
  of each value of `x` in the unique output `y`. In other words:


    y[idx[i]] = x[i] for i in [0, 1,...,rank(x) - 1]

  Example usage:

  >>> x = tf.constant([1, 1, 2, 4, 4, 4, 7, 8, 8])
  >>> y, idx = unique(x)
  >>> y
  <tf.Tensor: id=5, shape=(5,), dtype=int32,
  numpy=array([1, 2, 4, 7, 8], dtype=int32)>
  >>> idx
  <tf.Tensor: id=6, shape=(9,), dtype=int32,
  numpy=array([0, 0, 1, 2, 2, 2, 3, 4, 4], dtype=int32)>

  Args:
    x: A Tensor. 1-D.
    out_idx: An optional tf.DType from: tf.int32, tf.int64. Defaults to
      tf.int32.
    name: A name for the operation (optional).

  Returns:
    A tuple of Tensor objects (y, idx).
      y: A Tensor. Has the same type as x.
      idx: A Tensor of type out_idx.

  """
  # TODO(yongtang): switch to v2 once API deprecation
  # period (3 weeks) pass.
  # TODO(yongtang): The documentation should also
  # be updated when switch  to v2.
  return gen_array_ops.unique(x, out_idx, name)


unique.__doc__ = gen_array_ops.unique.__doc__


@tf_export("unique_with_counts")
@dispatch.add_dispatch_support
def unique_with_counts(x, out_idx=dtypes.int32, name=None):
  """Finds unique elements in a 1-D tensor.

  See also `tf.unique`.

  This operation returns a tensor `y` containing all the unique elements
  of `x` sorted in the same order that they occur in `x`. This operation
  also returns a tensor `idx` the same size as `x` that contains the index
  of each value of `x` in the unique output `y`. Finally, it returns a
  third tensor `count` that contains the count of each element of `y`
  in `x`. In other words:

    y[idx[i]] = x[i] for i in [0, 1,...,rank(x) - 1]

  Example usage:

  >>> x = tf.constant([1, 1, 2, 4, 4, 4, 7, 8, 8])
  >>> y, idx, count = unique_with_counts(x)
  >>> y
  <tf.Tensor: id=8, shape=(5,), dtype=int32,
  numpy=array([1, 2, 4, 7, 8], dtype=int32)>
  >>> idx
  <tf.Tensor: id=9, shape=(9,), dtype=int32,
  numpy=array([0, 0, 1, 2, 2, 2, 3, 4, 4], dtype=int32)>
  >>> count
  <tf.Tensor: id=10, shape=(5,), dtype=int32,
  numpy=array([2, 1, 3, 1, 2], dtype=int32)>

  Args:
    x: A Tensor. 1-D.
    out_idx: An optional tf.DType from: tf.int32, tf.int64. Defaults to
      tf.int32.
    name: A name for the operation (optional).

  Returns:
    A tuple of Tensor objects (y, idx, count).
      y: A Tensor. Has the same type as x.
      idx: A Tensor of type out_idx.
      count: A Tensor of type out_idx.

  """
  # TODO(yongtang): switch to v2 once API deprecation
  # period (3 weeks) pass.
  # TODO(yongtang): The documentation should also
  # be updated when switch  to v2.
  return gen_array_ops.unique_with_counts(x, out_idx, name)


unique_with_counts.__doc__ = gen_array_ops.unique_with_counts.__doc__


@tf_export("split")
@dispatch.add_dispatch_support
def split(value, num_or_size_splits, axis=0, num=None, name="split"):
  """Splits a tensor `value` into a list of sub tensors.

  See also `tf.unstack`.

  If `num_or_size_splits` is an `int`,  then it splits `value` along the
  dimension `axis` into `num_or_size_splits` smaller tensors. This requires that
  `value.shape[axis]` is divisible by `num_or_size_splits`.

  If `num_or_size_splits` is a 1-D Tensor (or list), then `value` is split into
  `len(num_or_size_splits)` elements. The shape of the `i`-th
  element has the same size as the `value` except along dimension `axis` where
  the size is `num_or_size_splits[i]`.

  For example:

  >>> x = tf.Variable(tf.random.uniform([5, 30], -1, 1))
  >>>
  >>> # Split `x` into 3 tensors along dimension 1
  >>> s0, s1, s2 = tf.split(x, num_or_size_splits=3, axis=1)
  >>> tf.shape(s0).numpy()
  array([ 5, 10], dtype=int32)
  >>>
  >>> # Split `x` into 3 tensors with sizes [4, 15, 11] along dimension 1
  >>> split0, split1, split2 = tf.split(x, [4, 15, 11], 1)
  >>> tf.shape(split0).numpy()
  array([5, 4], dtype=int32)
  >>> tf.shape(split1).numpy()
  array([ 5, 15], dtype=int32)
  >>> tf.shape(split2).numpy()
  array([ 5, 11], dtype=int32)

  Args:
    value: The `Tensor` to split.
    num_or_size_splits: Either an `int` indicating the number of splits
      along `axis` or a 1-D integer `Tensor` or Python list containing the sizes
      of each output tensor along `axis`. If an `int`, then it must evenly
      divide `value.shape[axis]`; otherwise the sum of sizes along the split
      axis must match that of the `value`.
    axis: An `int` or scalar `int32` `Tensor`. The dimension along which
      to split. Must be in the range `[-rank(value), rank(value))`. Defaults to
      0.
    num: Optional, an `int`, used to specify the number of outputs when it
      cannot be inferred from the shape of `size_splits`.
    name: A name for the operation (optional).

  Returns:
    if `num_or_size_splits` is an `int` returns a list of
    `num_or_size_splits` `Tensor` objects; if `num_or_size_splits` is a 1-D
    list or 1-D `Tensor` returns `num_or_size_splits.get_shape[0]`
    `Tensor` objects resulting from splitting `value`.

  Raises:
    ValueError: If `num` is unspecified and cannot be inferred.
    ValueError: If `num_or_size_splits` is a scalar `Tensor`.
  """
  if isinstance(num_or_size_splits,
                (numbers.Integral, tensor_shape.Dimension)):
    return gen_array_ops.split(
        axis=axis, num_split=num_or_size_splits, value=value, name=name)

  size_splits = ops.convert_to_tensor(num_or_size_splits)

  if size_splits._rank() == 0:
    raise ValueError(
        "Rank-0 tensors are not supported as the num_or_size_splits argument "
        "to split. Argument provided: %s" % (num_or_size_splits,))

  if num is None:
    size_splits_shape = size_splits._shape_tuple()
    if size_splits_shape:
      num = size_splits_shape[0]
    if num is None:
      raise ValueError(
          f"Cannot infer argument `num` from shape {num_or_size_splits}")

  return gen_array_ops.split_v(
      value=value, size_splits=size_splits, axis=axis, num_split=num, name=name)


@tf_export("transpose", v1=[])
@dispatch.add_dispatch_support
def transpose_v2(a, perm=None, conjugate=False, name="transpose"):
  """Transposes `a`, where `a` is a Tensor.

  Permutes the dimensions according to the value of `perm`.

  The returned tensor's dimension `i` will correspond to the input dimension
  `perm[i]`. If `perm` is not given, it is set to (n-1...0), where n is the rank
  of the input tensor. Hence, by default, this operation performs a regular
  matrix transpose on 2-D input Tensors.

  If conjugate is `True` and `a.dtype` is either `complex64` or `complex128`
  then the values of `a` are conjugated and transposed.

  @compatibility(numpy)
  In `numpy` transposes are memory-efficient constant time operations as they
  simply return a new view of the same data with adjusted `strides`.

  TensorFlow does not support strides, so `transpose` returns a new tensor with
  the items permuted.
  @end_compatibility

  For example:

  >>> x = tf.constant([[1, 2, 3], [4, 5, 6]])
  >>> tf.transpose(x)
  <tf.Tensor: shape=(3, 2), dtype=int32, numpy=
  array([[1, 4],
         [2, 5],
         [3, 6]], dtype=int32)>

  Equivalently, you could call `tf.transpose(x, perm=[1, 0])`.

  If `x` is complex, setting conjugate=True gives the conjugate transpose:

  >>> x = tf.constant([[1 + 1j, 2 + 2j, 3 + 3j],
  ...                  [4 + 4j, 5 + 5j, 6 + 6j]])
  >>> tf.transpose(x, conjugate=True)
  <tf.Tensor: shape=(3, 2), dtype=complex128, numpy=
  array([[1.-1.j, 4.-4.j],
         [2.-2.j, 5.-5.j],
         [3.-3.j, 6.-6.j]])>

  'perm' is more useful for n-dimensional tensors where n > 2:

  >>> x = tf.constant([[[ 1,  2,  3],
  ...                   [ 4,  5,  6]],
  ...                  [[ 7,  8,  9],
  ...                   [10, 11, 12]]])

  As above, simply calling `tf.transpose` will default to `perm=[2,1,0]`.

  To take the transpose of the matrices in dimension-0 (such as when you are
  transposing matrices where 0 is the batch dimension), you would set
  `perm=[0,2,1]`.

  >>> tf.transpose(x, perm=[0, 2, 1])
  <tf.Tensor: shape=(2, 3, 2), dtype=int32, numpy=
  array([[[ 1,  4],
          [ 2,  5],
          [ 3,  6]],
          [[ 7, 10],
          [ 8, 11],
          [ 9, 12]]], dtype=int32)>

  Note: This has a shorthand `linalg.matrix_transpose`):

  Args:
    a: A `Tensor`.
    perm: A permutation of the dimensions of `a`.  This should be a vector.
    conjugate: Optional bool. Setting it to `True` is mathematically equivalent
      to tf.math.conj(tf.transpose(input)).
    name: A name for the operation (optional).

  Returns:
    A transposed `Tensor`.
  """
  return transpose(a=a, perm=perm, name=name, conjugate=conjugate)


@tf_export(v1=["transpose"])
@dispatch.add_dispatch_support
def transpose(a, perm=None, name="transpose", conjugate=False):
  """Transposes `a`.

  Permutes the dimensions according to `perm`.

  The returned tensor's dimension i will correspond to the input dimension
  `perm[i]`. If `perm` is not given, it is set to (n-1...0), where n is
  the rank of the input tensor. Hence, by default, this operation performs a
  regular matrix transpose on 2-D input Tensors. If conjugate is True and
  `a.dtype` is either `complex64` or `complex128` then the values of `a`
  are conjugated and transposed.

  @compatibility(numpy)
  In `numpy` transposes are memory-efficient constant time operations as they
  simply return a new view of the same data with adjusted `strides`.

  TensorFlow does not support strides, so `transpose` returns a new tensor with
  the items permuted.
  @end_compatibility

  For example:

  ```python
  x = tf.constant([[1, 2, 3], [4, 5, 6]])
  tf.transpose(x)  # [[1, 4]
                   #  [2, 5]
                   #  [3, 6]]

  # Equivalently
  tf.transpose(x, perm=[1, 0])  # [[1, 4]
                                #  [2, 5]
                                #  [3, 6]]

  # If x is complex, setting conjugate=True gives the conjugate transpose
  x = tf.constant([[1 + 1j, 2 + 2j, 3 + 3j],
                   [4 + 4j, 5 + 5j, 6 + 6j]])
  tf.transpose(x, conjugate=True)  # [[1 - 1j, 4 - 4j],
                                   #  [2 - 2j, 5 - 5j],
                                   #  [3 - 3j, 6 - 6j]]

  # 'perm' is more useful for n-dimensional tensors, for n > 2
  x = tf.constant([[[ 1,  2,  3],
                    [ 4,  5,  6]],
                   [[ 7,  8,  9],
                    [10, 11, 12]]])

  # Take the transpose of the matrices in dimension-0
  # (this common operation has a shorthand `linalg.matrix_transpose`)
  tf.transpose(x, perm=[0, 2, 1])  # [[[1,  4],
                                   #   [2,  5],
                                   #   [3,  6]],
                                   #  [[7, 10],
                                   #   [8, 11],
                                   #   [9, 12]]]
  ```

  Args:
    a: A `Tensor`.
    perm: A permutation of the dimensions of `a`.
    name: A name for the operation (optional).
    conjugate: Optional bool. Setting it to `True` is mathematically equivalent
      to tf.math.conj(tf.transpose(input)).

  Returns:
    A transposed `Tensor`.
  """
  with ops.name_scope(name, "transpose", [a]) as name:
    if not tensor_util.is_tf_type(a):
      a = ops.convert_to_tensor(a, name="a")

    if conjugate and a.dtype.is_complex:
      transpose_fn = gen_array_ops.conjugate_transpose
    else:
      transpose_fn = gen_array_ops.transpose

    if perm is not None:
      return transpose_fn(a, perm, name=name)

    rank = a.shape.rank
    if rank is None:
      perm = gen_math_ops._range(gen_array_ops.rank(a) - 1, -1, -1)
    else:
      perm = np.arange(rank - 1, -1, -1, dtype=np.int32)
    return transpose_fn(a, perm, name=name)


# pylint: disable=invalid-name
@tf_export(
    "linalg.matrix_transpose",
    v1=["linalg.transpose", "linalg.matrix_transpose", "matrix_transpose"])
@dispatch.add_dispatch_support
@deprecation.deprecated_endpoints("matrix_transpose", "linalg.transpose")
def matrix_transpose(a, name="matrix_transpose", conjugate=False):
  """Transposes last two dimensions of tensor `a`.

  For example:

  ```python
  x = tf.constant([[1, 2, 3], [4, 5, 6]])
  tf.linalg.matrix_transpose(x)  # [[1, 4],
                                 #  [2, 5],
                                 #  [3, 6]]

  x = tf.constant([[1 + 1j, 2 + 2j, 3 + 3j],
                   [4 + 4j, 5 + 5j, 6 + 6j]])
  tf.linalg.matrix_transpose(x, conjugate=True)  # [[1 - 1j, 4 - 4j],
                                                 #  [2 - 2j, 5 - 5j],
                                                 #  [3 - 3j, 6 - 6j]]

  # Matrix with two batch dimensions.
  # x.shape is [1, 2, 3, 4]
  # tf.linalg.matrix_transpose(x) is shape [1, 2, 4, 3]
  ```

  Note that `tf.matmul` provides kwargs allowing for transpose of arguments.
  This is done with minimal cost, and is preferable to using this function. E.g.

  ```python
  # Good!  Transpose is taken at minimal additional cost.
  tf.matmul(matrix, b, transpose_b=True)

  # Inefficient!
  tf.matmul(matrix, tf.linalg.matrix_transpose(b))
  ```

  @compatibility(numpy)
  In `numpy` transposes are memory-efficient constant time operations as they
  simply return a new view of the same data with adjusted `strides`.

  TensorFlow does not support strides, `linalg.matrix_transpose` returns a new
  tensor with the items permuted.
  @end_compatibility

  Args:
    a: A `Tensor` with `rank >= 2`.
    name: A name for the operation (optional).
    conjugate: Optional bool. Setting it to `True` is mathematically equivalent
      to tf.math.conj(tf.linalg.matrix_transpose(input)).

  Returns:
    A transposed batch matrix `Tensor`.

  Raises:
    ValueError:  If `a` is determined statically to have `rank < 2`.
  """
  with ops.name_scope(name, values=[a]):
    a = ops.convert_to_tensor(a, name="a")

    # If we know the number of dimensions (statically), we can do two things:
    # 1. Check that `a` is a (batch) matrix.
    # 2. Use a Python list for perm.  This preserves static shape information
    #    and avoids extra computations.
    a_shape = a.get_shape()
    ndims = a_shape.ndims
    if ndims is not None:
      if ndims < 2:
        raise ValueError("Argument `a` should be a (batch) matrix with rank "
                         f">= 2.  Received `a` = {a} with shape: {a_shape}")
      perm = list(range(ndims - 2)) + [ndims - 1] + [ndims - 2]
    else:
      a_rank = rank(a)
      perm = concat(
          (gen_math_ops._range(0, a_rank - 2, 1), [a_rank - 1, a_rank - 2]), 0)

    return transpose(a, perm=perm, conjugate=conjugate)


@tf_export("linalg.diag", v1=["linalg.diag", "matrix_diag"])
@dispatch.add_dispatch_support
@deprecation.deprecated_endpoints("matrix_diag")
def matrix_diag(diagonal,
                name="diag",
                k=0,
                num_rows=-1,
                num_cols=-1,
                padding_value=0,
                align="RIGHT_LEFT"):
  """Returns a batched diagonal tensor with given batched diagonal values.

  Returns a tensor with the contents in `diagonal` as `k[0]`-th to `k[1]`-th
  diagonals of a matrix, with everything else padded with `padding`. `num_rows`
  and `num_cols` specify the dimension of the innermost matrix of the output. If
  both are not specified, the op assumes the innermost matrix is square and
  infers its size from `k` and the innermost dimension of `diagonal`. If only
  one of them is specified, the op assumes the unspecified value is the smallest
  possible based on other criteria.

  Let `diagonal` have `r` dimensions `[I, J, ..., L, M, N]`. The output tensor
  has rank `r+1` with shape `[I, J, ..., L, M, num_rows, num_cols]` when only
  one diagonal is given (`k` is an integer or `k[0] == k[1]`). Otherwise, it has
  rank `r` with shape `[I, J, ..., L, num_rows, num_cols]`.

  The second innermost dimension of `diagonal` has double meaning. When `k` is
  scalar or `k[0] == k[1]`, `M` is part of the batch size [I, J, ..., M], and
  the output tensor is:

  ```
  output[i, j, ..., l, m, n]
    = diagonal[i, j, ..., l, n-max(d_upper, 0)] ; if n - m == d_upper
      padding_value                             ; otherwise
  ```

  Otherwise, `M` is treated as the number of diagonals for the matrix in the
  same batch (`M = k[1]-k[0]+1`), and the output tensor is:

  ```
  output[i, j, ..., l, m, n]
    = diagonal[i, j, ..., l, diag_index, index_in_diag] ; if k[0] <= d <= k[1]
      padding_value                                     ; otherwise
  ```
  where `d = n - m`, `diag_index = k[1] - d`, and
  `index_in_diag = n - max(d, 0) + offset`.

  `offset` is zero except when the alignment of the diagonal is to the right.
  ```
  offset = max_diag_len - diag_len(d) ; if (`align` in {RIGHT_LEFT, RIGHT_RIGHT}
                                             and `d >= 0`) or
                                           (`align` in {LEFT_RIGHT, RIGHT_RIGHT}
                                             and `d <= 0`)
           0                          ; otherwise
  ```
  where `diag_len(d) = min(cols - max(d, 0), rows + min(d, 0))`.

  For example:

  ```
  # The main diagonal.
  diagonal = np.array([[1, 2, 3, 4],            # Input shape: (2, 4)
                       [5, 6, 7, 8]])
  tf.linalg.diag(diagonal) ==> [[[1, 0, 0, 0],  # Output shape: (2, 4, 4)
                                 [0, 2, 0, 0],
                                 [0, 0, 3, 0],
                                 [0, 0, 0, 4]],
                                [[5, 0, 0, 0],
                                 [0, 6, 0, 0],
                                 [0, 0, 7, 0],
                                 [0, 0, 0, 8]]]

  # A superdiagonal (per batch).
  diagonal = np.array([[1, 2, 3],  # Input shape: (2, 3)
                       [4, 5, 6]])
  tf.linalg.diag(diagonal, k = 1)
    ==> [[[0, 1, 0, 0],  # Output shape: (2, 4, 4)
          [0, 0, 2, 0],
          [0, 0, 0, 3],
          [0, 0, 0, 0]],
         [[0, 4, 0, 0],
          [0, 0, 5, 0],
          [0, 0, 0, 6],
          [0, 0, 0, 0]]]

  # A tridiagonal band (per batch).
  diagonals = np.array([[[8, 9, 0],  # Input shape: (2, 2, 3)
                         [1, 2, 3],
                         [0, 4, 5]],
                        [[2, 3, 0],
                         [6, 7, 9],
                         [0, 9, 1]]])
  tf.linalg.diag(diagonals, k = (-1, 1))
    ==> [[[1, 8, 0],  # Output shape: (2, 3, 3)
          [4, 2, 9],
          [0, 5, 3]],
         [[6, 2, 0],
          [9, 7, 3],
          [0, 1, 9]]]

  # RIGHT_LEFT alignment.
  diagonals = np.array([[[0, 8, 9],  # Input shape: (2, 2, 3)
                         [1, 2, 3],
                         [4, 5, 0]],
                        [[0, 2, 3],
                         [6, 7, 9],
                         [9, 1, 0]]])
  tf.linalg.diag(diagonals, k = (-1, 1), align="RIGHT_LEFT")
    ==> [[[1, 8, 0],  # Output shape: (2, 3, 3)
          [4, 2, 9],
          [0, 5, 3]],
         [[6, 2, 0],
          [9, 7, 3],
          [0, 1, 9]]]

  # Rectangular matrix.
  diagonal = np.array([1, 2])  # Input shape: (2)
  tf.linalg.diag(diagonal, k = -1, num_rows = 3, num_cols = 4)
    ==> [[0, 0, 0, 0],  # Output shape: (3, 4)
         [1, 0, 0, 0],
         [0, 2, 0, 0]]

  # Rectangular matrix with inferred num_cols and padding_value = 9.
  tf.linalg.diag(diagonal, k = -1, num_rows = 3, padding_value = 9)
    ==> [[9, 9],  # Output shape: (3, 2)
         [1, 9],
         [9, 2]]
  ```

  Args:
    diagonal: A `Tensor` with `rank k >= 1`.
    name: A name for the operation (optional).
    k: Diagonal offset(s). Positive value means superdiagonal, 0 refers to the
      main diagonal, and negative value means subdiagonals. `k` can be a single
      integer (for a single diagonal) or a pair of integers specifying the low
      and high ends of a matrix band. `k[0]` must not be larger than `k[1]`.
    num_rows: The number of rows of the output matrix. If it is not provided,
      the op assumes the output matrix is a square matrix and infers the matrix
      size from `d_lower`, `d_upper`, and the innermost dimension of `diagonal`.
    num_cols: The number of columns of the output matrix. If it is not provided,
      the op assumes the output matrix is a square matrix and infers the matrix
      size from `d_lower`, `d_upper`, and the innermost dimension of `diagonal`.
    padding_value: The value to fill the area outside the specified diagonal
      band with. Default is 0.
    align: Some diagonals are shorter than `max_diag_len` and need to be padded.
      `align` is a string specifying how superdiagonals and subdiagonals should
      be aligned, respectively. There are four possible alignments: "RIGHT_LEFT"
      (default), "LEFT_RIGHT", "LEFT_LEFT", and "RIGHT_RIGHT". "RIGHT_LEFT"
      aligns superdiagonals to the right (left-pads the row) and subdiagonals to
      the left (right-pads the row). It is the packing format LAPACK uses.
      cuSPARSE uses "LEFT_RIGHT", which is the opposite alignment.

  Returns:
    A Tensor. Has the same type as `diagonal`.
  """
  # Special case to sidestep the tf.constant conversion error:
  # TypeError: Expected bool, got 0 of type 'int' instead.
  if hasattr(diagonal, "dtype") and diagonal.dtype == "bool":
    padding_value = bool(padding_value)

  return gen_array_ops.matrix_diag_v3(
      diagonal=diagonal,
      k=k,
      num_rows=num_rows,
      num_cols=num_cols,
      padding_value=padding_value,
      align=align,
      name=name)


@tf_export("linalg.diag_part", v1=["linalg.diag_part", "matrix_diag_part"])
@dispatch.add_dispatch_support
@deprecation.deprecated_endpoints("matrix_diag_part")
def matrix_diag_part(
    input,  # pylint:disable=redefined-builtin
    name="diag_part",
    k=0,
    padding_value=0,
    align="RIGHT_LEFT"):
  """Returns the batched diagonal part of a batched tensor.

  Returns a tensor with the `k[0]`-th to `k[1]`-th diagonals of the batched
  `input`.

  Assume `input` has `r` dimensions `[I, J, ..., L, M, N]`.
  Let `max_diag_len` be the maximum length among all diagonals to be extracted,
  `max_diag_len = min(M + min(k[1], 0), N + min(-k[0], 0))`
  Let `num_diags` be the number of diagonals to extract,
  `num_diags = k[1] - k[0] + 1`.

  If `num_diags == 1`, the output tensor is of rank `r - 1` with shape
  `[I, J, ..., L, max_diag_len]` and values:

  ```
  diagonal[i, j, ..., l, n]
    = input[i, j, ..., l, n+y, n+x] ; if 0 <= n+y < M and 0 <= n+x < N,
      padding_value                 ; otherwise.
  ```
  where `y = max(-k[1], 0)`, `x = max(k[1], 0)`.

  Otherwise, the output tensor has rank `r` with dimensions
  `[I, J, ..., L, num_diags, max_diag_len]` with values:

  ```
  diagonal[i, j, ..., l, m, n]
    = input[i, j, ..., l, n+y, n+x] ; if 0 <= n+y < M and 0 <= n+x < N,
      padding_value                 ; otherwise.
  ```
  where `d = k[1] - m`, `y = max(-d, 0) - offset`, and `x = max(d, 0) - offset`.

  `offset` is zero except when the alignment of the diagonal is to the right.
  ```
  offset = max_diag_len - diag_len(d) ; if (`align` in {RIGHT_LEFT, RIGHT_RIGHT}
                                             and `d >= 0`) or
                                           (`align` in {LEFT_RIGHT, RIGHT_RIGHT}
                                             and `d <= 0`)
           0                          ; otherwise
  ```
  where `diag_len(d) = min(cols - max(d, 0), rows + min(d, 0))`.

  The input must be at least a matrix.

  For example:

  ```
  input = np.array([[[1, 2, 3, 4],  # Input shape: (2, 3, 4)
                     [5, 6, 7, 8],
                     [9, 8, 7, 6]],
                    [[5, 4, 3, 2],
                     [1, 2, 3, 4],
                     [5, 6, 7, 8]]])

  # A main diagonal from each batch.
  tf.linalg.diag_part(input) ==> [[1, 6, 7],  # Output shape: (2, 3)
                                  [5, 2, 7]]

  # A superdiagonal from each batch.
  tf.linalg.diag_part(input, k = 1)
    ==> [[2, 7, 6],  # Output shape: (2, 3)
         [4, 3, 8]]

  # A band from each batch.
  tf.linalg.diag_part(input, k = (-1, 2))
    ==> [[[3, 8, 0],  # Output shape: (2, 4, 3)
          [2, 7, 6],
          [1, 6, 7],
          [0, 5, 8]],
         [[3, 4, 0],
          [4, 3, 8],
          [5, 2, 7],
          [0, 1, 6]]]

  # RIGHT_LEFT alignment.
  tf.linalg.diag_part(input, k = (-1, 2), align="RIGHT_LEFT")
    ==> [[[0, 3, 8],  # Output shape: (2, 4, 3)
          [2, 7, 6],
          [1, 6, 7],
          [5, 8, 0]],
         [[0, 3, 4],
          [4, 3, 8],
          [5, 2, 7],
          [1, 6, 0]]]

  # max_diag_len can be shorter than the main diagonal.
  tf.linalg.diag_part(input, k = (-2, -1))
    ==> [[[5, 8],
          [0, 9]],
         [[1, 6],
          [0, 5]]]

  # padding_value = 9
  tf.linalg.diag_part(input, k = (1, 3), padding_value = 9)
    ==> [[[4, 9, 9],  # Output shape: (2, 3, 3)
          [3, 8, 9],
          [2, 7, 6]],
         [[2, 9, 9],
          [3, 4, 9],
          [4, 3, 8]]]

  ```

  Args:
    input: A `Tensor` with `rank k >= 2`.
    name: A name for the operation (optional).
    k: Diagonal offset(s). Positive value means superdiagonal, 0 refers to the
      main diagonal, and negative value means subdiagonals. `k` can be a single
      integer (for a single diagonal) or a pair of integers specifying the low
      and high ends of a matrix band. `k[0]` must not be larger than `k[1]`.
    padding_value: The value to fill the area outside the specified diagonal
      band with. Default is 0.
    align: Some diagonals are shorter than `max_diag_len` and need to be padded.
      `align` is a string specifying how superdiagonals and subdiagonals should
      be aligned, respectively. There are four possible alignments: "RIGHT_LEFT"
      (default), "LEFT_RIGHT", "LEFT_LEFT", and "RIGHT_RIGHT". "RIGHT_LEFT"
      aligns superdiagonals to the right (left-pads the row) and subdiagonals to
      the left (right-pads the row). It is the packing format LAPACK uses.
      cuSPARSE uses "LEFT_RIGHT", which is the opposite alignment.

  Returns:
    A Tensor containing diagonals of `input`. Has the same type as `input`.

  Raises:
    InvalidArgumentError: When `k` is out of bound or when `k[0]>k[1:]`.
  """
  # Special case to sidestep the tf.constant conversion error:
  # TypeError: Expected bool, got 0 of type 'int' instead.
  if hasattr(input, "dtype") and input.dtype == "bool":
    padding_value = bool(padding_value)

  return gen_array_ops.matrix_diag_part_v3(
      input=input, k=k, padding_value=padding_value, align=align, name=name)


@tf_export(
    "linalg.tensor_diag_part", v1=["linalg.tensor_diag_part", "diag_part"])
@dispatch.add_dispatch_support
@deprecation.deprecated_endpoints("diag_part")
def tensor_diag_part(
    input,  # pylint:disable=redefined-builtin
    name=None):
  """Returns the diagonal part of the tensor.

  This operation returns a tensor with the `diagonal` part
  of the `input`. The `diagonal` part is computed as follows:

  Assume `input` has dimensions `[D1,..., Dk, D1,..., Dk]`, then the output is a
  tensor of rank `k` with dimensions `[D1,..., Dk]` where:

  `diagonal[i1,..., ik] = input[i1, ..., ik, i1,..., ik]`.

  For a rank 2 tensor, `linalg.diag_part` and `linalg.tensor_diag_part`
  produce the same result. For rank 3 and higher, linalg.diag_part extracts
  the diagonal of each inner-most matrix in the tensor. An example where
  they differ is given below.

  >>> x = [[[[1111,1112],[1121,1122]],
  ...       [[1211,1212],[1221,1222]]],
  ...      [[[2111, 2112], [2121, 2122]],
  ...       [[2211, 2212], [2221, 2222]]]
  ...      ]
  >>> tf.linalg.tensor_diag_part(x)
  <tf.Tensor: shape=(2, 2), dtype=int32, numpy=
  array([[1111, 1212],
         [2121, 2222]], dtype=int32)>
  >>> tf.linalg.diag_part(x).shape
  TensorShape([2, 2, 2])

  Args:
    input: A `Tensor` with rank `2k`.
    name: A name for the operation (optional).

  Returns:
    A Tensor containing diagonals of `input`. Has the same type as `input`, and
    rank `k`.
  """
  return gen_array_ops.diag_part(input=input, name=name)


@tf_export("linalg.set_diag", v1=["linalg.set_diag", "matrix_set_diag"])
@dispatch.add_dispatch_support
@deprecation.deprecated_endpoints("matrix_set_diag")
def matrix_set_diag(
    input,  # pylint:disable=redefined-builtin
    diagonal,
    name="set_diag",
    k=0,
    align="RIGHT_LEFT"):
  """Returns a batched matrix tensor with new batched diagonal values.

  Given `input` and `diagonal`, this operation returns a tensor with the
  same shape and values as `input`, except for the specified diagonals of the
  innermost matrices. These will be overwritten by the values in `diagonal`.

  `input` has `r+1` dimensions `[I, J, ..., L, M, N]`. When `k` is scalar or
  `k[0] == k[1]`, `diagonal` has `r` dimensions `[I, J, ..., L, max_diag_len]`.
  Otherwise, it has `r+1` dimensions `[I, J, ..., L, num_diags, max_diag_len]`.
  `num_diags` is the number of diagonals, `num_diags = k[1] - k[0] + 1`.
  `max_diag_len` is the longest diagonal in the range `[k[0], k[1]]`,
  `max_diag_len = min(M + min(k[1], 0), N + min(-k[0], 0))`

  The output is a tensor of rank `k+1` with dimensions `[I, J, ..., L, M, N]`.
  If `k` is scalar or `k[0] == k[1]`:

  ```
  output[i, j, ..., l, m, n]
    = diagonal[i, j, ..., l, n-max(k[1], 0)] ; if n - m == k[1]
      input[i, j, ..., l, m, n]              ; otherwise
  ```

  Otherwise,

  ```
  output[i, j, ..., l, m, n]
    = diagonal[i, j, ..., l, diag_index, index_in_diag] ; if k[0] <= d <= k[1]
      input[i, j, ..., l, m, n]                         ; otherwise
  ```
  where `d = n - m`, `diag_index = k[1] - d`, and
  `index_in_diag = n - max(d, 0) + offset`.

  `offset` is zero except when the alignment of the diagonal is to the right.
  ```
  offset = max_diag_len - diag_len(d) ; if (`align` in {RIGHT_LEFT, RIGHT_RIGHT}
                                             and `d >= 0`) or
                                           (`align` in {LEFT_RIGHT, RIGHT_RIGHT}
                                             and `d <= 0`)
           0                          ; otherwise
  ```
  where `diag_len(d) = min(cols - max(d, 0), rows + min(d, 0))`.

  For example:

  ```
  # The main diagonal.
  input = np.array([[[7, 7, 7, 7],              # Input shape: (2, 3, 4)
                     [7, 7, 7, 7],
                     [7, 7, 7, 7]],
                    [[7, 7, 7, 7],
                     [7, 7, 7, 7],
                     [7, 7, 7, 7]]])
  diagonal = np.array([[1, 2, 3],               # Diagonal shape: (2, 3)
                       [4, 5, 6]])
  tf.matrix_set_diag(input, diagonal)
    ==> [[[1, 7, 7, 7],  # Output shape: (2, 3, 4)
          [7, 2, 7, 7],
          [7, 7, 3, 7]],
         [[4, 7, 7, 7],
          [7, 5, 7, 7],
          [7, 7, 6, 7]]]

  # A superdiagonal (per batch).
  tf.matrix_set_diag(input, diagonal, k = 1)
    ==> [[[7, 1, 7, 7],  # Output shape: (2, 3, 4)
          [7, 7, 2, 7],
          [7, 7, 7, 3]],
         [[7, 4, 7, 7],
          [7, 7, 5, 7],
          [7, 7, 7, 6]]]

  # A band of diagonals.
  diagonals = np.array([[[9, 1, 0],  # Diagonal shape: (2, 4, 3)
                         [6, 5, 8],
                         [1, 2, 3],
                         [0, 4, 5]],
                        [[1, 2, 0],
                         [5, 6, 4],
                         [6, 1, 2],
                         [0, 3, 4]]])
  tf.matrix_set_diag(input, diagonals, k = (-1, 2))
    ==> [[[1, 6, 9, 7],  # Output shape: (2, 3, 4)
          [4, 2, 5, 1],
          [7, 5, 3, 8]],
         [[6, 5, 1, 7],
          [3, 1, 6, 2],
          [7, 4, 2, 4]]]

  # RIGHT_LEFT alignment.
  diagonals = np.array([[[0, 9, 1],  # Diagonal shape: (2, 4, 3)
                         [6, 5, 8],
                         [1, 2, 3],
                         [4, 5, 0]],
                        [[0, 1, 2],
                         [5, 6, 4],
                         [6, 1, 2],
                         [3, 4, 0]]])
  tf.matrix_set_diag(input, diagonals, k = (-1, 2), align="RIGHT_LEFT")
    ==> [[[1, 6, 9, 7],  # Output shape: (2, 3, 4)
          [4, 2, 5, 1],
          [7, 5, 3, 8]],
         [[6, 5, 1, 7],
          [3, 1, 6, 2],
          [7, 4, 2, 4]]]

  ```

  Args:
    input: A `Tensor` with rank `k + 1`, where `k >= 1`.
    diagonal:  A `Tensor` with rank `k`, when `d_lower == d_upper`, or `k + 1`,
      otherwise. `k >= 1`.
    name: A name for the operation (optional).
    k: Diagonal offset(s). Positive value means superdiagonal, 0 refers to the
      main diagonal, and negative value means subdiagonals. `k` can be a single
      integer (for a single diagonal) or a pair of integers specifying the low
      and high ends of a matrix band. `k[0]` must not be larger than `k[1]`.
    align: Some diagonals are shorter than `max_diag_len` and need to be padded.
      `align` is a string specifying how superdiagonals and subdiagonals should
      be aligned, respectively. There are four possible alignments: "RIGHT_LEFT"
      (default), "LEFT_RIGHT", "LEFT_LEFT", and "RIGHT_RIGHT". "RIGHT_LEFT"
      aligns superdiagonals to the right (left-pads the row) and subdiagonals to
      the left (right-pads the row). It is the packing format LAPACK uses.
      cuSPARSE uses "LEFT_RIGHT", which is the opposite alignment.
  """
  return gen_array_ops.matrix_set_diag_v3(
      input=input, diagonal=diagonal, k=k, align=align, name=name)


# pylint: enable=invalid-name


def _constant_if_small(value, shape, dtype, name, layout=None):
  try:
    if np.prod(shape) < 1000:
      return d_api.call_with_layout(
          constant, layout, value, shape=shape, dtype=dtype, name=name
      )
  except (NotImplementedError, TypeError):
    # Happens when shape is a Tensor, list with Tensor elements, etc.
    pass
  return None


def _tag_zeros_tensor(fun):
  """ Tags the result of function by setting _is_zeros_tensor attribute.

  This is useful to compute Hessians of fused ops such as cross_entropy.
  """

  def wrapped(*args, **kwargs):
    tensor = fun(*args, **kwargs)
    tensor._is_zeros_tensor = True
    return tensor

  return tf_decorator.make_decorator(fun, wrapped)


@tf_export("zeros")
@dispatch.add_dispatch_support
@_tag_zeros_tensor
def zeros(shape, dtype=dtypes.float32, name=None, layout=None):
  """Creates a tensor with all elements set to zero.

  See also `tf.zeros_like`, `tf.ones`, `tf.fill`, `tf.eye`.

  This operation returns a tensor of type `dtype` with shape `shape` and
  all elements set to zero.

  >>> tf.zeros([3, 4], tf.int32)
  <tf.Tensor: shape=(3, 4), dtype=int32, numpy=
  array([[0, 0, 0, 0],
         [0, 0, 0, 0],
         [0, 0, 0, 0]], dtype=int32)>

  Args:
    shape: A `list` of integers, a `tuple` of integers, or a 1-D `Tensor` of
      type `int32`.
    dtype: The DType of an element in the resulting `Tensor`.
    name: Optional string. A name for the operation.
    layout: Optional, `tf.experimental.dtensor.Layout`. If provided, the result
      is a [DTensor](https://www.tensorflow.org/guide/dtensor_overview) with the
      provided layout.

  Returns:
    A `Tensor` with all elements set to zero.
  """
  dtype = dtypes.as_dtype(dtype).base_dtype
  with ops.name_scope(name, "zeros", [shape]) as name:
    if dtype == dtypes.bool:
      zero = False
    elif dtype == dtypes.string:
      zero = ""
    elif dtype.is_quantized:
      zero = np.zeros([]).astype(dtype.as_numpy_dtype)
    else:
      zero = 0

    if not isinstance(shape, tensor_lib.Tensor):
      try:
        if not context.executing_eagerly():
          # Create a constant if it won't be very big. Otherwise, create a fill
          # op to prevent serialized GraphDefs from becoming too large.
          output = _constant_if_small(zero, shape, dtype, name, layout=layout)
          if output is not None:
            return output

        # Go through tensor shapes to get int64-if-needed semantics
        shape = constant_op._tensor_shape_tensor_conversion_function(
            tensor_shape.TensorShape(shape))
      except (TypeError, ValueError, errors.UnimplementedError):
        # Happens when shape is a list with tensor elements
        shape = ops.convert_to_tensor(shape, dtype=dtypes.int32)
    if not shape._shape_tuple():
      shape = reshape(shape, [-1])  # Ensure it's a vector
    output = fill(shape, constant(zero, dtype=dtype), name=name, layout=layout)
  assert output.dtype.base_dtype == dtype
  return output


@tf_export(v1=["zeros_like"])
@dispatch.register_unary_elementwise_api
@dispatch.add_dispatch_support
def zeros_like(tensor, dtype=None, name=None, optimize=True):
  """Creates a tensor with all elements set to zero.

  See also `tf.zeros`.

  Given a single tensor (`tensor`), this operation returns a tensor of the
  same type and shape as `tensor` with all elements set to zero. Optionally,
  you can use `dtype` to specify a new type for the returned tensor.

  Examples:

    >>> tensor = tf.constant([[1, 2, 3], [4, 5, 6]])
    >>> tf.zeros_like(tensor)
    <tf.Tensor: shape=(2, 3), dtype=int32, numpy=
    array([[0, 0, 0],
           [0, 0, 0]], dtype=int32)>

    >>> tf.zeros_like(tensor, dtype=tf.float32)
    <tf.Tensor: shape=(2, 3), dtype=float32, numpy=
    array([[0., 0., 0.],
           [0., 0., 0.]], dtype=float32)>

  Args:
    tensor: A `Tensor`.
    dtype: A type for the returned `Tensor`. Must be `float16`, `float32`,
      `float64`, `int8`, `uint8`, `int16`, `uint16`, `int32`, `int64`,
      `complex64`, `complex128`, `bool` or `string`. (optional)
    name: A name for the operation (optional).
    optimize: if `True`, attempt to statically determine the shape of `tensor`
      and encode it as a constant. (optional, defaults to `True`)

  Returns:
    A `Tensor` with all elements set to zero.
  """
  return zeros_like_impl(tensor, dtype, name, optimize)


@tf_export("zeros_like", v1=[])
@dispatch.register_unary_elementwise_api
@dispatch.add_dispatch_support
def zeros_like_v2(
    input,  # pylint: disable=redefined-builtin
    dtype=None,
    name=None,
    layout=None,
):
  """Creates a tensor with all elements set to zero.

  See also `tf.zeros`.

  Given a single tensor or array-like object (`input`), this operation returns
  a tensor of the same type and shape as `input` with all elements set to zero.
  Optionally, you can use `dtype` to specify a new type for the returned tensor.

  Note that the layout of the input tensor is not preserved if the op
  is used inside tf.function. To obtain a tensor with the same layout as the
  input, chain the returned value to a `dtensor.relayout_like`.

  Examples:

    >>> tensor = tf.constant([[1, 2, 3], [4, 5, 6]])
    >>> tf.zeros_like(tensor)
    <tf.Tensor: shape=(2, 3), dtype=int32, numpy=
    array([[0, 0, 0],
           [0, 0, 0]], dtype=int32)>

    >>> tf.zeros_like(tensor, dtype=tf.float32)
    <tf.Tensor: shape=(2, 3), dtype=float32, numpy=
    array([[0., 0., 0.],
           [0., 0., 0.]], dtype=float32)>

    >>> tf.zeros_like([[1, 2, 3], [4, 5, 6]])
    <tf.Tensor: shape=(2, 3), dtype=int32, numpy=
    array([[0, 0, 0],
           [0, 0, 0]], dtype=int32)>

  Args:
    input: A `Tensor` or array-like object.
    dtype: A type for the returned `Tensor`. Must be `float16`, `float32`,
      `float64`, `int8`, `uint8`, `int16`, `uint16`, `int32`, `int64`,
      `complex64`, `complex128`, `bool` or `string` (optional).
    name: A name for the operation (optional).
    layout: Optional, `tf.experimental.dtensor.Layout`. If provided, the result
      is a [DTensor](https://www.tensorflow.org/guide/dtensor_overview) with the
      provided layout.

  Returns:
    A `Tensor` with all elements set to zero.
  """
  return zeros_like_impl(input, dtype, name, optimize=True, layout=layout)


def array_like_impl(
    array_fn, array_like_fn, tensor, dtype, name, optimize=True, layout=None
):
  """Internal implementation for ones_like and zeros_like API calls."""
  # Note: the output follows the layout argument, not the layout of the input.
  if not tensor_util.is_tf_type(tensor):
    tensor = ops.convert_to_tensor(tensor, name="tensor")
  tensor_shape = tensor.shape
  tensor_dtype = tensor.dtype

  if context.executing_eagerly():
    if dtype is not None and dtype != tensor_dtype:
      return array_fn(
          shape_internal(tensor, optimize=optimize),
          dtype=dtype,
          name=name,
          layout=layout,
      )
    return d_api.call_with_layout(array_like_fn, layout, tensor, name=name)

  # For now, variant types must be created via array_like_fn; as we need to
  # pass the input variant object to the proper array_fn callback.

  if (
      optimize
      and tensor_shape.is_fully_defined()
      and tensor_dtype != dtypes.variant
  ):
    # We can produce a result tensor independent of the value of 'tensor',
    # since the shape is known statically.
    return array_fn(
        tensor_shape, dtype=dtype or tensor_dtype, name=name, layout=layout
    )

  if dtype is not None and dtype != tensor_dtype and dtype != dtypes.variant:
    return array_fn(
        shape_internal(tensor, optimize=optimize),
        dtype=dtype,
        name=name,
        layout=layout,
    )
  return d_api.call_with_layout(array_like_fn, layout, tensor, name=name)


@_tag_zeros_tensor
def zeros_like_impl(tensor, dtype, name, optimize=True, layout=None):
  """Internal implementation for the v1/v2 zeros_like API calls."""
  with ops.name_scope(name, "zeros_like", [tensor]) as name:
    return array_like_impl(
        zeros,
        gen_array_ops.zeros_like,
        tensor,
        dtype,
        name,
        optimize=optimize,
        layout=layout,
    )


@tf_export(v1=["ones_like"])
@dispatch.register_unary_elementwise_api
@dispatch.add_dispatch_support
def ones_like(tensor, dtype=None, name=None, optimize=True):
  """Creates a tensor with all elements set to 1.

  See also `tf.ones`.

  Given a single tensor (`tensor`), this operation returns a tensor of the same
  type and shape as `tensor` with all elements set to 1. Optionally, you can
  specify a new type (`dtype`) for the returned tensor.

  For example:

  ```python
  tensor = tf.constant([[1, 2, 3], [4, 5, 6]])
  tf.ones_like(tensor)  # [[1, 1, 1], [1, 1, 1]]
  ```

  Args:
    tensor: A `Tensor`.
    dtype: A type for the returned `Tensor`. Must be `float32`, `float64`,
      `int8`, `uint8`, `int16`, `uint16`, `int32`, `int64`, `complex64`,
      `complex128` or `bool`.
    name: A name for the operation (optional).
    optimize: if true, attempt to statically determine the shape of 'tensor' and
      encode it as a constant.

  Returns:
    A `Tensor` with all elements set to 1.
  """
  return ones_like_impl(tensor, dtype, name, optimize)


@tf_export("ones_like", v1=[])
@dispatch.register_unary_elementwise_api
@dispatch.add_dispatch_support
def ones_like_v2(
    input,  # pylint: disable=redefined-builtin
    dtype=None,
    name=None,
    layout=None,
):
  """Creates a tensor of all ones that has the same shape as the input.

  See also `tf.ones`.

  Given a single tensor (`tensor`), this operation returns a tensor of the
  same type and shape as `tensor` with all elements set to 1. Optionally,
  you can use `dtype` to specify a new type for the returned tensor.

  For example:

  >>> tensor = tf.constant([[1, 2, 3], [4, 5, 6]])
  >>> tf.ones_like(tensor)
  <tf.Tensor: shape=(2, 3), dtype=int32, numpy=
    array([[1, 1, 1],
           [1, 1, 1]], dtype=int32)>

  Note that the layout of the input tensor is not preserved if the op
  is used inside tf.function. To obtain a tensor with the same layout as the
  input, chain the returned value to a `dtensor.relayout_like`.

  Args:
    input: A `Tensor`.
    dtype: A type for the returned `Tensor`. Must be `float16`, `float32`,
      `float64`, `int8`, `uint8`, `int16`, `uint16`, `int32`, `int64`,
      `complex64`, `complex128`, `bool` or `string`.
    name: A name for the operation (optional).
    layout: Optional, `tf.experimental.dtensor.Layout`. If provided, the result
      is a [DTensor](https://www.tensorflow.org/guide/dtensor_overview) with the
      provided layout.

  Returns:
    A `Tensor` with all elements set to one.
  """
  with ops.name_scope(name, "ones_like", [input]) as name:
    return array_like_impl(
        ones,
        gen_array_ops.ones_like,
        input,
        dtype,
        name,
        optimize=True,
        layout=layout,
    )


def ones_like_impl(tensor, dtype, name, optimize=True, layout=None):
  """Internal implementation for the v1/v2 ones_like API calls."""
  with ops.name_scope(name, "ones_like", [tensor]) as name:

    tensor = ops.convert_to_tensor(tensor, name="tensor")
    ones_shape = shape_internal(tensor, optimize=optimize)
    if dtype is None:
      dtype = tensor.dtype
    ret = ones(ones_shape, dtype=dtype, name=name, layout=layout)
    if not context.executing_eagerly():
      ret.set_shape(tensor.get_shape())
    return ret


@tf_export("ones")
@dispatch.add_dispatch_support
def ones(shape, dtype=dtypes.float32, name=None, layout=None):
  """Creates a tensor with all elements set to one (1).

  See also `tf.ones_like`, `tf.zeros`, `tf.fill`, `tf.eye`.

  This operation returns a tensor of type `dtype` with shape `shape` and
  all elements set to one.

  >>> tf.ones([3, 4], tf.int32)
  <tf.Tensor: shape=(3, 4), dtype=int32, numpy=
  array([[1, 1, 1, 1],
         [1, 1, 1, 1],
         [1, 1, 1, 1]], dtype=int32)>

  Args:
    shape: A `list` of integers, a `tuple` of integers, or a 1-D `Tensor` of
      type `int32`.
    dtype: Optional DType of an element in the resulting `Tensor`. Default is
      `tf.float32`.
    name: Optional string. A name for the operation.
    layout: Optional, `tf.experimental.dtensor.Layout`. If provided, the result
      is a [DTensor](https://www.tensorflow.org/guide/dtensor_overview) with the
      provided layout.

  Returns:
    A `Tensor` with all elements set to one (1).
  """
  dtype = dtypes.as_dtype(dtype).base_dtype
  with ops.name_scope(name, "ones", [shape]) as name:
    if dtype == dtypes.bool:
      one = True
    elif dtype.is_quantized:
      one = np.ones([]).astype(dtype.as_numpy_dtype)
    else:
      one = 1
    if not isinstance(shape, tensor_lib.Tensor):
      try:
        if not context.executing_eagerly():
          # Create a constant if it won't be very big. Otherwise, create a fill
          # op to prevent serialized GraphDefs from becoming too large.
          output = _constant_if_small(one, shape, dtype, name, layout=layout)
          if output is not None:
            return output

        # Go through tensor shapes to get int64-if-needed semantics
        shape = constant_op._tensor_shape_tensor_conversion_function(
            tensor_shape.TensorShape(shape))
      except (TypeError, ValueError):
        # Happens when shape is a list with tensor elements
        shape = ops.convert_to_tensor(shape, dtype=dtypes.int32)
    if not shape._shape_tuple():
      shape = reshape(shape, [-1])  # Ensure it's a vector
    output = fill(shape, constant(one, dtype=dtype), name=name, layout=layout)
  assert output.dtype.base_dtype == dtype
  return output


@tf_export(v1=["placeholder"])
def placeholder(dtype, shape=None, name=None):
  """Inserts a placeholder for a tensor that will be always fed.

  **Important**: This tensor will produce an error if evaluated. Its value must
  be fed using the `feed_dict` optional argument to `Session.run()`,
  `Tensor.eval()`, or `Operation.run()`.

  For example:

  ```python
  x = tf.compat.v1.placeholder(tf.float32, shape=(1024, 1024))
  y = tf.matmul(x, x)

  with tf.compat.v1.Session() as sess:
    print(sess.run(y))  # ERROR: will fail because x was not fed.

    rand_array = np.random.rand(1024, 1024)
    print(sess.run(y, feed_dict={x: rand_array}))  # Will succeed.
  ```

  Args:
    dtype: The type of elements in the tensor to be fed.
    shape: The shape of the tensor to be fed (optional). If the shape is not
      specified, you can feed a tensor of any shape.
    name: A name for the operation (optional).

  Returns:
    A `Tensor` that may be used as a handle for feeding a value, but not
    evaluated directly.

  Raises:
    RuntimeError: if eager execution is enabled

  @compatibility(TF2)
  This API is not compatible with eager execution and `tf.function`. To migrate
  to TF2, rewrite the code to be compatible with eager execution. Check the
  [migration
  guide](https://www.tensorflow.org/guide/migrate#1_replace_v1sessionrun_calls)
  on replacing `Session.run` calls. In TF2, you can just pass tensors directly
  into ops and layers. If you want to explicitly set up your inputs, also see
  [Keras functional API](https://www.tensorflow.org/guide/keras/functional) on
  how to use `tf.keras.Input` to replace `tf.compat.v1.placeholder`.
  `tf.function` arguments also do the job of `tf.compat.v1.placeholder`.
  For more details please read [Better
  performance with tf.function](https://www.tensorflow.org/guide/function).
  @end_compatibility
  """
  if context.executing_eagerly():
    raise RuntimeError("tf.placeholder() is not compatible with "
                       "eager execution.")

  return gen_array_ops.placeholder(dtype=dtype, shape=shape, name=name)


@tf_export(v1=["placeholder_with_default"])
def placeholder_with_default(input, shape, name=None):  # pylint: disable=redefined-builtin
  """A placeholder op that passes through `input` when its output is not fed.

  @compatibility(TF2)
  This API is strongly discouraged for use with eager execution and
  `tf.function`. The primary use of this API is for testing computation wrapped
  within a `tf.function` where the input tensors might not have statically known
  fully-defined shapes. The same can be achieved by creating a
  [concrete function](
  https://www.tensorflow.org/guide/function#obtaining_concrete_functions)
  from the `tf.function` with a `tf.TensorSpec` input which has partially
  defined shapes. For example, the code

  >>> @tf.function
  ... def f():
  ...   x = tf.compat.v1.placeholder_with_default(
  ...       tf.constant([[1., 2., 3.], [4., 5., 6.]]), [None, 3])
  ...   y = tf.constant([[1.],[2.], [3.]])
  ...   z = tf.matmul(x, y)
  ...   assert z.shape[0] == None
  ...   assert z.shape[1] == 1

  >>> f()

  can easily be replaced by

  >>> @tf.function
  ... def f(x):
  ...   y = tf.constant([[1.],[2.], [3.]])
  ...   z = tf.matmul(x, y)
  ...   assert z.shape[0] == None
  ...   assert z.shape[1] == 1

  >>> g = f.get_concrete_function(tf.TensorSpec([None, 3]))

  You can learn more about `tf.function` at [Better
  performance with tf.function](https://www.tensorflow.org/guide/function).
  @end_compatibility

  Args:
    input: A `Tensor`. The default value to produce when output is not fed.
    shape: A `tf.TensorShape` or list of `int`s. The (possibly partial) shape of
      the tensor.
    name: A name for the operation (optional).

  Returns:
    A `Tensor`. Has the same type as `input`.
  """
  return gen_array_ops.placeholder_with_default(input, shape, name)


@tf_export(v1=["sparse.placeholder", "sparse_placeholder"])
@deprecation.deprecated_endpoints("sparse_placeholder")
def sparse_placeholder(dtype, shape=None, name=None):
  """Inserts a placeholder for a sparse tensor that will be always fed.

  **Important**: This sparse tensor will produce an error if evaluated.
  Its value must be fed using the `feed_dict` optional argument to
  `Session.run()`, `Tensor.eval()`, or `Operation.run()`.

  For example:

  ```python
  x = tf.compat.v1.sparse.placeholder(tf.float32)
  y = tf.sparse.reduce_sum(x)

  with tf.compat.v1.Session() as sess:
    print(sess.run(y))  # ERROR: will fail because x was not fed.

    indices = np.array([[3, 2, 0], [4, 5, 1]], dtype=np.int64)
    values = np.array([1.0, 2.0], dtype=np.float32)
    shape = np.array([7, 9, 2], dtype=np.int64)
    print(sess.run(y, feed_dict={
      x: tf.compat.v1.SparseTensorValue(indices, values, shape)}))  # Will
      succeed.
    print(sess.run(y, feed_dict={
      x: (indices, values, shape)}))  # Will succeed.

    sp = tf.sparse.SparseTensor(indices=indices, values=values,
                                dense_shape=shape)
    sp_value = sp.eval(session=sess)
    print(sess.run(y, feed_dict={x: sp_value}))  # Will succeed.
  ```


  Args:
    dtype: The type of `values` elements in the tensor to be fed.
    shape: The shape of the tensor to be fed (optional). If the shape is not
      specified, you can feed a sparse tensor of any shape.
    name: A name for prefixing the operations (optional).

  Returns:
    A `SparseTensor` that may be used as a handle for feeding a value, but not
    evaluated directly.

  Raises:
    RuntimeError: if eager execution is enabled

  @compatibility(TF2)
  This API is not compatible with eager execution and `tf.function`. To migrate
  to TF2, rewrite the code to be compatible with eager execution. Check the
  [migration
  guide](https://www.tensorflow.org/guide/migrate#1_replace_v1sessionrun_calls)
  on replacing `Session.run` calls. In TF2, you can just pass tensors directly
  into ops and layers. If you want to explicitly set up your inputs, also see
  [Keras functional API](https://www.tensorflow.org/guide/keras/functional) on
  how to use `tf.keras.Input` to replace `tf.compat.v1.sparse_placeholder`.
  `tf.function` arguments also do the job of `tf.compat.v1.sparse_placeholder`.
  For more details please read [Better
  performance with tf.function](https://www.tensorflow.org/guide/function).
  @end_compatibility
  """
  if context.executing_eagerly():
    raise RuntimeError("`sparse_placeholder` is not compatible with "
                       "eager execution.")

  shape_name = (name + "/shape") if name is not None else None
  default_shape_name = (name + "/shape_default") if name is not None else None
  if shape is None:
    rank = None
    dense_shape = placeholder(dtypes.int64, shape=[rank], name=shape_name)
    dense_shape_default = tensor_util.constant_value_as_shape(dense_shape)
  else:
    if isinstance(shape, tensor_lib.Tensor):
      rank = shape.get_shape()[0]
      dense_shape_default = tensor_util.constant_value_as_shape(shape)
    else:
      rank = len(shape)
      # determine the shape, to override the `.shape` property of the
      # `SparseTensor`
      dense_shape_default = tensor_shape.TensorShape(
          tuple(None if dim == -1 else dim for dim in shape))
      shape = tuple(tensor_shape.dimension_value(dim) for dim in shape)
      shape = tuple(-1 if dim is None else dim for dim in shape)
      shape = ops.convert_to_tensor(
          shape, dtype=dtypes.int64, name=default_shape_name)

    # `dense_shape` needs to be feedable (for users that treat this as an
    # actual placeholder). `constant_value_as_shape` sets constants to
    # not-feedable. placeholder_with_default works, but blocks `SparseTensor`
    # from reading the default value back out.
    dense_shape = placeholder_with_default(
        shape, shape=shape.shape, name=shape_name)

  result = sparse_tensor.SparseTensor(
      values=placeholder(
          dtype,
          shape=[None],
          name=(name + "/values") if name is not None else None),
      indices=placeholder(
          dtypes.int64,
          shape=[None, rank],
          name=(name + "/indices") if name is not None else None),
      dense_shape=dense_shape)

  # Now the SparseTensor.shape is a list of `None`s, since it couldn't read the
  # default shape out of the placeholder. Override that
  # shape to be the value determined here, so partial shapes can be
  # propagated.
  result.set_shape(dense_shape_default)
  return result

# pylint: enable=redefined-outer-name


@tf_export("pad", v1=[])
@dispatch.add_dispatch_support
def pad_v2(tensor, paddings, mode="CONSTANT", constant_values=0, name=None):
  """Pads a tensor.

  This operation pads a `tensor` according to the `paddings` you specify.
  `paddings` is an integer tensor with shape `[n, 2]`, where n is the rank of
  `tensor`. For each dimension D of `input`, `paddings[D, 0]` indicates how
  many values to add before the contents of `tensor` in that dimension, and
  `paddings[D, 1]` indicates how many values to add after the contents of
  `tensor` in that dimension. If `mode` is "REFLECT" then both `paddings[D, 0]`
  and `paddings[D, 1]` must be no greater than `tensor.dim_size(D) - 1`. If
  `mode` is "SYMMETRIC" then both `paddings[D, 0]` and `paddings[D, 1]` must be
  no greater than `tensor.dim_size(D)`.

  The padded size of each dimension D of the output is:

  `paddings[D, 0] + tensor.dim_size(D) + paddings[D, 1]`

  For example:

  ```python
  t = tf.constant([[1, 2, 3], [4, 5, 6]])
  paddings = tf.constant([[1, 1,], [2, 2]])
  # 'constant_values' is 0.
  # rank of 't' is 2.
  tf.pad(t, paddings, "CONSTANT")  # [[0, 0, 0, 0, 0, 0, 0],
                                   #  [0, 0, 1, 2, 3, 0, 0],
                                   #  [0, 0, 4, 5, 6, 0, 0],
                                   #  [0, 0, 0, 0, 0, 0, 0]]

  tf.pad(t, paddings, "REFLECT")  # [[6, 5, 4, 5, 6, 5, 4],
                                  #  [3, 2, 1, 2, 3, 2, 1],
                                  #  [6, 5, 4, 5, 6, 5, 4],
                                  #  [3, 2, 1, 2, 3, 2, 1]]

  tf.pad(t, paddings, "SYMMETRIC")  # [[2, 1, 1, 2, 3, 3, 2],
                                    #  [2, 1, 1, 2, 3, 3, 2],
                                    #  [5, 4, 4, 5, 6, 6, 5],
                                    #  [5, 4, 4, 5, 6, 6, 5]]
  ```

  Args:
    tensor: A `Tensor`.
    paddings: A `Tensor` of type `int32`.
    mode: One of "CONSTANT", "REFLECT", or "SYMMETRIC" (case-insensitive)
    constant_values: In "CONSTANT" mode, the scalar pad value to use. Must be
      same type as `tensor`.
    name: A name for the operation (optional).

  Returns:
    A `Tensor`. Has the same type as `tensor`.

  Raises:
    ValueError: When mode is not one of "CONSTANT", "REFLECT", or "SYMMETRIC".
  """
  return pad(tensor, paddings, mode, name, constant_values)


@tf_export(v1=["pad"])
@dispatch.add_dispatch_support
def pad(tensor, paddings, mode="CONSTANT", name=None, constant_values=0):  # pylint: disable=invalid-name
  """Pads a tensor.

  This operation pads a `tensor` according to the `paddings` you specify.
  `paddings` is an integer tensor with shape `[n, 2]`, where n is the rank of
  `tensor`. For each dimension D of `input`, `paddings[D, 0]` indicates how
  many values to add before the contents of `tensor` in that dimension, and
  `paddings[D, 1]` indicates how many values to add after the contents of
  `tensor` in that dimension. If `mode` is "REFLECT" then both `paddings[D, 0]`
  and `paddings[D, 1]` must be no greater than `tensor.dim_size(D) - 1`. If
  `mode` is "SYMMETRIC" then both `paddings[D, 0]` and `paddings[D, 1]` must be
  no greater than `tensor.dim_size(D)`.

  The padded size of each dimension D of the output is:

  `paddings[D, 0] + tensor.dim_size(D) + paddings[D, 1]`

  For example:

  ```python
  t = tf.constant([[1, 2, 3], [4, 5, 6]])
  paddings = tf.constant([[1, 1,], [2, 2]])
  # 'constant_values' is 0.
  # rank of 't' is 2.
  tf.pad(t, paddings, "CONSTANT")  # [[0, 0, 0, 0, 0, 0, 0],
                                   #  [0, 0, 1, 2, 3, 0, 0],
                                   #  [0, 0, 4, 5, 6, 0, 0],
                                   #  [0, 0, 0, 0, 0, 0, 0]]

  tf.pad(t, paddings, "REFLECT")  # [[6, 5, 4, 5, 6, 5, 4],
                                  #  [3, 2, 1, 2, 3, 2, 1],
                                  #  [6, 5, 4, 5, 6, 5, 4],
                                  #  [3, 2, 1, 2, 3, 2, 1]]

  tf.pad(t, paddings, "SYMMETRIC")  # [[2, 1, 1, 2, 3, 3, 2],
                                    #  [2, 1, 1, 2, 3, 3, 2],
                                    #  [5, 4, 4, 5, 6, 6, 5],
                                    #  [5, 4, 4, 5, 6, 6, 5]]
  ```

  Args:
    tensor: A `Tensor`.
    paddings: A `Tensor` of type `int32`.
    mode: One of "CONSTANT", "REFLECT", or "SYMMETRIC" (case-insensitive)
    name: A name for the operation (optional).
    constant_values: In "CONSTANT" mode, the scalar pad value to use. Must be
      same type as `tensor`.

  Returns:
    A `Tensor`. Has the same type as `tensor`.

  Raises:
    ValueError: When mode is not one of "CONSTANT", "REFLECT", or "SYMMETRIC".
  """

  # Convert lower/mixed case to upper for NumPy compatibility
  # NumPy uses all lower-case modes.
  mode = mode.upper()
  if mode == "CONSTANT":
    # TODO(rjryan): Once the forward compatibility period (3 weeks) have passed
    # remove the "Pad" fallback here.
    if (not tensor_util.is_tf_type(constant_values) and
        np.ndim(constant_values) == 0 and
        constant_values == np.zeros_like(constant_values)):
      result = gen_array_ops.pad(tensor, paddings, name=name)
    else:
      result = gen_array_ops.pad_v2(
          tensor, paddings, constant_values, name=name)
  elif mode == "REFLECT":
    result = gen_array_ops.mirror_pad(
        tensor, paddings, mode="REFLECT", name=name)
  elif mode == "SYMMETRIC":
    result = gen_array_ops.mirror_pad(
        tensor, paddings, mode="SYMMETRIC", name=name)
  else:
    raise ValueError("Value of argument `mode` expected to be "
                     """one of "CONSTANT", "REFLECT", or "SYMMETRIC". """
                     f"Received `mode` = {mode}")

  # Restore shape information where possible.
  if not context.executing_eagerly():
    paddings_constant = _get_paddings_constant(paddings)
    input_shape = (
        tensor_shape.TensorShape(tensor.shape)
        if isinstance(tensor, tensor_lib.Tensor) else result.op.inputs[0].shape)
    if (input_shape.ndims is not None and
        not result.shape.is_fully_defined() and paddings_constant is not None):
      new_shape = []
      for padding, dim in zip(paddings_constant, input_shape.as_list()):
        if padding is None or dim is None or any((x is None for x in padding)):
          new_shape.append(None)
        else:
          new_shape.append(sum(padding) + dim)
      result.set_shape(new_shape)

  return result


def _get_paddings_constant(paddings):
  """Helper to get the constant values of the paddings arg to pad().

  Used under V1 graph mode to facilitate computation of the shape of the output
  tensor of `pad()`.

  Args:
    paddings: The same paddings arg as passed to pad(). Can be a Tensor, or
      a nested list or tuple of Tensor and/or numbers.

  Returns:
    A nested list or numbers or `None`, in which `None` indicates unknown
    padding size.
  """
  if isinstance(paddings, tensor_lib.Tensor):
    return tensor_util.constant_value(paddings, partial=True)
  elif isinstance(paddings, (list, tuple)):
    return [_get_paddings_constant(x) for x in paddings]
  else:
    return paddings


@tf_export("meshgrid")
@dispatch.add_dispatch_support
def meshgrid(*args, **kwargs):
  """Broadcasts parameters for evaluation on an N-D grid.

  Given N one-dimensional coordinate arrays `*args`, returns a list `outputs`
  of N-D coordinate arrays for evaluating expressions on an N-D grid.

  Notes:

  `meshgrid` supports cartesian ('xy') and matrix ('ij') indexing conventions.
  When the `indexing` argument is set to 'xy' (the default), the broadcasting
  instructions for the first two dimensions are swapped.

  Examples:

  Calling `X, Y = meshgrid(x, y)` with the tensors

  ```python
  x = [1, 2, 3]
  y = [4, 5, 6]
  X, Y = tf.meshgrid(x, y)
  # X = [[1, 2, 3],
  #      [1, 2, 3],
  #      [1, 2, 3]]
  # Y = [[4, 4, 4],
  #      [5, 5, 5],
  #      [6, 6, 6]]
  ```

  Args:
    *args: `Tensor`s with rank 1.
    **kwargs:
      - indexing: Either 'xy' or 'ij' (optional, default: 'xy').
      - name: A name for the operation (optional).

  Returns:
    outputs: A list of N `Tensor`s with rank N.

  Raises:
    TypeError: When no keyword arguments (kwargs) are passed.
    ValueError: When indexing keyword argument is not one of `xy` or `ij`.
  """

  indexing = kwargs.pop("indexing", "xy")
  name = kwargs.pop("name", "meshgrid")
  if kwargs:
    key = list(kwargs.keys())[0]
    raise TypeError("'{}' is an invalid keyword argument "
                    "for this function".format(key))

  if indexing not in ("xy", "ij"):
    raise ValueError("Argument `indexing` parameter must be either "
                     f"'xy' or 'ij', got '{indexing}'")

  with ops.name_scope(name, "meshgrid", args) as name:
    ndim = len(args)
    s0 = (1,) * ndim

    if not ndim:
      return []

    # Prepare reshape by inserting dimensions with size 1 where needed
    output = []
    for i, x in enumerate(args):
      output.append(
          reshape(array_ops_stack.stack(x), (s0[:i] + (-1,) + s0[i + 1::])))
    # Create parameters for broadcasting each tensor to the full size
    shapes = [size(x) for x in args]

    output_dtype = ops.convert_to_tensor(args[0]).dtype.base_dtype

    if indexing == "xy" and ndim > 1:
      output[0] = reshape(output[0], (1, -1) + (1,) * (ndim - 2))
      output[1] = reshape(output[1], (-1, 1) + (1,) * (ndim - 2))
      shapes[0], shapes[1] = shapes[1], shapes[0]

    # TODO(nolivia): improve performance with a broadcast
    mult_fact = ones(shapes, output_dtype)
    return [x * mult_fact for x in output]


NEW_AXIS = -1
SHRINK_AXIS = -2


# PEP-8 naming
# pylint: disable=invalid-name,redefined-outer-name
def _compute_size_of_strided_dim(shrink, spec, size):
  """Computes the size of a single strided slice dimension."""

  unknown = None  # Document what None means here.
  use_full_range = None  # Document other use of None.
  # if this is a shrink axis (i.e. a non-range index)
  # it either will produce an error or return 1
  if shrink:
    return 1
  if size is unknown or size.value is unknown:
    return unknown
  size = size.value
  stride = spec.step
  if stride is not unknown:
    if stride == 0:
      return unknown
    stride = spec.step
    valid_range = [0, size] if stride > 0 else [-1, size - 1]

    # PEP-8 naming
    # pylint: disable=invalid-name
    def canonical(x, c):
      if x is use_full_range:
        return valid_range[c] if stride > 0 else valid_range[(c + 1) & 1]
      else:
        x_fwd = size + x if x < 0 else x  # make negative indices positive
        return max(valid_range[0], min(valid_range[1], x_fwd))

    begin = canonical(spec.start, 0)
    end = canonical(spec.stop, 1)
    interval_length = end - begin
    if interval_length == 0 or ((interval_length < 0) != (stride < 0)):
      return 0
    else:
      remainder = 1 if interval_length % stride != 0 else 0
      return interval_length // stride + remainder
  else:
    return unknown  # unknown because stride is unknown


def _TileGradShape(op: ops.Operation):
  """Shape function for the TileGrad op."""
  multiples_shape = op.inputs[1].get_shape().with_rank(1)
  input_shape = op.inputs[0].get_shape().with_rank(multiples_shape[0])
  # NOTE(mrry): Represent `multiples` as a `TensorShape` because (i)
  # it is a vector of non-negative integers, and (ii) doing so allows
  # us to handle partially-known multiples.
  multiples = tensor_util.constant_value_as_shape(op.inputs[1]).with_rank(
      input_shape.ndims)
  if multiples.ndims is None:
    return [tensor_shape.unknown_shape()]
  else:
    output_dims = []
    for dim, multiple in zip(input_shape.dims, multiples.dims):
      output_dims.append(dim // multiple)
    return [tensor_shape.TensorShape(output_dims)]


@tf_export("edit_distance")
@dispatch.add_dispatch_support
def edit_distance(hypothesis, truth, normalize=True, name="edit_distance"):
  """Computes the Levenshtein distance between sequences.

  This operation takes variable-length sequences (`hypothesis` and `truth`),
  each provided as a `SparseTensor`, and computes the Levenshtein distance.
  You can normalize the edit distance by length of `truth` by setting
  `normalize` to true.

  For example:

  Given the following input,
  * `hypothesis` is a `tf.SparseTensor` of shape `[2, 1, 1]`
  * `truth` is a `tf.SparseTensor` of shape `[2, 2, 2]`

  >>> hypothesis = tf.SparseTensor(
  ...   [[0, 0, 0],
  ...    [1, 0, 0]],
  ...   ["a", "b"],
  ...   (2, 1, 1))
  >>> truth = tf.SparseTensor(
  ...   [[0, 1, 0],
  ...    [1, 0, 0],
  ...    [1, 0, 1],
  ...    [1, 1, 0]],
  ...    ["a", "b", "c", "a"],
  ...    (2, 2, 2))
  >>> tf.edit_distance(hypothesis, truth, normalize=True)
  <tf.Tensor: shape=(2, 2), dtype=float32, numpy=
  array([[inf, 1. ],
         [0.5, 1. ]], dtype=float32)>

  The operation returns a dense Tensor of shape `[2, 2]` with
  edit distances normalized by `truth` lengths.

  **Note**: It is possible to calculate edit distance between two
  sparse tensors with variable-length values. However, attempting to create
  them while eager execution is enabled will result in a `ValueError`.

  For the following  inputs,

  ```python
  # 'hypothesis' is a tensor of shape `[2, 1]` with variable-length values:
  #   (0,0) = ["a"]
  #   (1,0) = ["b"]
  hypothesis = tf.sparse.SparseTensor(
      [[0, 0, 0],
       [1, 0, 0]],
      ["a", "b"],
      (2, 1, 1))

  # 'truth' is a tensor of shape `[2, 2]` with variable-length values:
  #   (0,0) = []
  #   (0,1) = ["a"]
  #   (1,0) = ["b", "c"]
  #   (1,1) = ["a"]
  truth = tf.sparse.SparseTensor(
      [[0, 1, 0],
       [1, 0, 0],
       [1, 0, 1],
       [1, 1, 0]],
      ["a", "b", "c", "a"],
      (2, 2, 2))

  normalize = True

  # The output would be a dense Tensor of shape `(2,)`, with edit distances
  normalized by 'truth' lengths.
  # output => array([0., 0.5], dtype=float32)
  ```

  Args:
    hypothesis: A `SparseTensor` containing hypothesis sequences.
    truth: A `SparseTensor` containing truth sequences.
    normalize: A `bool`. If `True`, normalizes the Levenshtein distance by
      length of `truth.`
    name: A name for the operation (optional).

  Returns:
    A dense `Tensor` with rank `R - 1`, where R is the rank of the
    `SparseTensor` inputs `hypothesis` and `truth`.

  Raises:
    TypeError: If either `hypothesis` or `truth` are not a `SparseTensor`.
  """
  if not isinstance(
      hypothesis,
      (sparse_tensor.SparseTensor, sparse_tensor.SparseTensorValue)):
    raise TypeError("Hypothesis must be a SparseTensor.")
  if not isinstance(
      truth, (sparse_tensor.SparseTensor, sparse_tensor.SparseTensorValue)):
    raise TypeError("Truth must be a SparseTensor.")

  return gen_array_ops.edit_distance(
      hypothesis.indices,
      hypothesis.values,
      hypothesis.dense_shape,
      truth.indices,
      truth.values,
      truth.dense_shape,
      normalize=normalize,
      name=name)


@ops.RegisterGradient("FakeQuantWithMinMaxArgs")
def _FakeQuantWithMinMaxArgsGradient(op: ops.Operation, grad):
  """Gradient for FakeQuantWithMinMaxArgs op."""
  return fake_quant_with_min_max_args_gradient(
      grad,
      op.inputs[0],
      min=op.get_attr("min"),
      max=op.get_attr("max"),
      num_bits=op.get_attr("num_bits"),
      narrow_range=op.get_attr("narrow_range"))


@ops.RegisterGradient("FakeQuantWithMinMaxVars")
def _FakeQuantWithMinMaxVarsGradient(op: ops.Operation, grad):
  """Gradient for FakeQuantWithMinMaxVars op."""
  return fake_quant_with_min_max_vars_gradient(
      grad,
      op.inputs[0],
      op.inputs[1],
      op.inputs[2],
      num_bits=op.get_attr("num_bits"),
      narrow_range=op.get_attr("narrow_range"))


@ops.RegisterGradient("FakeQuantWithMinMaxVarsPerChannel")
def _FakeQuantWithMinMaxVarsPerChannelGradient(op: ops.Operation, grad):
  """Gradient for FakeQuantWithMinMaxVarsPerChannel op."""
  return fake_quant_with_min_max_vars_per_channel_gradient(
      grad,
      op.inputs[0],
      op.inputs[1],
      op.inputs[2],
      num_bits=op.get_attr("num_bits"),
      narrow_range=op.get_attr("narrow_range"))


@ops.RegisterGradient("QuantizeAndDequantizeV4")
def _QuantizeAndDequantizeV4Grad(op: ops.Operation, grad):
  """Gradient for QuantizeAndDequantizeV4 op."""
  return quantize_and_dequantize_v4_grad(
      grad,
      op.inputs[0],
      op.inputs[1],
      op.inputs[2],
      axis=op.get_attr("axis"))


@ops.RegisterGradient("QuantizeAndDequantizeV4Grad")
def _QuantizeAndDequantizeV4GradGrad(op: ops.Operation, grad):
  """Gradient for QuantizeAndDequantizeV4Grad op."""
  return _QuantizeAndDequantizeV4Grad(op, grad)


@tf_export("required_space_to_batch_paddings")
def required_space_to_batch_paddings(input_shape,
                                     block_shape,
                                     base_paddings=None,
                                     name=None):
  """Calculate padding required to make block_shape divide input_shape.

  This function can be used to calculate a suitable paddings argument for use
  with space_to_batch_nd and batch_to_space_nd.

  Args:
    input_shape: int32 Tensor of shape [N].
    block_shape: int32 Tensor of shape [N].
    base_paddings: Optional int32 Tensor of shape [N, 2].  Specifies the minimum
      amount of padding to use.  All elements must be >= 0.  If not specified,
      defaults to 0.
    name: string.  Optional name prefix.

  Returns:
    (paddings, crops), where:

    `paddings` and `crops` are int32 Tensors of rank 2 and shape [N, 2]
    satisfying:

        paddings[i, 0] = base_paddings[i, 0].
        0 <= paddings[i, 1] - base_paddings[i, 1] < block_shape[i]
        (input_shape[i] + paddings[i, 0] + paddings[i, 1]) % block_shape[i] == 0

        crops[i, 0] = 0
        crops[i, 1] = paddings[i, 1] - base_paddings[i, 1]

  Raises: ValueError if called with incompatible shapes.
  """
  with ops.name_scope(name, "required_space_to_batch_paddings",
                      [input_shape, block_shape]):
    input_shape = ops.convert_to_tensor(
        input_shape, dtype=dtypes.int32, name="input_shape")
    block_shape = ops.convert_to_tensor(
        block_shape, dtype=dtypes.int32, name="block_shape")

    block_shape.get_shape().assert_is_fully_defined()
    block_shape.get_shape().assert_has_rank(1)
    num_block_dims = block_shape.get_shape().dims[0].value
    if num_block_dims == 0:
      return zeros([0, 2], dtypes.int32), zeros([0, 2], dtypes.int32)

    input_shape.get_shape().assert_is_compatible_with([num_block_dims])

    if base_paddings is not None:
      base_paddings = ops.convert_to_tensor(
          base_paddings, dtype=dtypes.int32, name="base_paddings")
      base_paddings.get_shape().assert_is_compatible_with([num_block_dims, 2])
    else:
      base_paddings = zeros([num_block_dims, 2], dtypes.int32)

    const_block_shape = tensor_util.constant_value(block_shape)
    const_input_shape = tensor_util.constant_value(input_shape)
    const_base_paddings = tensor_util.constant_value(base_paddings)
    if (const_block_shape is not None and const_input_shape is not None and
        const_base_paddings is not None):
      block_shape = const_block_shape
      input_shape = const_input_shape
      base_paddings = const_base_paddings

    # Use same expression for both constant and non-constant case.
    pad_start = base_paddings[:, 0]
    orig_pad_end = base_paddings[:, 1]
    full_input_shape = input_shape + pad_start + orig_pad_end
    pad_end_extra = (block_shape - full_input_shape % block_shape) % block_shape
    pad_end = orig_pad_end + pad_end_extra

    result_paddings = array_ops_stack.stack(
        [[pad_start[i], pad_end[i]] for i in range(num_block_dims)],
        name="paddings")
    result_crops = array_ops_stack.stack(
        [[0, pad_end_extra[i]] for i in range(num_block_dims)], name="crops")
    return result_paddings, result_crops


@tf_export(v1=["nn.space_to_batch", "space_to_batch"])
@dispatch.add_dispatch_support
@deprecation.deprecated_endpoints("space_to_batch")
def space_to_batch(  # pylint: disable=missing-docstring
    input,  # pylint: disable=redefined-builtin
    paddings,
    block_size=None,
    name=None,
    block_shape=None):  # pylint: disable=redefined-builtin
  block_size = deprecation.deprecated_argument_lookup("block_shape",
                                                      block_shape, "block_size",
                                                      block_size)
  result = space_to_batch_nd(
      input,
      paddings=paddings,
      block_shape=np.array([block_size, block_size], dtype=np.int64),
      name=name)
  result.set_shape(result.get_shape().with_rank(4))
  return result


space_to_batch.__doc__ = gen_array_ops.space_to_batch.__doc__


@tf_export("space_to_batch", "nn.space_to_batch", v1=[])
@dispatch.add_dispatch_support
def space_to_batch_v2(input, block_shape, paddings, name=None):  # pylint: disable=redefined-builtin
  return space_to_batch_nd(input, block_shape, paddings, name)


space_to_batch_v2.__doc__ = gen_array_ops.space_to_batch_nd.__doc__


@tf_export(v1=["nn.space_to_depth", "space_to_depth"])
@dispatch.add_dispatch_support
@deprecation.deprecated_endpoints("space_to_depth")
def space_to_depth(input, block_size, name=None, data_format="NHWC"):  # pylint: disable=redefined-builtin
  return gen_array_ops.space_to_depth(input, block_size, data_format, name=name)


space_to_depth.__doc__ = gen_array_ops.space_to_depth.__doc__


@tf_export("nn.space_to_depth", v1=[])
@dispatch.add_dispatch_support
def space_to_depth_v2(input, block_size, data_format="NHWC", name=None):  # pylint: disable=redefined-builtin
  return gen_array_ops.space_to_depth(input, block_size, data_format, name=name)


space_to_depth_v2.__doc__ = gen_array_ops.space_to_depth.__doc__


@tf_export(v1=["nn.depth_to_space", "depth_to_space"])
@dispatch.add_dispatch_support
@deprecation.deprecated_endpoints("depth_to_space")
def depth_to_space(input, block_size, name=None, data_format="NHWC"):  # pylint: disable=redefined-builtin
  return gen_array_ops.depth_to_space(input, block_size, data_format, name=name)


depth_to_space.__doc__ = gen_array_ops.depth_to_space.__doc__


@tf_export("nn.depth_to_space", v1=[])
@dispatch.add_dispatch_support
def depth_to_space_v2(input, block_size, data_format="NHWC", name=None):  # pylint: disable=redefined-builtin
  return gen_array_ops.depth_to_space(input, block_size, data_format, name=name)


depth_to_space_v2.__doc__ = gen_array_ops.depth_to_space.__doc__


@tf_export(v1=["batch_to_space"])
@dispatch.add_dispatch_support
def batch_to_space(input, crops, block_size, name=None, block_shape=None):  # pylint: disable=redefined-builtin,missing-docstring
  block_size = deprecation.deprecated_argument_lookup("block_shape",
                                                      block_shape, "block_size",
                                                      block_size)
  result = batch_to_space_nd(
      input,
      crops=crops,
      block_shape=np.array([block_size, block_size], dtype=np.int64),
      name=name)
  result.set_shape(result.get_shape().with_rank(4))
  return result


batch_to_space.__doc__ = gen_array_ops.batch_to_space.__doc__


@tf_export("batch_to_space", v1=[])
@dispatch.add_dispatch_support
def batch_to_space_v2(input, block_shape, crops, name=None):  # pylint: disable=redefined-builtin
  """BatchToSpace for N-D tensors of type T.

  This operation reshapes the "batch" dimension 0 into `M + 1` dimensions of
  shape `block_shape + [batch]`, interleaves these blocks back into the grid
  defined by the spatial dimensions `[1, ..., M]`, to obtain a result with the
  same rank as the input.  The spatial dimensions of this intermediate result
  are then optionally cropped according to `crops` to produce the output.  This
  is the reverse of SpaceToBatch (see `tf.space_to_batch`).

  Args:
    input: A N-D `Tensor` with shape `input_shape = [batch] + spatial_shape +
      remaining_shape`, where `spatial_shape` has M dimensions.
    block_shape: A 1-D `Tensor` with shape [M]. Must be one of the following
      types: `int32`, `int64`. All values must be >= 1. For backwards
      compatibility with TF 1.0, this parameter may be an int, in which case it
      is converted to
      `numpy.array([block_shape, block_shape],
      dtype=numpy.int64)`.
    crops: A  2-D `Tensor` with shape `[M, 2]`. Must be one of the
      following types: `int32`, `int64`. All values must be >= 0.
      `crops[i] = [crop_start, crop_end]` specifies the amount to crop from
      input dimension `i + 1`, which corresponds to spatial dimension `i`.
      It is required that
      `crop_start[i] + crop_end[i] <= block_shape[i] * input_shape[i + 1]`.
      This operation is equivalent to the following steps:
      1. Reshape `input` to `reshaped` of shape: [block_shape[0], ...,
        block_shape[M-1], batch / prod(block_shape), input_shape[1], ...,
        input_shape[N-1]]
      2. Permute dimensions of `reshaped` to produce `permuted` of shape
         [batch / prod(block_shape),  input_shape[1], block_shape[0], ...,
         input_shape[M], block_shape[M-1], input_shape[M+1],
        ..., input_shape[N-1]]
      3. Reshape `permuted` to produce `reshaped_permuted` of shape
         [batch / prod(block_shape), input_shape[1] * block_shape[0], ...,
         input_shape[M] * block_shape[M-1], input_shape[M+1], ...,
         input_shape[N-1]]
      4. Crop the start and end of dimensions `[1, ..., M]` of
         `reshaped_permuted` according to `crops` to produce the output
         of shape:
         [batch / prod(block_shape),  input_shape[1] *
           block_shape[0] - crops[0,0] - crops[0,1], ..., input_shape[M] *
           block_shape[M-1] - crops[M-1,0] - crops[M-1,1],  input_shape[M+1],
           ..., input_shape[N-1]]
    name: A name for the operation (optional).

  Examples:

  1. For the following input of shape `[4, 1, 1, 1]`,
     `block_shape = [2, 2]`, and `crops = [[0, 0], [0, 0]]`:

     ```python
     [[[[1]]],
      [[[2]]],
      [[[3]]],
      [[[4]]]]
     ```

    The output tensor has shape `[1, 2, 2, 1]` and value:

     ```
     x = [[[[1], [2]],
         [[3], [4]]]]
     ```

  2. For the following input of shape `[4, 1, 1, 3]`,
     `block_shape = [2, 2]`, and `crops = [[0, 0], [0, 0]]`:

     ```python
     [[[1,  2,   3]],
      [[4,  5,   6]],
      [[7,  8,   9]],
      [[10, 11, 12]]]
     ```

    The output tensor has shape `[1, 2, 2, 3]` and value:

    ```python
     x = [[[[1, 2, 3], [4,  5,  6 ]],
           [[7, 8, 9], [10, 11, 12]]]]
     ```

  3. For the following
     input of shape `[4, 2, 2, 1]`,
     `block_shape = [2, 2]`, and `crops = [[0, 0], [0, 0]]`:

     ```python
     x = [[[[1], [3]], [[ 9], [11]]],
          [[[2], [4]], [[10], [12]]],
          [[[5], [7]], [[13], [15]]],
          [[[6], [8]], [[14], [16]]]]
     ```

    The output tensor has shape `[1, 4, 4, 1]` and value:

    ```python
     x = [[[1],  [2],  [ 3], [ 4]],
          [[5],  [6],  [ 7], [ 8]],
          [[9],  [10], [11], [12]],
          [[13], [14], [15], [16]]]
     ```

  4. For the following input of shape
      `[8, 1, 3, 1]`,
      `block_shape = [2, 2]`, and `crops = [[0, 0], [2, 0]]`:

      ```python
      x = [[[[0], [ 1], [ 3]]],
           [[[0], [ 9], [11]]],
           [[[0], [ 2], [ 4]]],
           [[[0], [10], [12]]],
           [[[0], [ 5], [ 7]]],
           [[[0], [13], [15]]],
           [[[0], [ 6], [ 8]]],
           [[[0], [14], [16]]]]
      ```

      The output tensor has shape `[2, 2, 4, 1]` and value:

      ```python
      x = [[[[ 1], [ 2], [ 3], [ 4]],
            [[ 5], [ 6], [ 7], [ 8]]],
           [[[ 9], [10], [11], [12]],
            [[13], [14], [15], [16]]]]
      ```

  Returns:
    A `Tensor`. Has the same type as `input`.
  """
  if isinstance(block_shape, int):
    block_shape = np.array([block_shape, block_shape], dtype=np.int64)

  return batch_to_space_nd(
      input=input, block_shape=block_shape, crops=crops, name=name)


@tf_export("one_hot")
@dispatch.add_dispatch_support
def one_hot(indices,
            depth,
            on_value=None,
            off_value=None,
            axis=None,
            dtype=None,
            name=None):
  """Returns a one-hot tensor.

  See also `tf.fill`, `tf.eye`.

  The locations represented by indices in `indices` take value `on_value`,
  while all other locations take value `off_value`.

  `on_value` and `off_value` must have matching data types. If `dtype` is also
  provided, they must be the same data type as specified by `dtype`.

  If `on_value` is not provided, it will default to the value `1` with type
  `dtype`

  If `off_value` is not provided, it will default to the value `0` with type
  `dtype`

  If the input `indices` is rank `N`, the output will have rank `N+1`. The
  new axis is created at dimension `axis` (default: the new axis is appended
  at the end).

  If `indices` is a scalar the output shape will be a vector of length `depth`

  If `indices` is a vector of length `features`, the output shape will be:

  ```
    features x depth if axis == -1
    depth x features if axis == 0
  ```

  If `indices` is a matrix (batch) with shape `[batch, features]`, the output
  shape will be:

  ```
    batch x features x depth if axis == -1
    batch x depth x features if axis == 1
    depth x batch x features if axis == 0
  ```

  If `indices` is a RaggedTensor, the 'axis' argument must be positive and refer
  to a non-ragged axis. The output will be equivalent to applying 'one_hot' on
  the values of the RaggedTensor, and creating a new RaggedTensor from the
  result.

  If `dtype` is not provided, it will attempt to assume the data type of
  `on_value` or `off_value`, if one or both are passed in. If none of
  `on_value`, `off_value`, or `dtype` are provided, `dtype` will default to the
  value `tf.float32`.

  Note: If a non-numeric data type output is desired (`tf.string`, `tf.bool`,
  etc.), both `on_value` and `off_value` _must_ be provided to `one_hot`.

  For example:

  ```python
  indices = [0, 1, 2]
  depth = 3
  tf.one_hot(indices, depth)  # output: [3 x 3]
  # [[1., 0., 0.],
  #  [0., 1., 0.],
  #  [0., 0., 1.]]

  indices = [0, 2, -1, 1]
  depth = 3
  tf.one_hot(indices, depth,
             on_value=5.0, off_value=0.0,
             axis=-1)  # output: [4 x 3]
  # [[5.0, 0.0, 0.0],  # one_hot(0)
  #  [0.0, 0.0, 5.0],  # one_hot(2)
  #  [0.0, 0.0, 0.0],  # one_hot(-1)
  #  [0.0, 5.0, 0.0]]  # one_hot(1)

  indices = [[0, 2], [1, -1]]
  depth = 3
  tf.one_hot(indices, depth,
             on_value=1.0, off_value=0.0,
             axis=-1)  # output: [2 x 2 x 3]
  # [[[1.0, 0.0, 0.0],   # one_hot(0)
  #   [0.0, 0.0, 1.0]],  # one_hot(2)
  #  [[0.0, 1.0, 0.0],   # one_hot(1)
  #   [0.0, 0.0, 0.0]]]  # one_hot(-1)

  indices = tf.ragged.constant([[0, 1], [2]])
  depth = 3
  tf.one_hot(indices, depth)  # output: [2 x None x 3]
  # [[[1., 0., 0.],
  #   [0., 1., 0.]],
  #  [[0., 0., 1.]]]
  ```

  Args:
    indices: A `Tensor` of indices.
    depth: A scalar defining the depth of the one hot dimension.
    on_value: A scalar defining the value to fill in output when `indices[j]
      = i`. (default: 1)
    off_value: A scalar defining the value to fill in output when `indices[j]
      != i`. (default: 0)
    axis: The axis to fill (default: -1, a new inner-most axis).
    dtype: The data type of the output tensor.
    name: A name for the operation (optional).

  Returns:
    output: The one-hot tensor.

  Raises:
    TypeError: If dtype of either `on_value` or `off_value` don't match `dtype`
    TypeError: If dtype of `on_value` and `off_value` don't match one another
  """
  with ops.name_scope(
      name, "one_hot",
      [indices, depth, on_value, off_value, axis, dtype]) as name:
    on_exists = on_value is not None
    off_exists = off_value is not None

    if on_exists:
      on_value = ops.convert_to_tensor(on_value, dtype_hint=dtype)
    if off_exists:
      off_value = ops.convert_to_tensor(off_value, dtype_hint=dtype)

    on_dtype = on_value.dtype.base_dtype if on_exists else None
    off_dtype = off_value.dtype.base_dtype if off_exists else None

    if on_exists or off_exists:
      if dtype is not None:
        # Ensure provided on_value and/or off_value match dtype
        if on_exists and on_dtype != dtype:
          raise TypeError("dtype {0} of on_value does not match "
                          "dtype parameter {1}".format(on_dtype, dtype))
        if off_exists and off_dtype != dtype:
          raise TypeError("dtype {0} of off_value does not match "
                          "dtype parameter {1}".format(off_dtype, dtype))
      else:
        # dtype not provided: automatically assign it
        dtype = on_dtype if on_exists else off_dtype
    elif dtype is None:
      # None of on_value, off_value, or dtype provided. Default dtype to float32
      dtype = dtypes.float32

    if not on_exists:
      # on_value not provided: assign to value 1 of type dtype
      on_value = ops.convert_to_tensor(1, dtype, name="on_value")
      on_dtype = dtype
    if not off_exists:
      # off_value not provided: assign to value 0 of type dtype
      off_value = ops.convert_to_tensor(0, dtype, name="off_value")
      off_dtype = dtype

    if on_dtype != off_dtype:
      raise TypeError("dtype {0} of on_value does not match "
                      "dtype {1} of off_value".format(on_dtype, off_dtype))

    return gen_array_ops.one_hot(indices, depth, on_value, off_value, axis,
                                 name)


def _all_dimensions(x):
  """Returns a 1D-tensor listing all dimensions in x."""
  # Fast path: avoid creating Rank and Range ops if ndims is known.
  if isinstance(x, tensor_lib.Tensor) and x.get_shape().ndims is not None:
    return constant_op.constant(
        np.arange(x.get_shape().ndims), dtype=dtypes.int32)
  if (isinstance(x, sparse_tensor.SparseTensor) and
      x.dense_shape.get_shape().is_fully_defined()):
    r = x.dense_shape.get_shape().dims[0].value  # sparse.dense_shape is 1-D.
    return constant_op.constant(np.arange(r), dtype=dtypes.int32)

  # Otherwise, we rely on `range` and `rank` to do the right thing at runtime.
  return gen_math_ops._range(0, rank(x), 1)


@tf_export("sequence_mask")
@dispatch.add_dispatch_support
def sequence_mask(lengths, maxlen=None, dtype=dtypes.bool, name=None):
  """Returns a mask tensor representing the first N positions of each cell.

  If `lengths` has shape `[d_1, d_2, ..., d_n]` the resulting tensor `mask` has
  dtype `dtype` and shape `[d_1, d_2, ..., d_n, maxlen]`, with

  ```
  mask[i_1, i_2, ..., i_n, j] = (j < lengths[i_1, i_2, ..., i_n])
  ```

  Examples:

  ```python
  tf.sequence_mask([1, 3, 2], 5)  # [[True, False, False, False, False],
                                  #  [True, True, True, False, False],
                                  #  [True, True, False, False, False]]

  tf.sequence_mask([[1, 3],[2,0]])  # [[[True, False, False],
                                    #   [True, True, True]],
                                    #  [[True, True, False],
                                    #   [False, False, False]]]
  ```

  Args:
    lengths: integer tensor, all its values <= maxlen.
    maxlen: scalar integer tensor, size of last dimension of returned tensor.
      Default is the maximum value in `lengths`.
    dtype: output type of the resulting tensor.
    name: name of the op.

  Returns:
    A mask tensor of shape `lengths.shape + (maxlen,)`, cast to specified dtype.
  Raises:
    ValueError: if `maxlen` is not a scalar.
  """
  with ops.name_scope(name, "SequenceMask", [lengths, maxlen]):
    lengths = ops.convert_to_tensor(lengths)

    if maxlen is None:
      maxlen = gen_math_ops._max(lengths, _all_dimensions(lengths))
      maxlen = gen_math_ops.maximum(constant(0, maxlen.dtype), maxlen)
    else:
      maxlen = ops.convert_to_tensor(maxlen)
    if maxlen.get_shape().ndims is not None and maxlen.get_shape().ndims != 0:
      raise ValueError("Argument `maxlen` must be scalar for sequence_mask, "
                       f"received `maxlen` = {maxlen} "
                       f"with shape '{maxlen.get_shape()}' instead")

    # The basic idea is to compare a range row vector of size maxlen:
    # [0, 1, 2, 3, 4]
    # to length as a matrix with 1 column: [[1], [3], [2]].
    # Because of broadcasting on both arguments this comparison results
    # in a matrix of size (len(lengths), maxlen)
    row_vector = gen_math_ops._range(
        constant(0, maxlen.dtype), maxlen, constant(1, maxlen.dtype))
    # Since maxlen >= max(lengths), it is safe to use maxlen as a cast
    # authoritative type. Whenever maxlen fits into tf.int32, so do the lengths.
    matrix = gen_math_ops.cast(expand_dims(lengths, -1), maxlen.dtype)
    result = row_vector < matrix
    if dtype is None or result.dtype.is_compatible_with(dtype):
      return result
    else:
      return gen_math_ops.cast(result, dtype)


@tf_export(v1=["squeeze"])
@dispatch.add_dispatch_support
@deprecation.deprecated_args(None, "Use the `axis` argument instead",
                             "squeeze_dims")
def squeeze(input, axis=None, name=None, squeeze_dims=None):
  # pylint: disable=redefined-builtin
  """Removes dimensions of size 1 from the shape of a tensor.

  Given a tensor `input`, this operation returns a tensor of the same type with
  all dimensions of size 1 removed. If you don't want to remove all size 1
  dimensions, you can remove specific size 1 dimensions by specifying
  `axis`.

  For example:

  >>> # 't' is a tensor of shape [1, 2, 1, 3, 1, 1]
  >>> t = tf.ones([1, 2, 1, 3, 1, 1])
  >>> print(tf.shape(tf.squeeze(t)).numpy())
  [2 3]

  Or, to remove specific size 1 dimensions:

  >>> # 't' is a tensor of shape [1, 2, 1, 3, 1, 1]
  >>> t = tf.ones([1, 2, 1, 3, 1, 1])
  >>> print(tf.shape(tf.squeeze(t, [2, 4])).numpy())
  [1 2 3 1]

  Note: if `input` is a `tf.RaggedTensor`, then this operation takes `O(N)`
  time, where `N` is the number of elements in the squeezed dimensions.

  Args:
    input: A `Tensor`. The `input` to squeeze.
    axis: An optional list of `ints`. Defaults to `[]`. If specified, only
      squeezes the dimensions listed. The dimension index starts at 0. It is an
      error to squeeze a dimension that is not 1. Must be in the range
      `[-rank(input), rank(input))`. Must be specified if `input` is a
      `RaggedTensor`.
    name: A name for the operation (optional).
    squeeze_dims: Deprecated keyword argument that is now axis.

  Returns:
    A `Tensor`. Has the same type as `input`.
    Contains the same data as `input`, but has one or more dimensions of
    size 1 removed.

  Raises:
    ValueError: When both `squeeze_dims` and `axis` are specified.
  """
  axis = deprecation.deprecated_argument_lookup("axis", axis, "squeeze_dims",
                                                squeeze_dims)
  if np.isscalar(axis):
    axis = [axis]
  return gen_array_ops.squeeze(input, axis, name)


@tf_export("squeeze", v1=[])
@dispatch.add_dispatch_support
def squeeze_v2(input, axis=None, name=None):
  """Removes dimensions of size 1 from the shape of a tensor.

  Given a tensor `input`, this operation returns a tensor of the same type with
  all dimensions of size 1 removed. If you don't want to remove all size 1
  dimensions, you can remove specific size 1 dimensions by specifying
  `axis`.

  For example:

  ```python
  # 't' is a tensor of shape [1, 2, 1, 3, 1, 1]
  tf.shape(tf.squeeze(t))  # [2, 3]
  ```

  Or, to remove specific size 1 dimensions:

  ```python
  # 't' is a tensor of shape [1, 2, 1, 3, 1, 1]
  tf.shape(tf.squeeze(t, [2, 4]))  # [1, 2, 3, 1]
  ```

  Unlike the older op `tf.compat.v1.squeeze`, this op does not accept a
  deprecated `squeeze_dims` argument.

  Note: if `input` is a `tf.RaggedTensor`, then this operation takes `O(N)`
  time, where `N` is the number of elements in the squeezed dimensions.

  Note: If squeeze is performed on dimensions of unknown sizes, then the
  returned Tensor will be of unknown shape. A common situation is when the
  first (batch) dimension is of size `None`, `tf.squeeze` returns
  `<unknown>` shape which may be a surprise. Specify the `axis=` argument
  to get the expected result, as illustrated in the following example:

  ```python
  @tf.function
  def func(x):
    print('x.shape:', x.shape)
    known_axes = [i for i, size in enumerate(x.shape) if size == 1]
    y = tf.squeeze(x, axis=known_axes)
    print('shape of tf.squeeze(x, axis=known_axes):', y.shape)
    y = tf.squeeze(x)
    print('shape of tf.squeeze(x):', y.shape)
    return 0

  _ = func.get_concrete_function(tf.TensorSpec([None, 1, 2], dtype=tf.int32))
  # Output is.
  # x.shape: (None, 1, 2)
  # shape of tf.squeeze(x, axis=known_axes): (None, 2)
  # shape of tf.squeeze(x): <unknown>
  ```

  Args:
    input: A `Tensor`. The `input` to squeeze.
    axis: An optional list of `ints`. Defaults to `[]`. If specified, only
      squeezes the dimensions listed. The dimension index starts at 0. It is an
      error to squeeze a dimension that is not 1. Must be in the range
      `[-rank(input), rank(input))`. Must be specified if `input` is a
      `RaggedTensor`.
    name: A name for the operation (optional).

  Returns:
    A `Tensor`. Has the same type as `input`.
    Contains the same data as `input`, but has one or more dimensions of
    size 1 removed.

  Raises:
    ValueError: The input cannot be converted to a tensor, or the specified
      axis cannot be squeezed.
  """
  # pylint: disable=redefined-builtin
  return squeeze(input, axis, name)


@tf_export(v1=["where"])
@dispatch.add_dispatch_support
def where(condition, x=None, y=None, name=None):
  """Return the elements, either from `x` or `y`, depending on the `condition`.

  If both `x` and `y` are None, then this operation returns the coordinates of
  true elements of `condition`.  The coordinates are returned in a 2-D tensor
  where the first dimension (rows) represents the number of true elements, and
  the second dimension (columns) represents the coordinates of the true
  elements. Keep in mind, the shape of the output tensor can vary depending on
  how many true values there are in input. Indices are output in row-major
  order.

  If both non-None, `x` and `y` must have the same shape.
  The `condition` tensor must be a scalar if `x` and `y` are scalar.
  If `x` and `y` are tensors of higher rank, then `condition` must be either a
  vector with size matching the first dimension of `x`, or must have the same
  shape as `x`.

  The `condition` tensor acts as a mask that chooses, based on the value at each
  element, whether the corresponding element / row in the output should be taken
  from `x` (if true) or `y` (if false).

  If `condition` is a vector and `x` and `y` are higher rank matrices, then it
  chooses which row (outer dimension) to copy from `x` and `y`. If `condition`
  has the same shape as `x` and `y`, then it chooses which element to copy from
  `x` and `y`.

  Args:
    condition: A `Tensor` of type `bool`
    x: A Tensor which may have the same shape as `condition`. If `condition` is
      rank 1, `x` may have higher rank, but its first dimension must match the
      size of `condition`.
    y: A `tensor` with the same shape and type as `x`.
    name: A name of the operation (optional)

  Returns:
    A `Tensor` with the same type and shape as `x`, `y` if they are non-None.
    Otherwise, a `Tensor` with shape `(num_true, rank(condition))`.

  Raises:
    ValueError: When exactly one of `x` or `y` is non-None.

  @compatibility(TF2)

  This API is compatible with eager execution and `tf.function`. However, this
  is still a legacy API endpoint originally designed for TF1. To migrate to
  fully-native TF2, please replace its usage with `tf.where` instead, which is
  directly backwards compatible with `tf.compat.v1.where`.

  However,`tf.compat.v1.where` is more restrictive than `tf.where`, requiring
  `x` and `y` to have the same shape, and returning a `Tensor` with the same
  type and shape as `x`, `y` (if they are both non-None).

  `tf.where` will accept `x`, `y` that are not the same shape as long as they
  are broadcastable with one another and with `condition`, and will return a
  `Tensor` with shape broadcast from `condition`, `x`, and `y`.

  For example, the following works with `tf.where` but not `tf.compat.v1.where`:

  >>> tf.where([True, False, False, True], [1,2,3,4], [100])
  <tf.Tensor: shape=(4,), dtype=int32, numpy=array([  1, 100, 100,   4],
  dtype=int32)>

  >>> tf.where(True, [1,2,3,4], 100)
  <tf.Tensor: shape=(4,), dtype=int32, numpy=array([1, 2, 3, 4],
  dtype=int32)>

  @end_compatibility
  """
  if x is None and y is None:
    with ops.name_scope(name, "Where", [condition]) as name:
      condition = ops.convert_to_tensor(
          condition, preferred_dtype=dtypes.bool, name="condition")
      return gen_array_ops.where(condition=condition, name=name)
  elif x is not None and y is not None:
    return gen_math_ops.select(condition=condition, x=x, y=y, name=name)
  else:
    raise ValueError("x and y must both be non-None or both be None.")


@tf_export("where", v1=["where_v2"])
@dispatch.add_dispatch_support
def where_v2(condition, x=None, y=None, name=None):
  """Returns the indices of non-zero elements, or multiplexes `x` and `y`.

  This operation has two modes:

  1. **Return the indices of non-zero elements** - When only
     `condition` is provided the result is an `int64` tensor where each row is
     the index of a non-zero element of `condition`. The result's shape
     is `[tf.math.count_nonzero(condition), tf.rank(condition)]`.
  2. **Multiplex `x` and `y`** - When both `x` and `y` are provided the
     result has the shape of `x`, `y`, and `condition` broadcast together. The
     result is taken from `x` where `condition` is non-zero
     or `y` where `condition` is zero.

  #### 1. Return the indices of non-zero elements

  Note: In this mode `condition` can have a dtype of `bool` or any numeric
  dtype.

  If `x` and `y` are not provided (both are None):

  `tf.where` will return the indices of `condition` that are non-zero,
  in the form of a 2-D tensor with shape `[n, d]`, where `n` is the number of
  non-zero elements in `condition` (`tf.count_nonzero(condition)`), and `d` is
  the number of axes of `condition` (`tf.rank(condition)`).

  Indices are output in row-major order. The `condition` can have a `dtype` of
  `tf.bool`, or any numeric `dtype`.

  Here `condition` is a 1-axis `bool` tensor with 2 `True` values. The result
  has a shape of `[2,1]`

  >>> tf.where([True, False, False, True]).numpy()
  array([[0],
         [3]])

  Here `condition` is a 2-axis integer tensor, with 3 non-zero values. The
  result has a shape of `[3, 2]`.

  >>> tf.where([[1, 0, 0], [1, 0, 1]]).numpy()
  array([[0, 0],
         [1, 0],
         [1, 2]])

  Here `condition` is a 3-axis float tensor, with 5 non-zero values. The output
  shape is `[5, 3]`.

  >>> float_tensor = [[[0.1, 0], [0, 2.2], [3.5, 1e6]],
  ...                 [[0,   0], [0,   0], [99,    0]]]
  >>> tf.where(float_tensor).numpy()
  array([[0, 0, 0],
         [0, 1, 1],
         [0, 2, 0],
         [0, 2, 1],
         [1, 2, 0]])

  These indices are the same that `tf.sparse.SparseTensor` would use to
  represent the condition tensor:

  >>> sparse = tf.sparse.from_dense(float_tensor)
  >>> sparse.indices.numpy()
  array([[0, 0, 0],
         [0, 1, 1],
         [0, 2, 0],
         [0, 2, 1],
         [1, 2, 0]])

  A complex number is considered non-zero if either the real or imaginary
  component is non-zero:

  >>> tf.where([complex(0.), complex(1.), 0+1j, 1+1j]).numpy()
  array([[1],
         [2],
         [3]])

  #### 2. Multiplex `x` and `y`

  Note: In this mode `condition` must have a dtype of `bool`.

  If `x` and `y` are also provided (both have non-None values) the `condition`
  tensor acts as a mask that chooses whether the corresponding
  element / row in the output should be taken from `x` (if the element in
  `condition` is `True`) or `y` (if it is `False`).

  The shape of the result is formed by
  [broadcasting](https://docs.scipy.org/doc/numpy/reference/ufuncs.html)
  together the shapes of `condition`, `x`, and `y`.

  When all three inputs have the same size, each is handled element-wise.

  >>> tf.where([True, False, False, True],
  ...          [1, 2, 3, 4],
  ...          [100, 200, 300, 400]).numpy()
  array([  1, 200, 300,   4], dtype=int32)

  There are two main rules for broadcasting:

  1. If a tensor has fewer axes than the others, length-1 axes are added to the
     left of the shape.
  2. Axes with length-1 are streched to match the coresponding axes of the other
     tensors.

  A length-1 vector is streched to match the other vectors:

  >>> tf.where([True, False, False, True], [1, 2, 3, 4], [100]).numpy()
  array([  1, 100, 100,   4], dtype=int32)

  A scalar is expanded to match the other arguments:

  >>> tf.where([[True, False], [False, True]], [[1, 2], [3, 4]], 100).numpy()
  array([[  1, 100], [100,   4]], dtype=int32)
  >>> tf.where([[True, False], [False, True]], 1, 100).numpy()
  array([[  1, 100], [100,   1]], dtype=int32)

  A scalar `condition` returns the complete `x` or `y` tensor, with
  broadcasting applied.

  >>> tf.where(True, [1, 2, 3, 4], 100).numpy()
  array([1, 2, 3, 4], dtype=int32)
  >>> tf.where(False, [1, 2, 3, 4], 100).numpy()
  array([100, 100, 100, 100], dtype=int32)

  For a non-trivial example of broadcasting, here `condition` has a shape of
  `[3]`, `x` has a shape of `[3,3]`, and `y` has a shape of `[3,1]`.
  Broadcasting first expands the shape of `condition` to `[1,3]`. The final
  broadcast shape is `[3,3]`. `condition` will select columns from `x` and `y`.
  Since `y` only has one column, all columns from `y` will be identical.

  >>> tf.where([True, False, True],
  ...          x=[[1, 2, 3],
  ...             [4, 5, 6],
  ...             [7, 8, 9]],
  ...          y=[[100],
  ...             [200],
  ...             [300]]
  ... ).numpy()
  array([[ 1, 100, 3],
         [ 4, 200, 6],
         [ 7, 300, 9]], dtype=int32)

  Note that if the gradient of either branch of the `tf.where` generates
  a `NaN`, then the gradient of the entire `tf.where` will be `NaN`. This is
  because the gradient calculation for `tf.where` combines the two branches, for
  performance reasons.

  A workaround is to use an inner `tf.where` to ensure the function has
  no asymptote, and to avoid computing a value whose gradient is `NaN` by
  replacing dangerous inputs with safe inputs.

  Instead of this,

  >>> x = tf.constant(0., dtype=tf.float32)
  >>> with tf.GradientTape() as tape:
  ...   tape.watch(x)
  ...   y = tf.where(x < 1., 0., 1. / x)
  >>> print(tape.gradient(y, x))
  tf.Tensor(nan, shape=(), dtype=float32)

  Although, the `1. / x` values are never used, its gradient is a `NaN` when
  `x = 0`. Instead, we should guard that with another `tf.where`

  >>> x = tf.constant(0., dtype=tf.float32)
  >>> with tf.GradientTape() as tape:
  ...   tape.watch(x)
  ...   safe_x = tf.where(tf.equal(x, 0.), 1., x)
  ...   y = tf.where(x < 1., 0., 1. / safe_x)
  >>> print(tape.gradient(y, x))
  tf.Tensor(0.0, shape=(), dtype=float32)

  See also:

  * `tf.sparse` - The indices returned by the first form of `tf.where` can be
     useful in `tf.sparse.SparseTensor` objects.
  * `tf.gather_nd`, `tf.scatter_nd`, and related ops - Given the
    list of indices returned from `tf.where` the `scatter` and `gather` family
    of ops can be used fetch values or insert values at those indices.
  * `tf.strings.length` - `tf.string` is not an allowed dtype for the
    `condition`. Use the string length instead.

  Args:
    condition: A `tf.Tensor` of dtype bool, or any numeric dtype. `condition`
      must have dtype `bool` when `x` and `y` are provided.
    x: If provided, a Tensor which is of the same type as `y`, and has a shape
      broadcastable with `condition` and `y`.
    y: If provided, a Tensor which is of the same type as `x`, and has a shape
      broadcastable with `condition` and `x`.
    name: A name of the operation (optional).

  Returns:
    If `x` and `y` are provided:
      A `Tensor` with the same type as `x` and `y`, and shape that
      is broadcast from `condition`, `x`, and `y`.
    Otherwise, a `Tensor` with shape `[tf.math.count_nonzero(condition),
    tf.rank(condition)]`.

  Raises:
    ValueError: When exactly one of `x` or `y` is non-None, or the shapes
      are not all broadcastable.
  """
  if x is None and y is None:
    with ops.name_scope(name, "Where", [condition]) as name:
      condition = ops.convert_to_tensor(
          condition, preferred_dtype=dtypes.bool, name="condition")
      return gen_array_ops.where(condition=condition, name=name)
  elif x is not None and y is not None:
    return gen_math_ops.select_v2(condition=condition, t=x, e=y, name=name)
  else:
    raise ValueError("x and y must both be non-None or both be None.")


# pylint: disable=redefined-builtin
@tf_export(v1=["reverse_sequence"])
@deprecation.deprecated_args(None,
                             "seq_dim is deprecated, use seq_axis instead",
                             "seq_dim")
@deprecation.deprecated_args(None,
                             "batch_dim is deprecated, use batch_axis instead",
                             "batch_dim")
def reverse_sequence(input,
                     seq_lengths,
                     seq_axis=None,
                     batch_axis=None,
                     name=None,
                     seq_dim=None,
                     batch_dim=None):
  """Reverses variable length slices.

  This op first slices `input` along the dimension `batch_axis`, and for
  each slice `i`, reverses the first `seq_lengths[i]` elements along the
  dimension `seq_axis`.

  The elements of `seq_lengths` must obey `seq_lengths[i] <=
  input.dims[seq_axis]`, and `seq_lengths` must be a vector of length
  `input.dims[batch_axis]`.

  The output slice `i` along dimension `batch_axis` is then given by
  input slice `i`, with the first `seq_lengths[i]` slices along
  dimension `seq_axis` reversed.

  Example usage:

  >>> seq_lengths = [7, 2, 3, 5]
  >>> input = [[1, 2, 3, 4, 5, 0, 0, 0], [1, 2, 0, 0, 0, 0, 0, 0],
  ...          [1, 2, 3, 4, 0, 0, 0, 0], [1, 2, 3, 4, 5, 6, 7, 8]]
  >>> output = tf.reverse_sequence(input, seq_lengths, seq_axis=1, batch_axis=0)
  >>> output
  <tf.Tensor: shape=(4, 8), dtype=int32, numpy=
  array([[0, 0, 5, 4, 3, 2, 1, 0],
         [2, 1, 0, 0, 0, 0, 0, 0],
         [3, 2, 1, 4, 0, 0, 0, 0],
         [5, 4, 3, 2, 1, 6, 7, 8]], dtype=int32)>

  Args:
    input: A `Tensor`. The input to reverse.
    seq_lengths: A `Tensor`. Must be one of the following types: `int32`,
      `int64`. 1-D with length `input.dims(batch_axis)` and `max(seq_lengths) <=
      input.dims(seq_axis)`
    seq_axis: An `int`. The dimension which is partially reversed.
    batch_axis: An optional `int`. Defaults to `0`. The dimension along which
      reversal is performed.
    name: A name for the operation (optional).

  Returns:
    A Tensor. Has the same type as input.
  """
  seq_axis = deprecation.deprecated_argument_lookup("seq_axis", seq_axis,
                                                    "seq_dim", seq_dim)
  batch_axis = deprecation.deprecated_argument_lookup("batch_axis", batch_axis,
                                                      "batch_dim", batch_dim)
  return gen_array_ops.reverse_sequence(
      input=input,
      seq_lengths=seq_lengths,
      seq_dim=seq_axis,
      batch_dim=batch_axis,
      name=name)


@tf_export("reverse_sequence", v1=[])
@dispatch.add_dispatch_support
def reverse_sequence_v2(input,
                        seq_lengths,
                        seq_axis=None,
                        batch_axis=None,
                        name=None):
  """Reverses variable length slices.

  This op first slices `input` along the dimension `batch_axis`, and for
  each slice `i`, reverses the first `seq_lengths[i]` elements along the
  dimension `seq_axis`.

  The elements of `seq_lengths` must obey `seq_lengths[i] <=
  input.dims[seq_axis]`, and `seq_lengths` must be a vector of length
  `input.dims[batch_axis]`.

  The output slice `i` along dimension `batch_axis` is then given by
  input slice `i`, with the first `seq_lengths[i]` slices along
  dimension `seq_axis` reversed.

  Example usage:

  >>> seq_lengths = [7, 2, 3, 5]
  >>> input = [[1, 2, 3, 4, 5, 0, 0, 0], [1, 2, 0, 0, 0, 0, 0, 0],
  ...          [1, 2, 3, 4, 0, 0, 0, 0], [1, 2, 3, 4, 5, 6, 7, 8]]
  >>> output = tf.reverse_sequence(input, seq_lengths, seq_axis=1, batch_axis=0)
  >>> output
  <tf.Tensor: shape=(4, 8), dtype=int32, numpy=
  array([[0, 0, 5, 4, 3, 2, 1, 0],
         [2, 1, 0, 0, 0, 0, 0, 0],
         [3, 2, 1, 4, 0, 0, 0, 0],
         [5, 4, 3, 2, 1, 6, 7, 8]], dtype=int32)>

  Args:
    input: A `Tensor`. The input to reverse.
    seq_lengths: A `Tensor`. Must be one of the following types: `int32`,
      `int64`. 1-D with length `input.dims(batch_axis)` and `max(seq_lengths) <=
      input.dims(seq_axis)`
    seq_axis: An `int`. The dimension which is partially reversed.
    batch_axis: An optional `int`. Defaults to `0`. The dimension along which
      reversal is performed.
    name: A name for the operation (optional).

  Returns:
    A Tensor. Has the same type as input.
  """
  return gen_array_ops.reverse_sequence(
      input=input,
      seq_lengths=seq_lengths,
      seq_dim=seq_axis,
      batch_dim=batch_axis,
      name=name)

# pylint: enable=redefined-builtin


@tf_export(v1=["gather"])
@dispatch.add_dispatch_support
@deprecation.deprecated_args(None,
                             ("The `validate_indices` argument has no effect. "
                              "Indices are always validated on CPU and never "
                              "validated on GPU."),
                             ("validate_indices", None))
def gather(params,
           indices,
           validate_indices=None,
           name=None,
           axis=None,
           batch_dims=0):  # pylint: disable=g-doc-args
  r"""Gather slices from params axis `axis` according to indices.

  Gather slices from `params` axis `axis` according to `indices`.  `indices`
  must be an integer tensor of any dimension (often 1-D).

  `Tensor.__getitem__` works for scalars, `tf.newaxis`, and
  [python slices](https://numpy.org/doc/stable/reference/arrays.indexing.html#basic-slicing-and-indexing)

  `tf.gather` extends indexing to handle tensors of indices.

  In the simplest case it's identical to scalar indexing:

  >>> params = tf.constant(['p0', 'p1', 'p2', 'p3', 'p4', 'p5'])
  >>> params[3].numpy()
  b'p3'
  >>> tf.gather(params, 3).numpy()
  b'p3'

  The most common case is to pass a single axis tensor of indices (this
  can't be expressed as a python slice because the indices are not sequential):

  >>> indices = [2, 0, 2, 5]
  >>> tf.gather(params, indices).numpy()
  array([b'p2', b'p0', b'p2', b'p5'], dtype=object)

  <div style="width:70%; margin:auto; margin-bottom:10px; margin-top:20px;">
  <img style="width:100%" src="https://www.tensorflow.org/images/Gather.png"
  alt>
  </div>

  The indices can have any shape. When the `params` has 1 axis, the
  output shape is equal to the input shape:

  >>> tf.gather(params, [[2, 0], [2, 5]]).numpy()
  array([[b'p2', b'p0'],
         [b'p2', b'p5']], dtype=object)

  The `params` may also have any shape. `gather` can select slices
  across any axis depending on the `axis` argument (which defaults to 0).
  Below it is used to gather first rows, then columns from a matrix:

  >>> params = tf.constant([[0, 1.0, 2.0],
  ...                       [10.0, 11.0, 12.0],
  ...                       [20.0, 21.0, 22.0],
  ...                       [30.0, 31.0, 32.0]])
  >>> tf.gather(params, indices=[3,1]).numpy()
  array([[30., 31., 32.],
         [10., 11., 12.]], dtype=float32)
  >>> tf.gather(params, indices=[2,1], axis=1).numpy()
  array([[ 2.,  1.],
         [12., 11.],
         [22., 21.],
         [32., 31.]], dtype=float32)

  More generally: The output shape has the same shape as the input, with the
  indexed-axis replaced by the shape of the indices.

  >>> def result_shape(p_shape, i_shape, axis=0):
  ...   return p_shape[:axis] + i_shape + p_shape[axis+1:]
  >>>
  >>> result_shape([1, 2, 3], [], axis=1)
  [1, 3]
  >>> result_shape([1, 2, 3], [7], axis=1)
  [1, 7, 3]
  >>> result_shape([1, 2, 3], [7, 5], axis=1)
  [1, 7, 5, 3]

  Here are some examples:

  >>> params.shape.as_list()
  [4, 3]
  >>> indices = tf.constant([[0, 2]])
  >>> tf.gather(params, indices=indices, axis=0).shape.as_list()
  [1, 2, 3]
  >>> tf.gather(params, indices=indices, axis=1).shape.as_list()
  [4, 1, 2]

  >>> params = tf.random.normal(shape=(5, 6, 7, 8))
  >>> indices = tf.random.uniform(shape=(10, 11), maxval=7, dtype=tf.int32)
  >>> result = tf.gather(params, indices, axis=2)
  >>> result.shape.as_list()
  [5, 6, 10, 11, 8]

  This is because each index takes a slice from `params`, and
  places it at the corresponding location in the output. For the above example

  >>> # For any location in indices
  >>> a, b = 0, 1
  >>> tf.reduce_all(
  ...     # the corresponding slice of the result
  ...     result[:, :, a, b, :] ==
  ...     # is equal to the slice of `params` along `axis` at the index.
  ...     params[:, :, indices[a, b], :]
  ... ).numpy().item()
  True

  ### Batching:

  The `batch_dims` argument lets you gather different items from each element
  of a batch.

  Using `batch_dims=1` is equivalent to having an outer loop over the first
  axis of `params` and `indices`:

  >>> params = tf.constant([
  ...     [0, 0, 1, 0, 2],
  ...     [3, 0, 0, 0, 4],
  ...     [0, 5, 0, 6, 0]])
  >>> indices = tf.constant([
  ...     [2, 4],
  ...     [0, 4],
  ...     [1, 3]])

  >>> tf.gather(params, indices, axis=1, batch_dims=1).numpy()
  array([[1, 2],
         [3, 4],
         [5, 6]], dtype=int32)

  This is equivalent to:

  >>> def manually_batched_gather(params, indices, axis):
  ...   batch_dims=1
  ...   result = []
  ...   for p,i in zip(params, indices):
  ...     r = tf.gather(p, i, axis=axis-batch_dims)
  ...     result.append(r)
  ...   return tf.stack(result)
  >>> manually_batched_gather(params, indices, axis=1).numpy()
  array([[1, 2],
         [3, 4],
         [5, 6]], dtype=int32)

  Higher values of `batch_dims` are equivalent to multiple nested loops over
  the outer axes of `params` and `indices`. So the overall shape function is

  >>> def batched_result_shape(p_shape, i_shape, axis=0, batch_dims=0):
  ...   return p_shape[:axis] + i_shape[batch_dims:] + p_shape[axis+1:]
  >>>
  >>> batched_result_shape(
  ...     p_shape=params.shape.as_list(),
  ...     i_shape=indices.shape.as_list(),
  ...     axis=1,
  ...     batch_dims=1)
  [3, 2]

  >>> tf.gather(params, indices, axis=1, batch_dims=1).shape.as_list()
  [3, 2]

  This comes up naturally if you need to use the indices of an operation like
  `tf.argsort`, or `tf.math.top_k` where the last dimension of the indices
  indexes into the last dimension of input, at the corresponding location.
  In this case you can use `tf.gather(values, indices, batch_dims=-1)`.

  See also:

  * `tf.Tensor.__getitem__`: The direct tensor index operation (`t[]`), handles
    scalars and python-slices `tensor[..., 7, 1:-1]`
  * `tf.scatter`: A collection of operations similar to `__setitem__`
    (`t[i] = x`)
  * `tf.gather_nd`: An operation similar to `tf.gather` but gathers across
    multiple axis at once (it can gather elements of a matrix instead of rows
    or columns)
  * `tf.boolean_mask`, `tf.where`: Binary indexing.
  * `tf.slice` and `tf.strided_slice`: For lower level access to the
    implementation of `__getitem__`'s python-slice handling (`t[1:-1:2]`)

  Args:
    params: The `Tensor` from which to gather values. Must be at least rank
      `axis + 1`.
    indices: The index `Tensor`.  Must be one of the following types: `int32`,
      `int64`. The values must be in range `[0, params.shape[axis])`.
    validate_indices: Deprecated, does nothing. Indices are always validated on
      CPU, never validated on GPU.

      Caution: On CPU, if an out of bound index is found, an error is raised.
      On GPU, if an out of bound index is found, a 0 is stored in the
      corresponding output value.
    axis: A `Tensor`. Must be one of the following types: `int32`, `int64`. The
      `axis` in `params` to gather `indices` from. Must be greater than or equal
      to `batch_dims`.  Defaults to the first non-batch dimension. Supports
      negative indexes.
    batch_dims: An `integer`.  The number of batch dimensions.  Must be less
      than or equal to `rank(indices)`.
    name: A name for the operation (optional).

  Returns:
    A `Tensor`. Has the same type as `params`.
  """
  del validate_indices

  if axis is None:
    axis = batch_dims
  if tensor_util.constant_value(axis) != 0:
    return gen_array_ops.gather_v2(
        params, indices, axis, batch_dims=batch_dims, name=name)
  try:
    # TODO(apassos) find a less bad way of detecting resource variables
    # without introducing a circular dependency.
    return params.sparse_read(indices, name=name)
  except AttributeError:
    return gen_array_ops.gather_v2(params, indices, axis, name=name)


@tf_export("gather", v1=[])
@dispatch.add_dispatch_support
def gather_v2(params,
              indices,
              validate_indices=None,
              axis=None,
              batch_dims=0,
              name=None):
  return gather(
      params,
      indices,
      validate_indices=validate_indices,
      name=name,
      axis=axis,
      batch_dims=batch_dims)


gather_v2.__doc__ = gather.__doc__


@tf_export(v1=["batch_gather"])
@dispatch.add_dispatch_support
@deprecation.deprecated(
    "2017-10-25", "`tf.batch_gather` is deprecated, please use `tf.gather` "
    "with `batch_dims=tf.rank(indices) - 1` instead.")  # pylint: disable=missing-docstring
def batch_gather(params, indices, name=None):
  """Gather slices from params according to indices with leading batch dims."""
  with ops.name_scope(name, "BatchGather", [params, indices]):
    indices = ops.convert_to_tensor(indices, name="indices")
    params = ops.convert_to_tensor(params, name="params")
    if indices.shape.ndims is None:
      raise ValueError(
          "batch_gather does not allow indices with unknown shape.")
    return _batch_gather(params, indices, batch_dims=indices.shape.ndims - 1)


def _batch_gather(params, indices, batch_dims, axis=None):
  r"""Gather slices from params according to indices with leading batch dims.

  This operation assumes that the leading `batch_dims` dimensions of `indices`
  and `params` are batch dimensions; and performs a `tf.gather` operation within
  each batch. (If `batch_dims` is not specified, then it defaults to
  `rank(indices)-1`.)  In the case in which `batch_dims==0`, this operation
  is equivalent to `tf.gather`.

  Args:
    params: A Tensor. The tensor from which to gather values.
    indices: A Tensor. Must be one of the following types: int32, int64. Index
      tensor. Must be in range `[0, params.shape[batch_dims]]`.
    batch_dims: An integer or none.  The number of batch dimensions.  Must be
      less than `rank(indices)`.  Defaults to `rank(indices) - 1` if None.
    axis: A `Tensor`. Must be one of the following types: `int32`, `int64`. The
      `axis` in `params` to gather `indices` from. Must be greater than or equal
      to `batch_dims`.  Defaults to the first non-batch dimension. Supports
      negative indexes.

  Returns:
    A Tensor. Has the same type as `params`.

  Raises:
    ValueError: if `indices` has an unknown shape.
  """
  if batch_dims is not None and not isinstance(batch_dims, int):
    raise TypeError("Argument `batch_dims` must be an int. "
                    f"Received `batch_dims` = {batch_dims} instead")
  indices = ops.convert_to_tensor(indices, name="indices")
  params = ops.convert_to_tensor(params, name="params")

  indices_ndims = indices.shape.ndims
  if indices_ndims is None:
    raise ValueError("tf.gather does not allow indices with unknown "
                     "rank when batch_dims is specified.")
  if batch_dims is None:
    batch_dims = indices_ndims - 1
  if batch_dims < 0:
    batch_dims += indices_ndims
  if batch_dims < 0 or batch_dims >= indices_ndims:
    raise ValueError(f"Argument `batch_dims` = {batch_dims} must be less than "
                     f"rank(`indices`) = {indices_ndims}")
  if params.shape.ndims is not None and batch_dims >= params.shape.ndims:
    raise ValueError(f"Argument `batch_dims` = {batch_dims} must be less than "
                     f"rank(`params`) = {params.shape.ndims}")

  # Handle axis by transposing the axis dimension to be the first non-batch
  # dimension, recursively calling batch_gather with axis=0, and then
  # transposing the result to put the pre-axis dimensions before the indices
  # dimensions.
  if axis is not None and axis != batch_dims:
    # Adjust axis to be positive.
    if not isinstance(axis, int):
      axis = tf.where(axis < 0, axis + array_ops.rank(params), axis)
    elif axis < 0 and params.shape.ndims is None:
      axis = axis + array_ops.rank(params)
    else:
      if (axis < -params.shape.ndims) or (axis >= params.shape.ndims):
        raise ValueError(f"Argument `axis` = {axis} out of range "
                         f"[{-params.shape.ndims}, {params.shape.ndims})")
      if axis < 0:
        axis += params.shape.ndims
      if axis < batch_dims:
        raise ValueError(f"Argument `batch_dims` = {batch_dims} must be less "
                         f"than or equal to argument `axis` = {axis}")

    # Move params[axis] up to params[batch_dims].
    perm = [
        list(range(batch_dims)), [axis],
        gen_math_ops._range(batch_dims, axis, 1),
        gen_math_ops._range(axis + 1, rank(params), 1)
    ]
    params = transpose(params, concat(perm, axis=0))

    result = _batch_gather(params, indices, batch_dims=batch_dims)

    # Move the result dimensions corresponding to params[batch_dims:axis]
    # to just before the dimensions corresponding to indices[batch_dims:].
    params_start = indices_ndims + axis - batch_dims
    perm = [
        list(range(batch_dims)),
        gen_math_ops._range(indices_ndims, params_start, 1),
        list(range(batch_dims, indices_ndims)),
        gen_math_ops._range(params_start, rank(result), 1)
    ]
    return transpose(result, perm=concat(perm, axis=0))

  indices_shape = shape(indices)
  params_shape = shape(params)
  batch_indices = indices
  indices_dtype = indices.dtype.base_dtype
  accum_dim_value = ones((), dtype=indices_dtype)
  # Use correct type for offset index computation
  casted_params_shape = gen_math_ops.cast(params_shape, indices_dtype)
  for dim in range(batch_dims, 0, -1):
    dim_value = casted_params_shape[dim - 1]
    accum_dim_value *= casted_params_shape[dim]
    start = zeros((), dtype=indices_dtype)
    step = ones((), dtype=indices_dtype)
    dim_indices = gen_math_ops._range(start, dim_value, step)
    dim_indices *= accum_dim_value
    dim_shape = array_ops_stack.stack(
        [1] * (dim - 1) + [dim_value] + [1] * (indices_ndims - dim), axis=0)
    batch_indices += reshape(dim_indices, dim_shape)

  flat_indices = reshape(batch_indices, [-1])
  outer_shape = params_shape[batch_dims + 1:]
  flat_inner_shape = gen_math_ops.prod(params_shape[:batch_dims + 1], [0],
                                       False)

  flat_params = reshape(params, concat([[flat_inner_shape], outer_shape],
                                       axis=0))
  flat_result = gather(flat_params, flat_indices)
  result = reshape(flat_result, concat([indices_shape, outer_shape], axis=0))
  final_shape = indices.get_shape()[:batch_dims].merge_with(
      params.get_shape()[:batch_dims])
  final_shape = final_shape.concatenate(indices.get_shape().dims[batch_dims:])
  final_shape = final_shape.concatenate(params.get_shape()[batch_dims + 1:])
  result.set_shape(final_shape)
  return result


@tf_export(v1=["gather_nd", "manip.gather_nd"])
@dispatch.add_dispatch_support
@deprecated_endpoints("manip.gather_nd")
def gather_nd(params, indices, name=None, batch_dims=0, bad_indices_policy=""):
  r"""Gather slices from `params` into a Tensor with shape specified by `indices`.

  `indices` is a `Tensor` of indices into `params`. The index vectors are
  arranged along the last axis of `indices`.

  This is similar to `tf.gather`, in which `indices` defines slices into the
  first dimension of `params`. In `tf.gather_nd`, `indices` defines slices into
  the first `N` dimensions of `params`, where `N = indices.shape[-1]`.

  ## Gathering scalars

  In the simplest case the vectors in `indices` index the full rank of `params`:

  >>> tf.gather_nd(
  ...     indices=[[0, 0],
  ...              [1, 1]],
  ...     params = [['a', 'b'],
  ...               ['c', 'd']]).numpy()
  array([b'a', b'd'], dtype=object)

  In this case the result has 1-axis fewer than `indices`, and each index vector
  is replaced by the scalar indexed from `params`.

  In this case the shape relationship is:

  ```
  index_depth = indices.shape[-1]
  assert index_depth == params.shape.rank
  result_shape = indices.shape[:-1]
  ```

  If `indices` has a rank of `K`, it is helpful to think `indices` as a
  (K-1)-dimensional tensor of indices into `params`.

  ## Gathering slices

  If the index vectors do not index the full rank of `params` then each location
  in the result contains a slice of params. This example collects rows from a
  matrix:

  >>> tf.gather_nd(
  ...     indices = [[1],
  ...                [0]],
  ...     params = [['a', 'b', 'c'],
  ...               ['d', 'e', 'f']]).numpy()
  array([[b'd', b'e', b'f'],
         [b'a', b'b', b'c']], dtype=object)

  Here `indices` contains `[2]` index vectors, each with a length of `1`.
  The index vectors each refer to rows of the `params` matrix. Each
  row has a shape of `[3]` so the output shape is `[2, 3]`.

  In this case, the relationship between the shapes is:

  ```
  index_depth = indices.shape[-1]
  outer_shape = indices.shape[:-1]
  assert index_depth <= params.shape.rank
  inner_shape = params.shape[index_depth:]
  output_shape = outer_shape + inner_shape
  ```

  It is helpful to think of the results in this case as tensors-of-tensors.
  The shape of the outer tensor is set by the leading dimensions of `indices`.
  While the shape of the inner tensors is the shape of a single slice.

  ## Batches

  Additionally, both `params` and `indices` can have `M` leading batch
  dimensions that exactly match. In this case `batch_dims` must be set to `M`.

  For example, to collect one row from each of a batch of matrices you could
  set the leading elements of the index vectors to be their location in the
  batch:

  >>> tf.gather_nd(
  ...     indices = [[0, 1],
  ...                [1, 0],
  ...                [2, 4],
  ...                [3, 2],
  ...                [4, 1]],
  ...     params=tf.zeros([5, 7, 3])).shape.as_list()
  [5, 3]

  The `batch_dims` argument lets you omit those leading location dimensions
  from the index:

  >>> tf.gather_nd(
  ...     batch_dims=1,
  ...     indices = [[1],
  ...                [0],
  ...                [4],
  ...                [2],
  ...                [1]],
  ...     params=tf.zeros([5, 7, 3])).shape.as_list()
  [5, 3]

  This is equivalent to caling a separate `gather_nd` for each location in the
  batch dimensions.


  >>> params=tf.zeros([5, 7, 3])
  >>> indices=tf.zeros([5, 1])
  >>> batch_dims = 1
  >>>
  >>> index_depth = indices.shape[-1]
  >>> batch_shape = indices.shape[:batch_dims]
  >>> assert params.shape[:batch_dims] == batch_shape
  >>> outer_shape = indices.shape[batch_dims:-1]
  >>> assert index_depth <= params.shape.rank
  >>> inner_shape = params.shape[batch_dims + index_depth:]
  >>> output_shape = batch_shape + outer_shape + inner_shape
  >>> output_shape.as_list()
  [5, 3]

  ### More examples

  Indexing into a 3-tensor:

  >>> tf.gather_nd(
  ...     indices = [[1]],
  ...     params = [[['a0', 'b0'], ['c0', 'd0']],
  ...               [['a1', 'b1'], ['c1', 'd1']]]).numpy()
  array([[[b'a1', b'b1'],
          [b'c1', b'd1']]], dtype=object)


  >>> tf.gather_nd(
  ...     indices = [[0, 1], [1, 0]],
  ...     params = [[['a0', 'b0'], ['c0', 'd0']],
  ...               [['a1', 'b1'], ['c1', 'd1']]]).numpy()
  array([[b'c0', b'd0'],
         [b'a1', b'b1']], dtype=object)


  >>> tf.gather_nd(
  ...     indices = [[0, 0, 1], [1, 0, 1]],
  ...     params = [[['a0', 'b0'], ['c0', 'd0']],
  ...               [['a1', 'b1'], ['c1', 'd1']]]).numpy()
  array([b'b0', b'b1'], dtype=object)

  The examples below are for the case when only indices have leading extra
  dimensions. If both 'params' and 'indices' have leading batch dimensions, use
  the 'batch_dims' parameter to run gather_nd in batch mode.

  Batched indexing into a matrix:

  >>> tf.gather_nd(
  ...     indices = [[[0, 0]], [[0, 1]]],
  ...     params = [['a', 'b'], ['c', 'd']]).numpy()
  array([[b'a'],
         [b'b']], dtype=object)


  Batched slice indexing into a matrix:

  >>> tf.gather_nd(
  ...     indices = [[[1]], [[0]]],
  ...     params = [['a', 'b'], ['c', 'd']]).numpy()
  array([[[b'c', b'd']],
         [[b'a', b'b']]], dtype=object)


  Batched indexing into a 3-tensor:

  >>> tf.gather_nd(
  ...     indices = [[[1]], [[0]]],
  ...     params = [[['a0', 'b0'], ['c0', 'd0']],
  ...               [['a1', 'b1'], ['c1', 'd1']]]).numpy()
  array([[[[b'a1', b'b1'],
           [b'c1', b'd1']]],
         [[[b'a0', b'b0'],
           [b'c0', b'd0']]]], dtype=object)


  >>> tf.gather_nd(
  ...     indices = [[[0, 1], [1, 0]], [[0, 0], [1, 1]]],
  ...     params = [[['a0', 'b0'], ['c0', 'd0']],
  ...               [['a1', 'b1'], ['c1', 'd1']]]).numpy()
  array([[[b'c0', b'd0'],
          [b'a1', b'b1']],
         [[b'a0', b'b0'],
          [b'c1', b'd1']]], dtype=object)

  >>> tf.gather_nd(
  ...     indices = [[[0, 0, 1], [1, 0, 1]], [[0, 1, 1], [1, 1, 0]]],
  ...     params = [[['a0', 'b0'], ['c0', 'd0']],
  ...               [['a1', 'b1'], ['c1', 'd1']]]).numpy()
  array([[b'b0', b'b1'],
         [b'd0', b'c1']], dtype=object)


  Examples with batched 'params' and 'indices':

  >>> tf.gather_nd(
  ...     batch_dims = 1,
  ...     indices = [[1],
  ...                [0]],
  ...     params = [[['a0', 'b0'],
  ...                ['c0', 'd0']],
  ...               [['a1', 'b1'],
  ...                ['c1', 'd1']]]).numpy()
  array([[b'c0', b'd0'],
         [b'a1', b'b1']], dtype=object)


  >>> tf.gather_nd(
  ...     batch_dims = 1,
  ...     indices = [[[1]], [[0]]],
  ...     params = [[['a0', 'b0'], ['c0', 'd0']],
  ...               [['a1', 'b1'], ['c1', 'd1']]]).numpy()
  array([[[b'c0', b'd0']],
         [[b'a1', b'b1']]], dtype=object)

  >>> tf.gather_nd(
  ...     batch_dims = 1,
  ...     indices = [[[1, 0]], [[0, 1]]],
  ...     params = [[['a0', 'b0'], ['c0', 'd0']],
  ...               [['a1', 'b1'], ['c1', 'd1']]]).numpy()
  array([[b'c0'],
         [b'b1']], dtype=object)


  See also `tf.gather`.

  Args:
    params: A `Tensor`. The tensor from which to gather values.
    indices: A `Tensor`. Must be one of the following types: `int32`, `int64`.
      Index tensor.
    name: A name for the operation (optional).
    batch_dims: An integer or a scalar 'Tensor'. The number of batch dimensions.
    bad_indices_policy: A string. If `""` or `"DEFAULT"`, the default behavior
      is used (error on CPU and ignore on GPU). If `"IGNORE"`, the bad indices
      are ignored and 0 is stored in the corresponding output value. If
      `"ERROR"`, an error is raised. Accelerators generally don't support
      `"ERROR"`.

  Returns:
    A `Tensor`. Has the same type as `params`.
  """
  batch_dims_ = tensor_util.constant_value(batch_dims)
  if batch_dims_ is not None:
    batch_dims = int(batch_dims_)
  if batch_dims == 0 and bad_indices_policy not in ("", "DEFAULT"):
    # TODO(cylai): also support `bad_indices_policy` for resource variables.
    return gen_array_ops.gather_nd(
        params, indices, name=name, bad_indices_policy=bad_indices_policy
    )
  if batch_dims == 0:
    try:
      # TODO(apassos) find a less bad way of detecting resource variables
      # without introducing a circular dependency.
      return params.gather_nd(indices, name=name)
    except AttributeError:
      return gen_array_ops.gather_nd(
          params, indices, name=name, bad_indices_policy=bad_indices_policy
      )
  else:
    return batch_gather_nd(
        params,
        indices,
        batch_dims=batch_dims,
        name=name,
        bad_indices_policy=bad_indices_policy,
    )


@tf_export("gather_nd", v1=[])
@dispatch.add_dispatch_support
def gather_nd_v2(
    params, indices, batch_dims=0, name=None, bad_indices_policy=""
):
  return gather_nd(
      params,
      indices,
      name=name,
      batch_dims=batch_dims,
      bad_indices_policy=bad_indices_policy,
  )


gather_nd_v2.__doc__ = gather_nd.__doc__


def batch_gather_nd(
    params, indices, batch_dims, name=None, bad_indices_policy=""
):
  """gather_nd implementation with batch support."""
  with ops.name_scope(name, "BatchGatherND", [params, indices]):
    indices = ops.convert_to_tensor(indices, name="indices")
    params = ops.convert_to_tensor(params, name="params")

    if not isinstance(batch_dims, int):
      raise TypeError(f"Argument `batch_dims` must be an int; got {batch_dims}")
    if batch_dims < 0:
      raise ValueError("tf.gather_nd does not allow negative batch_dims.")
    params_ndims = params.shape.ndims
    indices_ndims = indices.shape.ndims
    if indices_ndims is not None and batch_dims >= indices_ndims:
      raise ValueError(f"Argument `batch_dims` = {batch_dims} must be "
                       f"less than rank(`indices`) = {indices_ndims}")
    if params_ndims is not None and batch_dims >= params_ndims:
      raise ValueError(f"Argument `batch_dims` = {batch_dims} must be "
                       f"less than rank(`params`) = {params_ndims}")

    expand = batch_dims == 0
    if expand:
      # Normally gather_nd will be called when batch_dims == 0.
      # But if this function is called with batch_dims = 0, e.g. for testing
      # purposes, this adds a dummy batch dimension to make batch_dims = 1.
      params = expand_dims(params, axis=0)
      indices = expand_dims(indices, axis=0)
      batch_dims = 1

    params_shape = shape(params)
    indices_shape = shape(indices)
    batch_shape = params_shape[:batch_dims]
    batch_size = gen_math_ops.prod(batch_shape, [0])
    index_internal_ndims = rank(indices) - batch_dims - 1
    indices_internal_shape = indices_shape[batch_dims:-1]

    # Assuming a 'params' with shape [b1, ..., bM, g1, ..., gN] and an 'indices'
    # with shape [b1, ..., bM, i1, ..., iK, C], where C <= N, we need to modify
    # 'indices' s.t. it has shape [i1, ..., iK, D], where D <= M + N and slices
    # to the entire 'params' tensor.
    # Assuming we have a batch of shape [B1, B2], we use meshgrid to create a
    # grid of size B1 x B2.
    batch_dim_list = array_ops_stack.unstack(batch_shape, axis=0)
    dim_ranges = [
        gen_math_ops.cast(
            gen_math_ops._range(0, gen_math_ops.cast(x, dtypes.int32), 1),
            indices.dtype,
        )
        for x in batch_dim_list
    ]
    mesh_list = meshgrid(*dim_ranges, indexing="ij") if dim_ranges else []
    # Then we flatten and stack the tensors to form a (B1.B2) by 2 matrix.
    flat_list = [reshape(x, shape=(-1,)) for x in mesh_list]
    index_grid = transpose(array_ops_stack.stack(flat_list, axis=0))
    # We need to concatenate these batch coordinates with the internal indices.
    # concat -> index_grid [B1.B2, 2] with indices [i1, ..., iK, C]
    # So we reshape them both to [(B1.B2), i1, ..., iK, *]
    index_grid_shape = shape(index_grid)
    index_grid = reshape(
        index_grid,
        concat(
            [
                index_grid_shape[:1],
                ones(index_internal_ndims, dtype=index_grid_shape.dtype),
                index_grid_shape[1:],
            ],
            axis=0,
        ),
    )
    tile_shape = concat(((1,), indices_internal_shape, (1,)), axis=0)
    index_grid = tile(index_grid, multiples=tile_shape)
    # index_grid now has shape [(B1.B2), i1, ..., iK, 2]
    flat_shape = concat(([batch_size], indices_shape[batch_dims:]), axis=0)
    flat_indices = reshape(indices, shape=flat_shape)
    # flat_indices now has shape [(B1.B2), i1, ..., iK, C]
    indices = concat((index_grid, flat_indices), axis=-1)
    # indices has shape [(B1.B2), i1, ..., iK, 2+C]
    out = gen_array_ops.gather_nd(
        params, indices, bad_indices_policy=bad_indices_policy
    )
    # out has shape [(B1.B2), i1, ..., iK, N-C]. Now we reshape batch to
    # its original form.
    out_shape = shape(out)
    out = reshape(out, shape=concat((batch_shape, out_shape[1:]), axis=0))
    if expand:
      out = squeeze(out, axis=0)
  return out


@deprecation.deprecated_endpoints("tensor_scatter_update")
@tf_export(
    "tensor_scatter_nd_update",
    v1=["tensor_scatter_nd_update", "tensor_scatter_update"])
@dispatch.add_dispatch_support
def tensor_scatter_nd_update(tensor, indices, updates, name=None):
  """Scatter `updates` into an existing tensor according to `indices`.

  This operation creates a new tensor by applying sparse `updates` to the
  input `tensor`. This is similar to an index assignment.

  ```
  # Not implemented: tensors cannot be updated inplace.
  tensor[indices] = updates
  ```

  If an out of bound index is found on CPU, an error is returned.

  > **WARNING**: There are some GPU specific semantics for this operation.
  >
  > - If an out of bound index is found, the index is ignored.
  > - The order in which updates are applied is nondeterministic, so the output
  >   will be nondeterministic if `indices` contains duplicates.

  This operation is very similar to `tf.scatter_nd`, except that the updates are
  scattered onto an existing tensor (as opposed to a zero-tensor). If the memory
  for the existing tensor cannot be re-used, a copy is made and updated.

  In general:

  * `indices` is an integer tensor - the indices to update in `tensor`.
  * `indices` has **at least two** axes, the last axis is the depth of the
    index vectors.
  * For each index vector in `indices` there is a corresponding entry in
    `updates`.
  * If the length of the index vectors matches the rank of the `tensor`, then
    the index vectors each point to scalars in `tensor` and each update is a
    scalar.
  * If the length of the index vectors is less than the rank of `tensor`, then
    the index vectors each point to the slices of `tensor` and shape of the updates
    must match that slice.

  Overall this leads to the following shape constraints:

  ```
  assert tf.rank(indices) >= 2
  index_depth = indices.shape[-1]
  batch_shape = indices.shape[:-1]
  assert index_depth <= tf.rank(tensor)
  outer_shape = tensor.shape[:index_depth]
  inner_shape = tensor.shape[index_depth:]
  assert updates.shape == batch_shape + inner_shape
  ```

  Typical usage is often much simpler than this general form, and it
  can be better understood starting with simple examples:

  ### Scalar updates

  The simplest usage inserts scalar elements into a tensor by index.
  In this case, the `index_depth` must equal the rank of the
  input `tensor`, slice each column of `indices` is an index into an axis of the
  input `tensor`.

  In this simplest case the shape constraints are:

  ```
  num_updates, index_depth = indices.shape.as_list()
  assert updates.shape == [num_updates]
  assert index_depth == tf.rank(tensor)`
  ```

  For example, to insert 4 scattered elements in a rank-1 tensor with
  8 elements.

  <div style="width:70%; margin:auto; margin-bottom:10px; margin-top:20px;">
  <img style="width:100%"
    src="https://www.tensorflow.org/images/ScatterNd1.png">
  </div>

  This scatter operation would look like this:

  >>> tensor = [0, 0, 0, 0, 0, 0, 0, 0]    # tf.rank(tensor) == 1
  >>> indices = [[1], [3], [4], [7]]       # num_updates == 4, index_depth == 1
  >>> updates = [9, 10, 11, 12]            # num_updates == 4
  >>> print(tf.tensor_scatter_nd_update(tensor, indices, updates))
  tf.Tensor([ 0 9  0 10  11  0  0 12], shape=(8,), dtype=int32)

  The length (first axis) of `updates` must equal the length of the `indices`:
  `num_updates`. This is the number of updates being inserted. Each scalar
  update is inserted into `tensor` at the indexed location.

  For a higher rank input `tensor` scalar updates can be inserted by using an
  `index_depth` that matches `tf.rank(tensor)`:

  >>> tensor = [[1, 1], [1, 1], [1, 1]]    # tf.rank(tensor) == 2
  >>> indices = [[0, 1], [2, 0]]           # num_updates == 2, index_depth == 2
  >>> updates = [5, 10]                    # num_updates == 2
  >>> print(tf.tensor_scatter_nd_update(tensor, indices, updates))
  tf.Tensor(
      [[ 1  5]
       [ 1  1]
       [10  1]], shape=(3, 2), dtype=int32)

  ### Slice updates

  When the input `tensor` has more than one axis scatter can be used to update
  entire slices.

  In this case it's helpful to think of the input `tensor` as being a two level
  array-of-arrays. The shape of this two level array is split into the
  `outer_shape` and the `inner_shape`.

  `indices` indexes into the outer level of the input tensor (`outer_shape`).
  and replaces the sub-array at that location with the corresponding item from
  the `updates` list. The shape of each update is `inner_shape`.

  When updating a list of slices the shape constraints are:

  ```
  num_updates, index_depth = indices.shape.as_list()
  outer_shape = tensor.shape[:index_depth]
  inner_shape = tensor.shape[index_depth:]
  assert updates.shape == [num_updates, inner_shape]
  ```

  For example, to update rows of a `(6, 3)` `tensor`:

  >>> tensor = tf.zeros([6, 3], dtype=tf.int32)

  Use an index depth of one.

  >>> indices = tf.constant([[2], [4]])     # num_updates == 2, index_depth == 1
  >>> num_updates, index_depth = indices.shape.as_list()

  The `outer_shape` is `6`, the inner shape is `3`:

  >>> outer_shape = tensor.shape[:index_depth]
  >>> inner_shape = tensor.shape[index_depth:]

  2 rows are being indexed so 2 `updates` must be supplied.
  Each update must be shaped to match the `inner_shape`.

  >>> # num_updates == 2, inner_shape==3
  >>> updates = tf.constant([[1, 2, 3],
  ...                        [4, 5, 6]])

  Altogether this gives:

  >>> tf.tensor_scatter_nd_update(tensor, indices, updates).numpy()
  array([[0, 0, 0],
         [0, 0, 0],
         [1, 2, 3],
         [0, 0, 0],
         [4, 5, 6],
         [0, 0, 0]], dtype=int32)

  #### More slice update examples

  A tensor representing a batch of uniformly sized video clips naturally has 5
  axes: `[batch_size, time, width, height, channels]`.

  For example:

  >>> batch_size, time, width, height, channels = 13,11,7,5,3
  >>> video_batch = tf.zeros([batch_size, time, width, height, channels])

  To replace a selection of video clips:
    * Use an `index_depth` of 1 (indexing the `outer_shape`: `[batch_size]`)
    * Provide updates each with a shape matching the `inner_shape`:
      `[time, width, height, channels]`.

  To replace the first two clips with ones:

  >>> indices = [[0],[1]]
  >>> new_clips = tf.ones([2, time, width, height, channels])
  >>> tf.tensor_scatter_nd_update(video_batch, indices, new_clips)

  To replace a selection of frames in the videos:

  * `indices` must have an `index_depth` of 2 for the `outer_shape`:
    `[batch_size, time]`.
  * `updates` must be shaped like a list of images.  Each update must have a
    shape, matching the `inner_shape`: `[width, height, channels]`.

  To replace the first frame of the first three video clips:

  >>> indices = [[0, 0], [1, 0], [2, 0]] # num_updates=3, index_depth=2
  >>> new_images = tf.ones([
  ...   # num_updates=3, inner_shape=(width, height, channels)
  ...   3, width, height, channels])
  >>> tf.tensor_scatter_nd_update(video_batch, indices, new_images)

  ### Folded indices

  In simple cases it's convenient to think of `indices` and `updates` as
  lists, but this is not a strict requirement. Instead of a flat `num_updates`,
  the `indices` and `updates` can be folded into a `batch_shape`. This
  `batch_shape` is all axes of the `indices`, except for the innermost
  `index_depth` axis.

  ```
  index_depth = indices.shape[-1]
  batch_shape = indices.shape[:-1]
  ```

  Note: The one exception is that the `batch_shape` cannot be `[]`. You can't
  update a single index by passing indices with shape `[index_depth]`.

  `updates` must have a matching `batch_shape` (the axes before `inner_shape`).

  ```
  assert updates.shape == batch_shape + inner_shape
  ```

  Note: The result is equivalent to flattening the `batch_shape` axes of
  `indices` and `updates`. This generalization just avoids the need
  for reshapes when it is more natural to construct "folded" indices and
  updates.

  With this generalization the full shape constraints are:

  ```
  assert tf.rank(indices) >= 2
  index_depth = indices.shape[-1]
  batch_shape = indices.shape[:-1]
  assert index_depth <= tf.rank(tensor)
  outer_shape = tensor.shape[:index_depth]
  inner_shape = tensor.shape[index_depth:]
  assert updates.shape == batch_shape + inner_shape
  ```

  For example, to draw an `X` on a `(5,5)` matrix start with these indices:

  >>> tensor = tf.zeros([5,5])
  >>> indices = tf.constant([
  ...  [[0,0],
  ...   [1,1],
  ...   [2,2],
  ...   [3,3],
  ...   [4,4]],
  ...  [[0,4],
  ...   [1,3],
  ...   [2,2],
  ...   [3,1],
  ...   [4,0]],
  ... ])
  >>> indices.shape.as_list()  # batch_shape == [2, 5], index_depth == 2
  [2, 5, 2]

  Here the `indices` do not have a shape of `[num_updates, index_depth]`, but a
  shape of `batch_shape+[index_depth]`.

  Since the `index_depth` is equal to the rank of `tensor`:

  * `outer_shape` is `(5,5)`
  * `inner_shape` is `()` - each update is scalar
  * `updates.shape` is `batch_shape + inner_shape == (5,2) + ()`

  >>> updates = [
  ...   [1,1,1,1,1],
  ...   [1,1,1,1,1],
  ... ]

  Putting this together gives:

  >>> tf.tensor_scatter_nd_update(tensor, indices, updates).numpy()
  array([[1., 0., 0., 0., 1.],
         [0., 1., 0., 1., 0.],
         [0., 0., 1., 0., 0.],
         [0., 1., 0., 1., 0.],
         [1., 0., 0., 0., 1.]], dtype=float32)

  Args:
    tensor: Tensor to copy/update.
    indices: Indices to update.
    updates: Updates to apply at the indices.
    name: Optional name for the operation.

  Returns:
    A new tensor with the given shape and updates applied according to the
    indices.
  """
  return gen_array_ops.tensor_scatter_update(
      tensor=tensor, indices=indices, updates=updates, name=name)


# Define quantize_v2 here in order to make name the second-to-last attribute,
# because round_mode was added later.
# (And also now because of 'axis' processing).
@tf_export(v1=["quantize_v2"])
@dispatch.add_dispatch_support
@deprecation.deprecated(
    "2017-10-25",
    "`tf.quantize_v2` is deprecated, please use `tf.quantization.quantize` "
    "instead.")  # pylint: disable=missing-docstring
def quantize_v2(
    input,  # pylint: disable=redefined-builtin
    min_range,
    max_range,
    T,
    mode="MIN_COMBINED",
    name=None,
    round_mode="HALF_AWAY_FROM_ZERO",
    narrow_range=False,
    axis=None,
    ensure_minimum_range=0.01):
  if axis is None:
    axis = -1
  elif axis < 0:
    if input.shape.ndims is None:
      raise ValueError("input should have known rank to use negative axis.")
    axis %= input.shape.ndims

  if ensure_minimum_range != 0.01:
    return gen_array_ops.quantize_v2(
        input,
        min_range,
        max_range,
        T=T,
        mode=mode,
        name=name,
        round_mode=round_mode,
        narrow_range=narrow_range,
        axis=axis,
        ensure_minimum_range=ensure_minimum_range)
  return gen_array_ops.quantize_v2(
      input,
      min_range,
      max_range,
      T=T,
      mode=mode,
      name=name,
      round_mode=round_mode,
      narrow_range=narrow_range,
      axis=axis)


quantize_v2.__doc__ = """Please use `tf.quantization.quantize` instead."""


# We want to expose tf.quantization.quantize instead of
# tf.quantization.quantize; we can deprecate tf.quantization.quantize in next
# version of TensorFlow.
@tf_export("quantization.quantize", v1=["quantization.quantize", "quantize"])
@dispatch.add_dispatch_support
@deprecation.deprecated_endpoints("quantize")
def quantize(
    input,  # pylint: disable=redefined-builtin
    min_range,
    max_range,
    T,
    mode="MIN_COMBINED",
    round_mode="HALF_AWAY_FROM_ZERO",
    name=None,
    narrow_range=False,
    axis=None,
    ensure_minimum_range=0.01):
  """Quantize the input tensor."""
  if ensure_minimum_range != 0.01:
    return quantize_v2(
        input,
        min_range,
        max_range,
        T,
        mode=mode,
        round_mode=round_mode,
        name=name,
        narrow_range=narrow_range,
        axis=axis,
        ensure_minimum_range=ensure_minimum_range)
  return quantize_v2(
      input,
      min_range,
      max_range,
      T,
      mode=mode,
      round_mode=round_mode,
      name=name,
      narrow_range=narrow_range,
      axis=axis)


@tf_export("quantization.dequantize", v1=["quantization.dequantize",
                                          "dequantize"])
@dispatch.add_dispatch_support
@deprecation.deprecated_endpoints("dequantize")
def dequantize(  # pylint: disable=missing-docstring
    input,  # pylint: disable=redefined-builtin
    min_range,
    max_range,
    mode="MIN_COMBINED",
    name=None,
    axis=None,
    narrow_range=False,
    dtype=dtypes.float32):
  if axis is None:
    axis = -1
  elif axis < 0:
    if input.shape.ndims is None:
      raise ValueError("input should have known rank to use negative axis.")
    axis %= input.shape.ndims

  if axis >= 0 or narrow_range:
    return gen_array_ops.dequantize(
        input,
        min_range,
        max_range,
        mode=mode,
        name=name,
        narrow_range=narrow_range,
        axis=axis,
        dtype=dtype)
  return gen_array_ops.dequantize(
      input, min_range, max_range, mode=mode, name=name, dtype=dtype)


dequantize.__doc__ = gen_array_ops.dequantize.__doc__


@tf_export("quantization.quantize_and_dequantize")
@dispatch.add_dispatch_support
@deprecation.deprecated(None,
                        "This Op has been deprecated, use" +
                        "`quantize_and_dequantize_v2` instead. To " +
                        "To simulate the V1 the behavior of " +
                        "tf.quantization.quantize_and_dequantize(...) use " +
                        "tf.grad_pass_through(" +
                        "tf.quantization.quantize_and_dequantize_v2)(...).")
def quantize_and_dequantize(
    input,  # pylint: disable=redefined-builtin
    input_min,
    input_max,
    signed_input=True,
    num_bits=8,
    range_given=False,
    round_mode="HALF_TO_EVEN",
    name=None,
    narrow_range=False,
    axis=None):
  """Quantizes then dequantizes a tensor.

  Args:
    input: A `Tensor` to quantize and dequantize.
    input_min: If range_given=True, the minimum input value, that needs to be
      represented in the quantized representation. If axis is specified, this
      should be a vector of minimum values for each slice along axis.
    input_max: If range_given=True, the maximum input value that needs to be
      represented in the quantized representation. If axis is specified, this
      should be a vector of maximum values for each slice along axis.
    signed_input: True if the quantization is signed or unsigned.
    num_bits: The bitwidth of the quantization.
    range_given: If true use `input_min` and `input_max` for the range of the
      input, otherwise determine min and max from the input `Tensor`.
    round_mode: Rounding mode when rounding from float values to quantized ones.
      one of ['HALF_TO_EVEN', 'HALF_UP']
    name: Optional name for the operation.
    narrow_range: If true, then the absolute value of the quantized minimum
      value is the same as the quantized maximum value, instead of 1 greater.
      i.e. for 8 bit quantization, the minimum value is -127 instead of -128.
    axis: Integer. If specified, refers to a dimension of the input tensor, such
      that quantization will be per slice along that dimension.

  Returns:
    A `Tensor`. Each element is the result of quantizing and dequantizing the
    corresponding element of `input`.
  """
  if axis is None:
    axis = -1
  elif axis < 0:
    if input.shape.ndims is None:
      raise ValueError("input should have known rank to use negative axis.")
    axis %= input.shape.ndims

  return gen_array_ops.quantize_and_dequantize_v2(
      input,
      input_min=input_min,
      input_max=input_max,
      signed_input=signed_input,
      num_bits=num_bits,
      range_given=range_given,
      round_mode=round_mode,
      narrow_range=narrow_range,
      axis=axis,
      name=name)


@tf_export("quantization.quantize_and_dequantize_v2")
@dispatch.add_dispatch_support
def quantize_and_dequantize_v2(
    input,  # pylint: disable=redefined-builtin
    input_min,
    input_max,
    signed_input=True,
    num_bits=8,
    range_given=False,
    round_mode="HALF_TO_EVEN",
    name=None,
    narrow_range=False,
    axis=None):
  """Quantizes then dequantizes a tensor.

  Updates the gradient definition for quantization that is outside the range to
  be 0.To simulate the V1 the behavior of
  tf.quantization.quantize_and_dequantize(...) use
  tf.grad_pass_through(tf.quantization.quantize_and_dequantize_v2)(...).

  Example usage:

  ```python
  def getQuantizeOp(input):
      input_tensor = tf.placeholder(tf.float32, shape=[4, 4])
      net = tf.quantization.quantize_and_dequantize(input,
                                                    input_min=min_threshold,
                                                    input_max=max_threshold,
                                                    range_given=True)

  To simulate v1 behavior:

  def testDecomposeQuantizeDequantize(self):
      def f(input_tensor):
        return tf.quantization.quantize_and_dequantize_v2(input_tensor,
                                                          input_min = 5.0,
                                                          input_max= -10.0,
                                                          range_given=True)
      input_tensor = tf.placeholder(tf.float32, shape=[4, 4])
      net = tf.grad_pass_through(f)(input_tensor)
  ```

  Args:
    input: A `Tensor` to quantize and dequantize.
    input_min: If range_given=True, the minimum input value, that needs to be
      represented in the quantized representation. If axis is specified, this
      should be a vector of minimum values for each slice along axis.
    input_max: If range_given=True, the maximum input value that needs to be
      represented in the quantized representation. If axis is specified, this
      should be a vector of maximum values for each slice along axis.
    signed_input: True if the quantization is signed or unsigned.
    num_bits: The bitwidth of the quantization.
    range_given: If true use `input_min` and `input_max` for the range of the
      input, otherwise determine min and max from the input `Tensor`.
    round_mode: Rounding mode when rounding from float values to quantized ones.
      one of ['HALF_TO_EVEN', 'HALF_UP']
    name: Optional name for the operation.
    narrow_range: If true, then the absolute value of the quantized minimum
      value is the same as the quantized maximum value, instead of 1 greater.
      i.e. for 8 bit quantization, the minimum value is -127 instead of -128.
    axis: Integer. If specified, refers to a dimension of the input tensor, such
      that quantization will be per slice along that dimension.

  Returns:
    A `Tensor`. Each element is the result of quantizing and dequantizing the
    corresponding element of `input`.
  """
  if axis is None:
    axis = -1
  elif axis < 0:
    if input.shape.ndims is None:
      raise ValueError("input should have known rank to use negative axis.")
    axis %= input.shape.ndims

  return gen_array_ops.quantize_and_dequantize_v4(
      input,
      input_min=input_min,
      input_max=input_max,
      signed_input=signed_input,
      num_bits=num_bits,
      range_given=range_given,
      round_mode=round_mode,
      narrow_range=narrow_range,
      axis=axis,
      name=name)


@tf_export("searchsorted")
@dispatch.add_dispatch_support
def searchsorted(sorted_sequence,
                 values,
                 side="left",
                 out_type=dtypes.int32,
                 name=None):
  """Searches for where a value would go in a sorted sequence.

  This is not a method for checking containment (like python `in`).

  The typical use case for this operation is "binning", "bucketing", or
  "discretizing". The `values` are assigned to bucket-indices based on the
  **edges** listed in `sorted_sequence`. This operation
  returns the bucket-index for each value.

  >>> edges = [-1, 3.3, 9.1, 10.0]
  >>> values = [0.0, 4.1, 12.0]
  >>> tf.searchsorted(edges, values).numpy()
  array([1, 2, 4], dtype=int32)

  The `side` argument controls which index is returned if a value lands exactly
  on an edge:

  >>> seq = [0, 3, 9, 10, 10]
  >>> values = [0, 4, 10]
  >>> tf.searchsorted(seq, values).numpy()
  array([0, 2, 3], dtype=int32)
  >>> tf.searchsorted(seq, values, side="right").numpy()
  array([1, 2, 5], dtype=int32)

  The `axis` is not settable for this operation. It always operates on the
  innermost dimension (`axis=-1`). The operation will accept any number of
  outer dimensions. Here it is applied to the rows of a matrix:

  >>> sorted_sequence = [[0., 3., 8., 9., 10.],
  ...                    [1., 2., 3., 4., 5.]]
  >>> values = [[9.8, 2.1, 4.3],
  ...           [0.1, 6.6, 4.5, ]]
  >>> tf.searchsorted(sorted_sequence, values).numpy()
  array([[4, 1, 2],
         [0, 5, 4]], dtype=int32)

  Note: This operation assumes that `sorted_sequence` **is sorted** along the
  innermost axis, maybe using `tf.sort(..., axis=-1)`. **If the sequence is not
  sorted, no error is raised** and the content of the returned tensor is not well
  defined.

  Args:
    sorted_sequence: N-D `Tensor` containing a sorted sequence.
    values: N-D `Tensor` containing the search values.
    side: 'left' or 'right'; 'left' corresponds to lower_bound and 'right' to
      upper_bound.
    out_type: The output type (`int32` or `int64`).  Default is `tf.int32`.
    name: Optional name for the operation.

  Returns:
    An N-D `Tensor` the size of `values` containing the result of applying
    either lower_bound or upper_bound (depending on side) to each value.  The
    result is not a global index to the entire `Tensor`, but the index in the
    last dimension.

  Raises:
    ValueError: If the last dimension of `sorted_sequence >= 2^31-1` elements.
                If the total size of `values` exceeds `2^31 - 1` elements.
                If the first `N-1` dimensions of the two tensors don't match.
  """
  sequence_size = shape_internal(sorted_sequence)[-1]
  values_size = shape_internal(values)[-1]
  sorted_sequence_2d = reshape(sorted_sequence, [-1, sequence_size])
  values_2d = reshape(values, [-1, values_size])
  if side == "right":
    output = gen_array_ops.upper_bound(sorted_sequence_2d, values_2d, out_type,
                                       name)
  elif side == "left":
    output = gen_array_ops.lower_bound(sorted_sequence_2d, values_2d, out_type,
                                       name)
  else:
    raise ValueError("Argument `side` must be either 'right' or 'left'. "
                     f"Received: `side` = '{side}'.")
  return reshape(output, shape_internal(values))


quantize.__doc__ = gen_array_ops.quantize_v2.__doc__


@tf_export("image.extract_patches")
@dispatch.add_dispatch_support
def extract_image_patches_v2(images, sizes, strides, rates, padding, name=None):
  r"""Extract `patches` from `images`.

  This op collects patches from the input image, as if applying a
  convolution. All extracted patches are stacked in the depth (last) dimension
  of the output.

  Specifically, the op extracts patches of shape `sizes` which are `strides`
  apart in the input image. The output is subsampled using the `rates` argument,
  in the same manner as "atrous" or "dilated" convolutions.

  The result is a 4D tensor which is indexed by batch, row, and column.
  `output[i, x, y]` contains a flattened patch of size `sizes[1], sizes[2]`
  which is taken from the input starting at
  `images[i, x*strides[1], y*strides[2]]`.

  Each output patch can be reshaped to `sizes[1], sizes[2], depth`, where
  `depth` is `images.shape[3]`.

  The output elements are taken from the input at intervals given by the `rate`
  argument, as in dilated convolutions.

  The `padding` argument has no effect on the size of each patch, it determines
  how many patches are extracted. If `VALID`, only patches which are fully
  contained in the input image are included. If `SAME`, all patches whose
  starting point is inside the input are included, and areas outside the input
  default to zero.

  Example:

  ```
    n = 10
    # images is a 1 x 10 x 10 x 1 array that contains the numbers 1 through 100
    images = [[[[x * n + y + 1] for y in range(n)] for x in range(n)]]

    # We generate two outputs as follows:
    # 1. 3x3 patches with stride length 5
    # 2. Same as above, but the rate is increased to 2
    tf.image.extract_patches(images=images,
                             sizes=[1, 3, 3, 1],
                             strides=[1, 5, 5, 1],
                             rates=[1, 1, 1, 1],
                             padding='VALID')

    # Yields:
    [[[[ 1  2  3 11 12 13 21 22 23]
       [ 6  7  8 16 17 18 26 27 28]]
      [[51 52 53 61 62 63 71 72 73]
       [56 57 58 66 67 68 76 77 78]]]]
  ```

  If we mark the pixels in the input image which are taken for the output with
  `*`, we see the pattern:

  ```
     *  *  *  4  5  *  *  *  9 10
     *  *  * 14 15  *  *  * 19 20
     *  *  * 24 25  *  *  * 29 30
    31 32 33 34 35 36 37 38 39 40
    41 42 43 44 45 46 47 48 49 50
     *  *  * 54 55  *  *  * 59 60
     *  *  * 64 65  *  *  * 69 70
     *  *  * 74 75  *  *  * 79 80
    81 82 83 84 85 86 87 88 89 90
    91 92 93 94 95 96 97 98 99 100
  ```

  ```
    tf.image.extract_patches(images=images,
                             sizes=[1, 3, 3, 1],
                             strides=[1, 5, 5, 1],
                             rates=[1, 2, 2, 1],
                             padding='VALID')

    # Yields:
    [[[[  1   3   5  21  23  25  41  43  45]
       [  6   8  10  26  28  30  46  48  50]]

      [[ 51  53  55  71  73  75  91  93  95]
       [ 56  58  60  76  78  80  96  98 100]]]]
  ```

  We can again draw the effect, this time using the symbols `*`, `x`, `+` and
  `o` to distinguish the patches:

  ```
     *  2  *  4  *  x  7  x  9  x
    11 12 13 14 15 16 17 18 19 20
     * 22  * 24  *  x 27  x 29  x
    31 32 33 34 35 36 37 38 39 40
     * 42  * 44  *  x 47  x 49  x
     + 52  + 54  +  o 57  o 59  o
    61 62 63 64 65 66 67 68 69 70
     + 72  + 74  +  o 77  o 79  o
    81 82 83 84 85 86 87 88 89 90
     + 92  + 94  +  o 97  o 99  o
  ```

  Args:
    images: A 4-D Tensor with shape `[batch, in_rows, in_cols, depth]`.
    sizes: The size of the extracted patches. Must be
      `[1, size_rows, size_cols, 1]`.
    strides: A 1-D Tensor of length 4. How far the centers of two consecutive
      patches are in the images. Must be: `[1, stride_rows, stride_cols, 1]`.
    rates: A 1-D Tensor of length 4. Must be: `[1, rate_rows, rate_cols, 1]`.
      This is the input stride, specifying how far two consecutive patch samples
      are in the input. Equivalent to extracting patches with `patch_sizes_eff =
      patch_sizes + (patch_sizes - 1) * (rates - 1)`, followed by subsampling
      them spatially by a factor of `rates`. This is equivalent to `rate` in
      dilated (a.k.a. Atrous) convolutions.
    padding: The type of padding algorithm to use.
    name: A name for the operation (optional).

  Returns:
    A 4-D Tensor of the same type as the input.
  """
  return gen_array_ops.extract_image_patches(images, sizes, strides, rates,
                                             padding, name)


@tf_export(v1=["image.extract_image_patches", "extract_image_patches"])
@dispatch.add_dispatch_support
@deprecation.deprecated_args(None, "ksizes is deprecated, use sizes instead",
                             "ksizes")
def extract_image_patches(  # pylint: disable=missing-docstring
    images,
    ksizes=None,
    strides=None,
    rates=None,
    padding=None,
    name=None,
    sizes=None):
  """Extract patches from images and put them in the "depth" output dimension.

  Args:
    `images`: A `Tensor`. Must be one of the following types: `float32`,
      `float64`, `int32`, `uint8`, `int16`, `int8`, `int64`, `bfloat16`,
      `uint16`, `half`, `uint32`, `uint64`. 4-D Tensor with shape
    `[batch, in_rows, in_cols, depth]`. `ksizes`: A list of `ints` that has
      length `>= 4`. The size of the sliding window for each
    dimension of `images`. `strides`: A list of `ints` that has length `>= 4`.
      1-D of length 4. How far the centers of two consecutive
    patches are in the images. Must be:
    `[1, stride_rows, stride_cols, 1]`. `rates`: A list of `ints`
    that has length `>= 4`. 1-D of length 4. Must be: `[1, rate_rows, rate_cols,
      1]`. This is the input stride, specifying how far two consecutive patch
      samples are in the input. Equivalent to extracting patches with
      `patch_sizes_eff = patch_sizes + (patch_sizes - 1) * (rates - 1)`,
      followed by subsampling them spatially by a factor of `rates`. This is
      equivalent to `rate` in dilated (a.k.a. Atrous) convolutions.
    `padding`: A `string` from: "SAME", "VALID". The type of padding algorithm
      to use.
    We specify the size-related attributes as:  ``` ksizes = [1, ksize_rows,
      ksize_cols, 1] strides = [1, strides_rows, strides_cols, 1] rates = [1,
      rates_rows, rates_cols, 1]
    name: A name for the operation (optional). ```

  Returns:
    A Tensor. Has the same type as images.
  """
  ksizes = deprecation.deprecated_argument_lookup("sizes", sizes, "ksizes",
                                                  ksizes)
  return gen_array_ops.extract_image_patches(images, ksizes, strides, rates,
                                             padding, name)


extract_image_patches.__doc__ = gen_array_ops.extract_image_patches.__doc__


@tf_export("fingerprint")
@dispatch.add_dispatch_support
def fingerprint(data, method="farmhash64", name=None):
  r"""Generates fingerprint values.

  Generates fingerprint values of `data`.

  Fingerprint op considers the first dimension of `data` as the batch dimension,
  and `output[i]` contains the fingerprint value generated from contents in
  `data[i, ...]` for all `i`.

  Fingerprint op writes fingerprint values as byte arrays. For example, the
  default method `farmhash64` generates a 64-bit fingerprint value at a time.
  This 8-byte value is written out as an `tf.uint8` array of size 8, in
  little-endian order.

  For example, suppose that `data` has data type `tf.int32` and shape (2, 3, 4),
  and that the fingerprint method is `farmhash64`. In this case, the output
  shape is (2, 8), where 2 is the batch dimension size of `data`, and 8 is the
  size of each fingerprint value in bytes. `output[0, :]` is generated from
  12 integers in `data[0, :, :]` and similarly `output[1, :]` is generated from
  other 12 integers in `data[1, :, :]`.

  Note that this op fingerprints the raw underlying buffer, and it does not
  fingerprint Tensor's metadata such as data type and/or shape. For example, the
  fingerprint values are invariant under reshapes and bitcasts as long as the
  batch dimension remain the same:

  ```python
  tf.fingerprint(data) == tf.fingerprint(tf.reshape(data, ...))
  tf.fingerprint(data) == tf.fingerprint(tf.bitcast(data, ...))
  ```

  For string data, one should expect `tf.fingerprint(data) !=
  tf.fingerprint(tf.string.reduce_join(data))` in general.

  Args:
    data: A `Tensor`. Must have rank 1 or higher.
    method: A `Tensor` of type `tf.string`. Fingerprint method used by this op.
      Currently, available method is `farmhash64`.
    name: A name for the operation (optional).

  Returns:
    A two-dimensional `Tensor` of type `tf.uint8`. The first dimension equals to
    `data`'s first dimension, and the second dimension size depends on the
    fingerprint algorithm.
  """
  return gen_array_ops.fingerprint(data, method, name)


def convert_to_int_tensor(tensor, name, dtype=dtypes.int32):
  """Converts the given value to an integer Tensor."""
  tensor = ops.convert_to_tensor(
      tensor, name=name, preferred_dtype=dtype or dtypes.int32)
  if tensor.dtype.is_integer:
    if dtype is not None:
      tensor = gen_math_ops.cast(tensor, dtype)
  else:
    raise TypeError(f"Argument `tensor` (name: {name}) must be of type integer."
                    f" Received `tensor` = {tensor} of dtype: {tensor.dtype}")
  return tensor


def get_positive_axis(axis, ndims, axis_name="axis", ndims_name="ndims"):
  """Validate an `axis` parameter, and normalize it to be positive.

  If `ndims` is known (i.e., not `None`), then check that `axis` is in the
  range `-ndims <= axis < ndims`, and return `axis` (if `axis >= 0`) or
  `axis + ndims` (otherwise).
  If `ndims` is not known, and `axis` is positive, then return it as-is.
  If `ndims` is not known, and `axis` is negative, then report an error.

  Args:
    axis: An integer constant
    ndims: An integer constant, or `None`
    axis_name: The name of `axis` (for error messages).
    ndims_name: The name of `ndims` (for error messages).

  Returns:
    The normalized `axis` value.

  Raises:
    ValueError: If `axis` is out-of-bounds, or if `axis` is negative and
      `ndims is None`.
  """
  if not isinstance(axis, int):
    raise TypeError(f"{axis_name} must be an int; got {type(axis).__name__}")
  if ndims is not None:
    if 0 <= axis < ndims:
      return axis
    elif -ndims <= axis < 0:
      return axis + ndims
    else:
      raise ValueError(f"{axis_name}={axis} out of bounds: "
                       f"expected {-ndims}<={axis_name}<{ndims}")
  elif axis < 0:
    raise ValueError(f"{axis_name}={axis} may only be negative "
                     f"if {ndims_name} is statically known.")
  return axis


# This op is intended to exactly match the semantics of numpy.repeat, with
# one exception: numpy.repeat has special (and somewhat non-intuitive) behavior
# when axis is not specified.  Rather than implement that special behavior, we
# simply make `axis` be a required argument.
#
# External (OSS) `tf.repeat` feature request:
# https://github.com/tensorflow/tensorflow/issues/8246
def repeat_with_axis(data, repeats, axis, name=None):
  """Repeats elements of `data`.

  Args:
    data: An `N`-dimensional tensor.
    repeats: A 1-D integer tensor specifying how many times each element in
      `axis` should be repeated.  `len(repeats)` must equal `data.shape[axis]`.
      Supports broadcasting from a scalar value.
    axis: `int`.  The axis along which to repeat values.  Must be less than
      `max(N, 1)`.
    name: A name for the operation.

  Returns:
    A tensor with `max(N, 1)` dimensions.  Has the same shape as `data`,
    except that dimension `axis` has size `sum(repeats)`.

  Example usage:

  >>> repeat(['a', 'b', 'c'], repeats=[3, 0, 2], axis=0)
  <tf.Tensor: shape=(5,), dtype=string,
  numpy=array([b'a', b'a', b'a', b'c', b'c'], dtype=object)>
  >>> repeat([[1, 2], [3, 4]], repeats=[2, 3], axis=0)
  <tf.Tensor: shape=(5, 2), dtype=int32, numpy=
  array([[1, 2],
         [1, 2],
         [3, 4],
         [3, 4],
         [3, 4]], dtype=int32)>
  >>> repeat([[1, 2], [3, 4]], repeats=[2, 3], axis=1)
  <tf.Tensor: shape=(2, 5), dtype=int32, numpy=
  array([[1, 1, 2, 2, 2],
         [3, 3, 4, 4, 4]], dtype=int32)>

  """
  # Whether the execution uses the optimized non-XLA implementation below.
  # TODO(b/236387200): Separate the implementations at a lower level, so that
  # non-XLA path gets the performance benefits and the XLA path is not broken
  # after loading a saved model with the optimization.
  use_optimized_non_xla_implementation = False

  if not isinstance(axis, int):
    raise TypeError("Argument `axis` must be an int. "
                    f"Received `axis` = {axis} of type {type(axis).__name__}")

  with ops.name_scope(name, "Repeat", [data, repeats]):
    data = ops.convert_to_tensor(data, name="data")
    # Note: We want to pass dtype=None to convert_to_int_tensor so that the
    # existing type is maintained instead of force-casting to int32. However,
    # this is not compatible with the implementation used on the XLA path.
    if not use_optimized_non_xla_implementation:
      repeats = convert_to_int_tensor(repeats, name="repeats")
    else:
      repeats = convert_to_int_tensor(repeats, name="repeats", dtype=None)

    repeats.shape.with_rank_at_most(1)

    # If `data` is a scalar, then upgrade it to a vector.
    data = _with_nonzero_rank(data)
    data_shape = shape(data, out_type=repeats.dtype)

    # If `axis` is negative, then convert it to a positive value.
    axis = get_positive_axis(axis, data.shape.rank, ndims_name="rank(data)")

    # If we know that `repeats` is a scalar, then we can just tile & reshape.
    if repeats.shape.num_elements() == 1:
      repeats = reshape(repeats, [])
      expanded = expand_dims(data, axis + 1)
      tiled = tile_one_dimension(expanded, axis + 1, repeats)
      result_shape = concat([
          data_shape[:axis], [repeats * data_shape[axis]], data_shape[axis + 1:]
      ],
                            axis=0)
      return reshape(tiled, result_shape)

    # Check data Tensor shapes.
    if repeats.shape.ndims == 1:
      data.shape.dims[axis].assert_is_compatible_with(repeats.shape[0])

    repeats = broadcast_to(repeats, [data_shape[axis]])

    # The implementation on the else branch has better performance. However, it
    # does not work on the XLA path since it relies on the range op with a
    # shape that is not a compile-time constant.
    if not use_optimized_non_xla_implementation:
      repeats_original = repeats

      # Broadcast the `repeats` tensor so rank(repeats) == axis + 1.
      if repeats.shape.ndims != axis + 1:
        repeats_shape = shape(repeats)
        repeats_ndims = rank(repeats)
        broadcast_shape = concat(
            [data_shape[:axis + 1 - repeats_ndims], repeats_shape], axis=0)
        repeats = broadcast_to(repeats, broadcast_shape)
        repeats.set_shape([None] * (axis + 1))

      # Create a "sequence mask" based on `repeats`, where slices across `axis`
      # contain one `True` value for each repetition.  E.g., if
      # `repeats = [3, 1, 2]`, then `mask = [[1, 1, 1], [1, 0, 0], [1, 1, 0]]`.
      max_repeat = gen_math_ops._max(repeats, _all_dimensions(repeats))
      max_repeat = gen_math_ops.maximum(
          ops.convert_to_tensor(0, name="zero", dtype=max_repeat.dtype),
          max_repeat)

      mask = sequence_mask(repeats, max_repeat)

      # Add a new dimension around each value that needs to be repeated, and
      # then tile that new dimension to match the maximum number of repetitions.
      expanded = expand_dims(data, axis + 1)
      tiled = tile_one_dimension(expanded, axis + 1, max_repeat)

      # Use `boolean_mask` to discard the extra repeated values.  This also
      # flattens all dimensions up through `axis`.
      masked = boolean_mask(tiled, mask)

      # Reshape the output tensor to add the outer dimensions back.
      if axis == 0:
        result = masked
      else:
        repeated_dim_size = gen_math_ops._sum(
            repeats_original,
            axis=gen_math_ops._range(0, rank(repeats_original), 1))
        result_shape = concat(
            [data_shape[:axis], [repeated_dim_size], data_shape[axis + 1:]],
            axis=0)
        result = reshape(masked, result_shape)

      # Preserve shape information.
      if data.shape.ndims is not None:
        new_axis_size = 0 if repeats.shape[0] == 0 else None
        result.set_shape(data.shape[:axis].concatenate(
            [new_axis_size]).concatenate(data.shape[axis + 1:]))

      return result

    else:
      # Non-XLA path implementation
      # E.g., repeats = [3, 4, 0, 2, 1].
      # E.g., repeats_scan = [3, 7, 7, 9, 10].
      repeats_scan = gen_math_ops.cumsum(repeats)
      # This concat just prepends 0 to handle the case when repeats are empty.
      # E.g., output_size = [0, 3, 7, 7, 9, 10][-1] = 10.
      output_size = concat([zeros(1, dtype=repeats_scan.dtype), repeats_scan],
                           axis=0)[-1]
      # E.g., output_indices = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9].
      output_indices = gen_math_ops.range(output_size, dtype=repeats.dtype)
      # E.g., gather_indices = [0, 0, 0, 1, 1, 1, 1, 3, 3, 4].
      gather_indices = searchsorted(
          repeats_scan, output_indices, side="right", out_type=repeats.dtype)
      return gather(data, gather_indices, axis=axis)


def tile_one_dimension(data, axis, multiple):
  """Tiles a single dimension of a tensor."""
  # Assumes axis is a nonnegative int.
  if data.shape.ndims is not None:
    multiples = [1] * data.shape.ndims
    multiples[axis] = multiple
  else:
    ones_value = ones(rank(data), dtypes.int32)
    multiples = concat([ones_value[:axis], [multiple], ones_value[axis + 1:]],
                       axis=0)
  return tile(data, multiples)


def _with_nonzero_rank(data):
  """If `data` is scalar, then add a dimension; otherwise return as-is."""
  if data.shape.ndims is not None:
    if data.shape.ndims == 0:
      return array_ops_stack.stack([data])
    else:
      return data
  else:
    data_shape = shape(data)
    data_ndims = rank(data)
    return reshape(data, concat([[1], data_shape], axis=0)[-data_ndims:])


@tf_export("repeat")
@dispatch.add_dispatch_support
def repeat(input, repeats, axis=None, name=None):  # pylint: disable=redefined-builtin
  """Repeat elements of `input`.

  See also `tf.concat`, `tf.stack`, `tf.tile`.

  Args:
    input: An `N`-dimensional Tensor.
    repeats: An 1-D `int` Tensor. The number of repetitions for each element.
      repeats is broadcasted to fit the shape of the given axis. `len(repeats)`
      must equal `input.shape[axis]` if axis is not None.
    axis: An int. The axis along which to repeat values. By default, (axis=None),
      use the flattened input array, and return a flat output array.
    name: A name for the operation.

  Returns:
    A Tensor which has the same shape as `input`, except along the given axis.
      If axis is None then the output array is flattened to match the flattened
      input array.

  Example usage:

  >>> repeat(['a', 'b', 'c'], repeats=[3, 0, 2], axis=0)
  <tf.Tensor: shape=(5,), dtype=string,
  numpy=array([b'a', b'a', b'a', b'c', b'c'], dtype=object)>

  >>> repeat([[1, 2], [3, 4]], repeats=[2, 3], axis=0)
  <tf.Tensor: shape=(5, 2), dtype=int32, numpy=
  array([[1, 2],
         [1, 2],
         [3, 4],
         [3, 4],
         [3, 4]], dtype=int32)>

  >>> repeat([[1, 2], [3, 4]], repeats=[2, 3], axis=1)
  <tf.Tensor: shape=(2, 5), dtype=int32, numpy=
  array([[1, 1, 2, 2, 2],
         [3, 3, 4, 4, 4]], dtype=int32)>

  >>> repeat(3, repeats=4)
  <tf.Tensor: shape=(4,), dtype=int32, numpy=array([3, 3, 3, 3], dtype=int32)>

  >>> repeat([[1,2], [3,4]], repeats=2)
  <tf.Tensor: shape=(8,), dtype=int32,
  numpy=array([1, 1, 2, 2, 3, 3, 4, 4], dtype=int32)>

  """
  if axis is None:
    input = reshape(input, [-1])
    axis = 0
  return repeat_with_axis(input, repeats, axis, name)


@tf_export("guarantee_const")
@deprecation.deprecated(None, "Not for public use.")
def guarantee_const(input, name=None):    # pylint: disable=redefined-builtin
  """Promise to the TF runtime that the input tensor is a constant.

  The runtime is then free to make optimizations based on this.

  Returns the input tensor without modification.

  Args:
    input: A `Tensor`.
    name: A name for this operation.

  Returns:
    A `Tensor`. Has the same dtype as `input`.
  """
  return gen_array_ops.guarantee_const(input=input, name=name)


@tf_export("stop_gradient")
@dispatch.add_dispatch_support
def stop_gradient(input, name=None):  # pylint: disable=redefined-builtin
  """Stops gradient computation.

  NOTE: This docstring is patched out below. See
  tensorflow/core/api_def/base_api/api_def_StopGradient.pbtxt for the full
  docstring. That file determines the public documentation page.

  Args:
    input: A `Tensor`.
    name: A name for this operation.

  Returns:
    A `Tensor`. Has the same dtype as `input`.
  """
  # Don't expand ResourceVariables, so stop_gradient(variable) will return a
  # Tensor.
  if (isinstance(input, composite_tensor.CompositeTensor) and
      not _pywrap_utils.IsResourceVariable(input)):
    return nest.map_structure(stop_gradient, input, expand_composites=True)
  # The StopGradient op has a gradient function registered which returns None
  # (meaning statically known to be zero). For correctness, that's all we
  # need. However, tf.GradientTape often makes decisions about what to keep in
  # memory based on which forward-pass tensors are currently being watched, and
  # returning None in a gradient is not sufficient to stop watching a tensor
  # since the backward function doesn't run in the forward pass. Pausing the
  # tape around this op instructs any tf.GradientTapes to ignore the
  # forward-pass output of StopGradient, which may be much more efficient.
  with record.stop_recording():
    return gen_array_ops.stop_gradient(input, name=name)


stop_gradient.__doc__ = gen_array_ops.stop_gradient.__doc__


# Register elementwise ops that don't have Python wrappers.
dispatch.register_unary_elementwise_api(gen_array_ops.check_numerics)