def BuildLoop(self, pred, body, loop_vars):
    """Add the loop termination condition and body to the graph."""

    loop_vars = ops.convert_n_to_tensor_or_indexed_slices(loop_vars)
    # Let the context know the loop variabes so the _Enter nodes below
    # would be added into the context correctly.
    self._values = set([x.name for x in loop_vars])
    if self._outer_context is not None:
      real_vars = [self._outer_context.AddValue(x) for x in loop_vars]
    else:
      real_vars = loop_vars
    enter_vars = [_Enter(x, self._name, is_constant=False,
                         parallel_iterations=self._parallel_iterations)
                  for x in real_vars]
    self._values = set([x.name for x in enter_vars])

    merge_vars = [merge([x, x])[0] for x in enter_vars]
    self._pivot_for_pred = merge_vars[0]

    # Build the graph for pred.
    c = ops.convert_to_tensor(pred(*merge_vars))
    self._pivot = loop_cond(c, name="LoopCond")
    switch_vars = [_SwitchRefOrTensor(x, self._pivot) for x in merge_vars]

    # Build the graph for body.
    vars_for_body = [_Identity(x[1]) for x in switch_vars]
    self._pivot_for_body = vars_for_body[0]

    body_result = body(*vars_for_body)
    if not isinstance(body_result, (list, _basetuple)):
      body_result = [body_result]
    result = ops.convert_n_to_tensor_or_indexed_slices(body_result)
    next_vars = [next_iteration(x) for x in result]

    # Add the back edges to complete the loop.
    assert len(merge_vars) == len(next_vars)
    for x in zip(merge_vars, next_vars):
      x[0].op._update_input(1, x[1])

    # Add the exit ops.
    exit_vars = [exit(x[0]) for x in switch_vars]

    for m_var, n_var, e_var in zip(merge_vars, next_vars, exit_vars):
      if m_var.get_shape().is_compatible_with(n_var.get_shape()):
        e_var.set_shape(m_var.get_shape().merge_with(n_var.get_shape()))

    # Exit the loop.
    self.ExitResult(exit_vars)
    self.Exit()
    return exit_vars[0] if len(exit_vars) == 1 else exit_vars

def _DefaultGradYs(grad_ys, ys, colocate_gradients_with_ops):
  """Fill in default values for grad_ys.

  Args:
    grad_ys: List of gradients, can contain None.
    ys: List of tensors.
    colocate_gradients_with_ops: If True, try colocating gradients with
      the corresponding op.

  Returns:
    A list of gradients to use, without None.

  Raises:
    ValueError: If one of the grad_ys is invalid.
  """
  if len(grad_ys) != len(ys):
    raise ValueError("Passed %d grad_ys for %d ys" % (len(grad_ys), len(ys)))
  grad_ys = ops.convert_n_to_tensor_or_indexed_slices(grad_ys, name="grad_y")
  for i in xrange(len(grad_ys)):
    grad_y = grad_ys[i]
    y = ys[i]
    if grad_y is None:
      with _maybe_colocate_with(y.op, colocate_gradients_with_ops):
        grad_ys[i] = array_ops.fill(
            array_ops.shape(y), constant_op.constant(
                1, dtype=y.dtype))
    else:
      if grad_y.dtype != y.dtype:
        raise ValueError("Y and ys_grad must be of the same type, "
                         "not y: %s, ys_grad: %s " %
                         (dtypes.as_dtype(y.dtype).name,
                          dtypes.as_dtype(grad_y.dtype).name))
  return grad_ys

def slice_input_producer(tensor_list, num_epochs=None, shuffle=True, seed=None, capacity=32, name=None):
    """Produces a slice of each `Tensor` in `tensor_list`.

  Implemented using a Queue -- a `QueueRunner` for the Queue
  is added to the current `Graph`'s `QUEUE_RUNNER` collection.

  Args:
    tensor_list: A list of `Tensor` objects. Every `Tensor` in
      `tensor_list` must have the same size in the first dimension.
    num_epochs: An integer (optional). If specified, `slice_input_producer`
      produces each slice `num_epochs` times before generating
      an `OutOfRange` error. If not specified, `slice_input_producer` can cycle
      through the slices an unlimited number of times.
    seed: An integer (optional). Seed used if shuffle == True.
    capacity: An integer. Sets the queue capacity.
    name: A name for the operations (optional).

  Returns:
    A list of tensors, one for each element of `tensor_list`.  If the tensor
    in `tensor_list` has shape `[N, a, b, .., z]`, then the corresponding output
    tensor will have shape `[a, b, ..., z]`.
  """
    with ops.op_scope(tensor_list, name, "input_producer"):
        tensor_list = ops.convert_n_to_tensor_or_indexed_slices(tensor_list)
        if not tensor_list:
            raise ValueError("Expected at least one tensor in slice_input_producer().")
        range_size = array_ops.shape(tensor_list[0])[0]
        # TODO(josh11b): Add an assertion that the first dimension of
        # everything in TensorList matches. Maybe just check the inferred shapes?
        queue = range_input_producer(range_size, num_epochs=num_epochs, shuffle=shuffle, seed=seed, capacity=capacity)
        index = queue.dequeue()
        output = [array_ops.gather(t, index) for t in tensor_list]
        return output

def _DefaultGradYs(grad_ys, ys, colocate_gradients_with_ops):
  """Fill in default values for grad_ys.

  Args:
    grad_ys: List of gradients, can contain None.
    ys: List of tensors.
    colocate_gradients_with_ops: If True, try colocating gradients with
      the corresponding op.

  Returns:
    A list of gradients to use, without None.

  Raises:
    ValueError: If sizes of gradients and inputs don't match
    TypeError: If type of any gradient is not valid for its input.
  """
  if len(grad_ys) != len(ys):
    raise ValueError("Passed %d grad_ys for %d ys" % (len(grad_ys), len(ys)))
  grad_ys = ops.convert_n_to_tensor_or_indexed_slices(grad_ys, name="grad_y")
  for i in xrange(len(grad_ys)):
    grad_y = grad_ys[i]
    y = ys[i]
    if grad_y is None:
      if y.dtype.is_complex:
        raise TypeError(
            "Gradients of complex tensors must set grad_ys (y.dtype = %r)" %
            y.dtype)
      with _maybe_colocate_with(y.op, colocate_gradients_with_ops):
        grad_ys[i] = array_ops.fill(
            array_ops.shape(y), constant_op.constant(
                1, dtype=y.dtype))
      continue
    if y.dtype.is_floating or y.dtype.is_integer:
      if not grad_y.dtype.is_floating and not grad_y.dtype.is_integer:
        raise TypeError("Gradient type %s generated for real or "
                         "integer-valued tensor %s with type %s must be "
                         "real or integer" %
                         (dtypes.as_dtype(grad_y.dtype).name, y,
                          dtypes.as_dtype(y.dtype).name))
    elif y.dtype.is_complex:
      if not grad_y.dtype.is_complex:
        raise TypeError("Gradient type %s generated for complex-valued "
                         "tensor %s with type %s must be real" %
                         (dtypes.as_dtype(grad_y.dtype).name, y,
                          dtypes.as_dtype(y.dtype).name))
    else:
      raise TypeError("Tensor %s with type %s must be numeric "
                      "to obtain a default gradient" %
                      (y, dtypes.as_dtype(y.dtype).name))
  return grad_ys

def rejection_sample(tensors, accept_prob_fn, batch_size, queue_threads=1,
                     enqueue_many=False, prebatch_capacity=16,
                     prebatch_threads=1, runtime_checks=False, name=None):
  """Stochastically creates batches by rejection sampling.

  Each list of non-batched tensors is evaluated by `accept_prob_fn`, to produce
  a scalar tensor between 0 and 1. This tensor corresponds to the probability of
  being accepted. When `batch_size` tensor groups have been accepted, the batch
  queue will return a mini-batch.

  Args:
    tensors: List of tensors for data. All tensors are either one item or a
        batch, according to enqueue_many.
    accept_prob_fn: A python lambda that takes a non-batch tensor from each
        item in `tensors`, and produces a scalar tensor.
    batch_size: Size of batch to be returned.
    queue_threads: The number of threads for the queue that will hold the final
      batch.
    enqueue_many: Bool. If true, interpret input tensors as having a batch
        dimension.
    prebatch_capacity: Capacity for the large queue that is used to convert
      batched tensors to single examples.
    prebatch_threads: Number of threads for the large queue that is used to
      convert batched tensors to single examples.
    runtime_checks: Bool. If true, insert runtime checks on the output of
        `accept_prob_fn`. Using `True` might have a performance impact.
    name: Optional prefix for ops created by this function.
  Raises:
    ValueError: enqueue_many is True and labels doesn't have a batch
        dimension, or if enqueue_many is False and labels isn't a scalar.
    ValueError: enqueue_many is True, and batch dimension on data and labels
        don't match.
    ValueError: if a zero initial probability class has a nonzero target
        probability.
  Returns:
    A list of tensors of the same length as `tensors`, with batch dimension
    `batch_size`.

  Example:
    # Get tensor for a single data and label example.
    data, label = data_provider.Get(['data', 'label'])

    # Get stratified batch according to data tensor.
    accept_prob_fn = lambda x: (tf.tanh(x[0]) + 1) / 2
    data_batch = tf.contrib.training.rejection_sample(
        [data, label], accept_prob_fn, 16)

    # Run batch through network.
    ...
  """
  with variable_scope.variable_scope(name, 'rejection_sample', tensors):
    tensor_list = ops.convert_n_to_tensor_or_indexed_slices(tensors)
    # Reduce the case of a batched example to that of a batch of a single
    # example by taking a batch of size one.
    if enqueue_many:
      # Validate that batch dimension of the input is consistent.
      tensor_list = _verify_data_inputs(tensor_list)

      # Make a single queue to hold input examples. Reshape output so examples
      # don't have singleton batch dimension.
      batched = input_ops.batch(tensor_list,
                                batch_size=1,
                                num_threads=prebatch_threads,
                                capacity=prebatch_capacity,
                                enqueue_many=True)
      tensor_list = [array_ops.squeeze(x, [0]) for x in batched]

    # Set up a queue containing batches that have the distribution.
    cur_prob = accept_prob_fn(tensor_list)
    if runtime_checks:
      cur_prob = array_ops.identity(control_flow_ops.with_dependencies(
          [check_ops.assert_less_equal(0.0, cur_prob),
           check_ops.assert_less_equal(cur_prob, 1.0)],
          cur_prob), name='prob_with_checks')
    keep_input = random_ops.random_uniform([]) < cur_prob
    return _conditional_batch(
        tensor_list, keep_input, batch_size, num_threads=queue_threads)

def stratified_sample_unknown_dist(tensors, labels, probs, batch_size,
                                   enqueue_many=False, queue_capacity=16,
                                   threads_per_queue=1, name=None):
  """Stochastically creates batches based on per-class probabilities.

  **NOTICE** This sampler can be significantly slower than `stratified_sample`
  due to each thread discarding all examples not in its assigned class.

  This uses a number of threads proportional to the number of classes. See
  `stratified_sample` for an implementation that discards fewer examples and
  uses a fixed number of threads. This function's only advantage over
  `stratified_sample` is that the class data-distribution doesn't need to be
  known ahead of time.

  Args:
    tensors: List of tensors for data. All tensors are either one item or a
        batch, according to enqueue_many.
    labels: Tensor for label of data. Label is a single integer or a batch,
        depending on enqueue_many. It is not a one-hot vector.
    probs: Target class probabilities. An object whose type has a registered
        Tensor conversion function.
    batch_size: Size of batch to be returned.
    enqueue_many: Bool. If true, interpret input tensors as having a batch
        dimension.
    queue_capacity: Capacity of each per-class queue.
    threads_per_queue: Number of threads for each per-class queue.
    name: Optional prefix for ops created by this function.
  Raises:
    ValueError: enqueue_many is True and labels doesn't have a batch
        dimension, or if enqueue_many is False and labels isn't a scalar.
    ValueError: enqueue_many is True, and batch dimension of data and labels
        don't match.
    ValueError: if probs don't sum to one.
    TFAssertion: if labels aren't integers in [0, num classes).
  Returns:
    (data_batch, label_batch), where data_batch is a list of tensors of the same
        length as `tensors`

  Example:
    # Get tensor for a single data and label example.
    data, label = data_provider.Get(['data', 'label'])

    # Get stratified batch according to per-class probabilities.
    init_probs = [1.0/NUM_CLASSES for _ in range(NUM_CLASSES)]
    [data_batch], labels = (
        tf.contrib.training.stratified_sample_unknown_dist(
            [data], label, init_probs, 16))

    # Run batch through network.
    ...
  """
  with ops.name_scope(name, 'stratified_sample_unknown_dist',
                      tensors + [labels]):
    tensor_list = ops.convert_n_to_tensor_or_indexed_slices(tensors)
    labels = ops.convert_to_tensor(labels)
    probs = ops.convert_to_tensor(probs, dtype=dtypes.float32)
    # Reduce the case of a single example to that of a batch of size 1.
    if not enqueue_many:
      tensor_list = [array_ops.expand_dims(tensor, 0) for tensor in tensor_list]
      labels = array_ops.expand_dims(labels, 0)

    # Validate that input is consistent.
    tensor_list, labels, [probs] = _verify_input(tensor_list, labels, [probs])

    # Make per-class queues.
    per_class_queues = _make_per_class_queues(
        tensor_list, labels, probs.get_shape().num_elements(), queue_capacity,
        threads_per_queue)

    # Use the per-class queues to generate stratified batches.
    return _get_batch_from_per_class_queues(
        per_class_queues, probs, batch_size)

def stratified_sample(tensors, labels, target_probs, batch_size,
                      init_probs=None, enqueue_many=False, queue_capacity=16,
                      threads_per_queue=1, name=None):
  """Stochastically creates batches based on per-class probabilities.

  This method discards examples. Internally, it creates one queue to amortize
  the cost of disk reads, and one queue to hold the properly-proportioned
  batch. See `stratified_sample_unknown_dist` for a function that performs
  stratified sampling with one queue per class and doesn't require knowing the
  class data-distribution ahead of time.

  Args:
    tensors: List of tensors for data. All tensors are either one item or a
        batch, according to enqueue_many.
    labels: Tensor for label of data. Label is a single integer or a batch,
        depending on enqueue_many. It is not a one-hot vector.
    target_probs: Target class proportions in batch. An object whose type has a
        registered Tensor conversion function.
    batch_size: Size of batch to be returned.
    init_probs: Class proportions in the data. An object whose type has a
        registered Tensor conversion function, or `None` for estimating the
        initial distribution.
    enqueue_many: Bool. If true, interpret input tensors as having a batch
        dimension.
    queue_capacity: Capacity of the large queue that holds input examples.
    threads_per_queue: Number of threads for the large queue that holds input
        examples and for the final queue with the proper class proportions.
    name: Optional prefix for ops created by this function.
  Raises:
    ValueError: enqueue_many is True and labels doesn't have a batch
        dimension, or if enqueue_many is False and labels isn't a scalar.
    ValueError: enqueue_many is True, and batch dimension on data and labels
        don't match.
    ValueError: if probs don't sum to one.
    ValueError: if a zero initial probability class has a nonzero target
        probability.
    TFAssertion: if labels aren't integers in [0, num classes).
  Returns:
    (data_batch, label_batch), where data_batch is a list of tensors of the same
        length as `tensors`

  Example:
    # Get tensor for a single data and label example.
    data, label = data_provider.Get(['data', 'label'])

    # Get stratified batch according to per-class probabilities.
    target_probs = [...distribution you want...]
    [data_batch], labels = tf.contrib.training.stratified_sample(
        [data], label, target_probs)

    # Run batch through network.
    ...
  """
  with ops.name_scope(name, 'stratified_sample', tensors + [labels]):
    tensor_list = ops.convert_n_to_tensor_or_indexed_slices(tensors)
    labels = ops.convert_to_tensor(labels)
    target_probs = ops.convert_to_tensor(target_probs, dtype=dtypes.float32)
    # Reduce the case of a single example to that of a batch of size 1.
    if not enqueue_many:
      tensor_list = [array_ops.expand_dims(tensor, 0) for tensor in tensor_list]
      labels = array_ops.expand_dims(labels, 0)

    # If `init_probs` is `None`, set up online estimation of data distribution.
    if init_probs is None:
      # We use `target_probs` to get the number of classes, so its shape must be
      # fully defined at graph construction time.
      target_probs.get_shape().assert_is_fully_defined()
      init_probs = _estimate_data_distribution(
          labels, target_probs.get_shape().num_elements())
    else:
      init_probs = ops.convert_to_tensor(init_probs, dtype=dtypes.float32)

    # Validate that input is consistent.
    tensor_list, labels, [init_probs, target_probs] = _verify_input(
        tensor_list, labels, [init_probs, target_probs])

    # Check that all zero initial probabilities also have zero target
    # probabilities.
    assert_op = control_flow_ops.Assert(
        math_ops.reduce_all(math_ops.logical_or(
            math_ops.not_equal(init_probs, 0),
            math_ops.equal(target_probs, 0))),
        ['All classes with zero initial probability must also have zero target '
         'probability: ', init_probs, target_probs])
    init_probs = control_flow_ops.with_dependencies([assert_op], init_probs)

    # Calculate acceptance sampling probabilities.
    accept_probs = _calculate_acceptance_probabilities(init_probs, target_probs)
    proportion_rejected = math_ops.reduce_sum((1 - accept_probs) * init_probs)
    accept_probs = control_flow_ops.cond(
        math_ops.less(proportion_rejected, .5),
        lambda: accept_probs,
        lambda: logging_ops.Print(  # pylint: disable=g-long-lambda
            accept_probs, [accept_probs],
            message='Proportion of examples rejected by sampler is high.',
            first_n=10))

    # Make a single queue to hold input examples. Reshape output so examples
    # don't have singleton batch dimension.
    batched = input_ops.batch(tensor_list + [labels],
#.........这里部分代码省略.........

def _GradientsHelper(ys, xs, grad_ys, name, colocate_gradients_with_ops,
                     gate_gradients, aggregation_method, stop_gradients):
  """Implementation of gradients()."""
  if context.executing_eagerly():
    raise RuntimeError("tf.gradients not supported when eager execution "
                       "is enabled. Use tf.contrib.eager.GradientTape "
                       "instead.")
  ys = _AsList(ys)
  xs = _AsList(xs)
  stop_gradients = [] if stop_gradients is None else _AsList(stop_gradients)
  if grad_ys is None:
    grad_ys = [None] * len(ys)
  else:
    grad_ys = _AsList(grad_ys)

  with ops.name_scope(
      name, "gradients",
      list(ys) + list(xs) + list(stop_gradients) + list(grad_ys)) as grad_scope:
    ys = ops.convert_n_to_tensor_or_indexed_slices(ys, name="y")
    xs = [
        x.handle if resource_variable_ops.is_resource_variable(x) else x
        for x in xs
    ]
    xs = ops.internal_convert_n_to_tensor_or_indexed_slices(
        xs, name="x", as_ref=True)
    grad_ys = _DefaultGradYs(grad_ys, ys, colocate_gradients_with_ops)

    # The approach we take here is as follows: Create a list of all ops in the
    # subgraph between the ys and xs.  Visit these ops in reverse order of ids
    # to ensure that when we visit an op the gradients w.r.t its outputs have
    # been collected.  Then aggregate these gradients if needed, call the op's
    # gradient function, and add the generated gradients to the gradients for
    # its input.

    # Initialize the pending count for ops in the connected subgraph from ys
    # to the xs.
    if len(ys) > 1:
      ys = [array_ops.identity(y) if y.consumers() else y for y in ys]
    to_ops = [t.op for t in ys]
    from_ops = [t.op for t in xs]
    stop_gradient_ops = [t.op for t in stop_gradients]
    pending_count, loop_state = _PendingCount(
        ops.get_default_graph(), to_ops, from_ops, colocate_gradients_with_ops)

    # Iterate over the collected ops.
    #
    # grads: op => list of gradients received on each output endpoint of the
    # op.  The gradients for each endpoint are initially collected as a list.
    # When it is time to call the op's gradient function, for each endpoint we
    # aggregate the list of received gradients into a Add() Operation if there
    # is more than one.
    grads = {}

    # Add the initial gradients for the ys.
    for y, grad_y in zip(ys, grad_ys):
      _SetGrad(grads, y, grad_y)

    # Initialize queue with to_ops.
    queue = collections.deque()
    # Add the ops in 'to_ops' into the queue.
    to_ops_set = set()
    for op in to_ops:
      # 'ready' handles the case where one output gradient relies on
      # another output's gradient.
      # pylint: disable=protected-access
      ready = (pending_count[op._id] == 0)
      if ready and op._id not in to_ops_set:
        to_ops_set.add(op._id)
        queue.append(op)
      # pylint: enable=protected-access

    if loop_state:
      loop_exits = loop_state.ProcessUnusedLoopExits(pending_count, to_ops_set)
      for y in loop_exits:
        if _IsTrainable(y):
          _SetGrad(grads, y, loop_state.ZerosLikeForExit(y))
          queue.append(y.op)

    stop_ops = _StopOps(from_ops, stop_gradient_ops, pending_count)
    while queue:
      # generate gradient subgraph for op.
      op = queue.popleft()
      with _maybe_colocate_with(op, colocate_gradients_with_ops):
        if loop_state:
          loop_state.EnterGradWhileContext(op, before=True)
        out_grads = _AggregatedGrads(grads, op, loop_state, aggregation_method)
        if loop_state:
          loop_state.ExitGradWhileContext(op, before=True)

        grad_fn = None
        # pylint: disable=protected-access
        func_call = None
        is_func_call = ops.get_default_graph()._is_function(op.type)
        has_out_grads = any(isinstance(g, ops.Tensor) or g for g in out_grads)
        if has_out_grads and (op._id not in stop_ops):
          if is_func_call:
            func_call = ops.get_default_graph()._get_function(op.type)
            grad_fn = func_call.python_grad_func
            # pylint: enable=protected-access
          else:
#.........这里部分代码省略.........

def _validate(tensor_list):
  tensor_list = ops.convert_n_to_tensor_or_indexed_slices(tensor_list)
  if not tensor_list:
    raise ValueError("Expected at least one tensor in batch().")
  return tensor_list

def accumulate_n(inputs, shape=None, tensor_dtype=None, name=None):
  """Returns the element-wise sum of a list of tensors.

  Optionally, pass `shape` and `tensor_dtype` for shape and type checking,
  otherwise, these are inferred.

  For example:

  ```python
  # tensor 'a' is [[1, 2], [3, 4]]
  # tensor `b` is [[5, 0], [0, 6]]
  tf.accumulate_n([a, b, a]) ==> [[7, 4], [6, 14]]

  # Explicitly pass shape and type
  tf.accumulate_n([a, b, a], shape=[2, 2], tensor_dtype=tf.int32)
    ==> [[7, 4], [6, 14]]
  ```

  Args:
    inputs: A list of `Tensor` objects, each with same shape and type.
    shape: Shape of elements of `inputs`.
    tensor_dtype: The type of `inputs`.
    name: A name for the operation (optional).

  Returns:
    A `Tensor` of same shape and type as the elements of `inputs`.

  Raises:
    ValueError: If `inputs` don't all have same shape and dtype or the shape
    cannot be inferred.
  """
  if tensor_dtype is None:
    if not inputs or not isinstance(inputs, (list, tuple)):
      raise ValueError("inputs must be a list of at least one Tensor with the "
                       "same dtype and shape")
    inputs = ops.convert_n_to_tensor_or_indexed_slices(inputs)
    if not all(isinstance(x, ops.Tensor) for x in inputs):
      raise ValueError("inputs must be a list of at least one Tensor with the "
                       "same dtype and shape")
    if not all(x.dtype == inputs[0].dtype for x in inputs):
      raise ValueError("inputs must be a list of at least one Tensor with the "
                       "same dtype and shape")
    tensor_dtype = inputs[0].dtype
  if shape is not None:
    shape = tensor_shape.as_shape(shape)
  else:
    shape = tensor_shape.unknown_shape()
    for input_tensor in inputs:
      if isinstance(input_tensor, ops.Tensor):
        shape = shape.merge_with(input_tensor.get_shape())
  if not shape.is_fully_defined():
    # TODO(pbar): Make a version of assign_add that accepts an uninitialized
    # lvalue, and takes its shape from that? This would allow accumulate_n to
    # work in all situations that add_n currently works.
    raise ValueError("Cannot infer the shape of the accumulator for "
                     "accumulate_n. Pass the shape argument, or set the shape "
                     "of at least one of the inputs.")
  with ops.op_scope(inputs, name, "AccumulateN") as name:
    var = gen_state_ops._temporary_variable(shape=shape, dtype=tensor_dtype)
    var_name = var.op.name
    var = state_ops.assign(var, array_ops.zeros_like(inputs[0]))
    update_ops = []
    for input_tensor in inputs:
      op = state_ops.assign_add(var, input_tensor, use_locking=True)
      update_ops.append(op)
    with ops.control_dependencies(update_ops):
      return gen_state_ops._destroy_temporary_variable(var,
                                                       var_name=var_name,
                                                       name=name)

def gradients(ys,
              xs,
              grad_ys=None,
              name="gradients",
              colocate_gradients_with_ops=False,
              gate_gradients=False,
              aggregation_method=None,
              stop_gradients=None):
  """Constructs symbolic derivatives of sum of `ys` w.r.t. x in `xs`.

  `ys` and `xs` are each a `Tensor` or a list of tensors.  `grad_ys`
  is a list of `Tensor`, holding the gradients received by the
  `ys`. The list must be the same length as `ys`.

  `gradients()` adds ops to the graph to output the derivatives of `ys` with
  respect to `xs`.  It returns a list of `Tensor` of length `len(xs)` where
  each tensor is the `sum(dy/dx)` for y in `ys`.

  `grad_ys` is a list of tensors of the same length as `ys` that holds
  the initial gradients for each y in `ys`.  When `grad_ys` is None,
  we fill in a tensor of '1's of the shape of y for each y in `ys`.  A
  user can provide their own initial `grad_ys` to compute the
  derivatives using a different initial gradient for each y (e.g., if
  one wanted to weight the gradient differently for each value in
  each y).

  `stop_gradients` is a `Tensor` or a list of tensors to be considered constant
  with respect to all `xs`. These tensors will not be backpropagated through,
  as though they had been explicitly disconnected using `stop_gradient`.  Among
  other things, this allows computation of partial derivatives as opposed to
  total derivatives. For example:

  ```python
  a = tf.constant(0.)
  b = 2 * a
  g = tf.gradients(a + b, [a, b], stop_gradients=[a, b])
  ```

  Here the partial derivatives `g` evaluate to `[1.0, 1.0]`, compared to the
  total derivatives `tf.gradients(a + b, [a, b])`, which take into account the
  influence of `a` on `b` and evaluate to `[3.0, 1.0]`.  Note that the above is
  equivalent to:

  ```python
  a = tf.stop_gradient(tf.constant(0.))
  b = tf.stop_gradient(2 * a)
  g = tf.gradients(a + b, [a, b])
  ```

  `stop_gradients` provides a way of stopping gradient after the graph has
  already been constructed, as compared to `tf.stop_gradient` which is used
  during graph construction.  When the two approaches are combined,
  backpropagation stops at both `tf.stop_gradient` nodes and nodes in
  `stop_gradients`, whichever is encountered first.

  Args:
    ys: A `Tensor` or list of tensors to be differentiated.
    xs: A `Tensor` or list of tensors to be used for differentiation.
    grad_ys: Optional. A `Tensor` or list of tensors the same size as
      `ys` and holding the gradients computed for each y in `ys`.
    name: Optional name to use for grouping all the gradient ops together.
      defaults to 'gradients'.
    colocate_gradients_with_ops: If True, try colocating gradients with
      the corresponding op.
    gate_gradients: If True, add a tuple around the gradients returned
      for an operations.  This avoids some race conditions.
    aggregation_method: Specifies the method used to combine gradient terms.
      Accepted values are constants defined in the class `AggregationMethod`.
    stop_gradients: Optional. A `Tensor` or list of tensors not to differentiate
      through.

  Returns:
    A list of `sum(dy/dx)` for each x in `xs`.

  Raises:
    LookupError: if one of the operations between `x` and `y` does not
      have a registered gradient function.
    ValueError: if the arguments are invalid.
    RuntimeError: if called in Eager mode.

  """
  if context.in_eager_mode():
    raise RuntimeError("tf.gradients not supported in EAGER mode. Use "
                       "functions in tf.contrib.eager.backprop instead.")
  ys = _AsList(ys)
  xs = _AsList(xs)
  stop_gradients = [] if stop_gradients is None else _AsList(stop_gradients)
  if grad_ys is None:
    grad_ys = [None] * len(ys)
  else:
    grad_ys = _AsList(grad_ys)

  with ops.name_scope(
      name, "gradients",
      list(ys) + list(xs) + list(stop_gradients) + list(grad_ys)) as grad_scope:
    ys = ops.convert_n_to_tensor_or_indexed_slices(ys, name="y")
    xs = [
        x.handle if isinstance(x, resource_variable_ops.ResourceVariable) else x
        for x in xs
    ]
#.........这里部分代码省略.........

def embedding_lookup(params, ids, name=None):
  """Looks up `ids` in a list of embedding tensors.

  This function is used to perform parallel lookups on the list of
  tensors in `params`.  It is a generalization of
  [`tf.gather()`](../../api_docs/python/array_ops.md#gather), where `params` is
  interpreted as a partition of a larger embedding tensor.

  If `len(params) > 1`, each element `id` of `ids` is partitioned between
  the elements of `params` by computing `p = id % len(params)`, and is
  then used to look up the slice `params[p][id // len(params), ...]`.

  The results of the lookup are then concatenated into a dense
  tensor. The returned tensor has shape `shape(ids) + shape(params)[1:]`.

  Args:
    params: A list of tensors with the same shape and type.
    ids: A `Tensor` with type `int32` containing the ids to be looked
      up in `params`.
    name: A name for the operation (optional).

  Returns:
    A `Tensor` with the same type as the tensors in `params`.

  Raises:
    ValueError: If `params` is empty.
  """
  if not isinstance(params, list):
    params = [params]
  with ops.op_scope(params + [ids], name, "embedding_lookup") as name:
    if not params:
      raise ValueError("Need at least one param")
    np = len(params)  # Number of partitions
    params = ops.convert_n_to_tensor_or_indexed_slices(params, name="params")
    if np == 1:
      with ops.device(params[0].device):
        return array_ops.gather(params[0], ids, name=name)
    else:
      ids = ops.convert_to_tensor(ids, name="ids")
      flat_ids = array_ops.reshape(ids, [-1])
      original_indices = math_ops.range(0, array_ops.size(flat_ids))
      # Compute flat_ids % partitions for each id
      ids_mod_p = flat_ids % np
      if ids_mod_p.dtype != types.int32:
        ids_mod_p = math_ops.cast(ids_mod_p, types.int32)
      # Partition single list of ids based on ids % np into np separate lists
      plist = data_flow_ops.dynamic_partition(flat_ids, ids_mod_p, np)
      # Similarly, partition the original indices.
      pindices = data_flow_ops.dynamic_partition(original_indices, ids_mod_p,
                                                 np)
      # Do np separate lookups, finding embeddings for plist[p] in params[p]
      partitioned_result = []
      for p in xrange(np):
        # TODO(agarwal): handle device allocations here and later in the
        # colocate code.
        gather_ids = plist[p] // np
        with ops.device(params[p].device):
          partitioned_result.append(array_ops.gather(params[p], gather_ids))
      # Stitch these back together
      ret = data_flow_ops.dynamic_stitch(pindices, partitioned_result,
                                         name=name)
      # Reshape to reverse the flattening of ids.
      # It's important that we compute params[0].shape on the right device
      # to avoid data motion.
      with ops.device(params[0].device):
        params_shape = array_ops.shape(params[0])
      ret = array_ops.reshape(ret, array_ops.concat(0, [
          array_ops.shape(ids), array_ops.slice(params_shape, [1], [-1])]))
      # output shape = ids.shape + params[*].shape[1:]
      # Normally the reshape is sufficient, but setting shape explicitly
      # teaches shape inference that params[1:].get_shape() matters.
      element_shape = params[0].get_shape()[1:]
      for p in params[1:]:
        element_shape = element_shape.merge_with(p.get_shape()[1:])
      ret.set_shape(ids.get_shape().concatenate(element_shape))
      return ret

def _embedding_lookup_and_transform(params,
                                    ids,
                                    partition_strategy="mod",
                                    name=None,
                                    max_norm=None,
                                    transform_fn=None):
  """Helper function for embedding_lookup and _compute_sampled_logits.

  This function is a generalization of embedding_lookup that optionally
  applies a caller-specified transformation to each embedding. This is
  done through the `transform_fn` argument. If provided, the function is
  applied to each partitioned tensor of retrieved embeddings, colocated
  with the embeddings. This function will be called with a single `Tensor`
  argument of the same type as the `params` tensor and should return a
  `Tensor`. The shape of the argument will be the same as `params` except
  for the size of the first dimension. The first dimension of the result's
  shape must be the same size as the argument's.

  Args:
    params: See embedding_lookup.
    ids: See embedding_lookup.
    partition_strategy: See embedding_lookup.
    name: See embedding_lookup.
    max_norm: See embedding_lookup.
    transform_fn: An optional function to apply to each retrieved embedding.
      If max_norm is provided, transform_fn is applied to the norm-limited
      embeddings.

  Returns:
    See embedding_lookup for details.
  Raises:
    ValueError: If `params` is empty.
  """
  if params is None or params in ((), []):
    raise ValueError("Need at least one param")
  if isinstance(params, variables.PartitionedVariable):
    params = list(params)  # Iterate to get the underlying Variables.
  if not isinstance(params, list):
    params = [params]

  with ops.name_scope(name, "embedding_lookup", params + [ids]) as name:
    np = len(params)  # Number of partitions
    # Preserve the resource variable status to avoid accidental dense reads.
    if not any(
        isinstance(p, resource_variable_ops.ResourceVariable) for p in params):
      params = ops.convert_n_to_tensor_or_indexed_slices(params, name="params")
    ids = ops.convert_to_tensor(ids, name="ids")
    if np == 1 and (not transform_fn or ids.get_shape().ndims == 1):
      with ops.colocate_with(params[0]):
        result = _clip(array_ops.gather(params[0], ids, name=name),
                       ids, max_norm)
        if transform_fn:
          result = transform_fn(result)
      # Make sure the final result does not have colocation contraints on the
      # params. Similar to the case np > 1 where parallel_dynamic_stitch is
      # outside the scioe of all with ops.colocate_with(params[p]).
      return array_ops.identity(result)
    else:
      # Flatten the ids. There are two cases where we need to do this.
      # - There is more than one params tensor.
      # - There is a transform_fn and ids is not statically known to be 1-D.
      #   We must flatten in this case because transform_fn expects a flat
      #   tensor of embeddings.
      flat_ids = array_ops.reshape(ids, [-1])
      original_indices = math_ops.range(array_ops.size(flat_ids))

      # Create p_assignments and set new_ids depending on the strategy.
      if partition_strategy == "mod":
        p_assignments = flat_ids % np
        new_ids = flat_ids // np
      elif partition_strategy == "div":
        # Compute num_total_ids as the sum of dim-0 of params, then assign to
        # partitions based on a constant number of ids per partition. Optimize
        # if we already know the full shape statically.
        dim_0_size = tensor_shape.Dimension(tensor_shape.dimension_value(
            params[0].get_shape()[0]))
        for p in xrange(1, np):
          dim_0_size += tensor_shape.Dimension(tensor_shape.dimension_value(
              params[p].get_shape()[0]))
        if dim_0_size.value:
          num_total_ids = constant_op.constant(dim_0_size.value, flat_ids.dtype)
        else:
          dim_0_sizes = []
          for p in xrange(np):
            param_p_dim = tensor_shape.dimension_value(params[p].get_shape()[0])
            if param_p_dim is not None:
              dim_0_sizes.append(param_p_dim)
            else:
              with ops.colocate_with(params[p]):
                dim_0_sizes.append(array_ops.shape(params[p])[0])
          num_total_ids = math_ops.reduce_sum(
              math_ops.cast(array_ops.stack(dim_0_sizes), flat_ids.dtype))
        ids_per_partition = num_total_ids // np
        extras = num_total_ids % np

        p_assignments = math_ops.maximum(
            flat_ids // (ids_per_partition + 1),
            (flat_ids - extras) // ids_per_partition)

        # Emulate a conditional using a boolean indicator tensor
#.........这里部分代码省略.........

def gradients(ys, xs, grad_ys=None, name="gradients",
              colocate_gradients_with_ops=False,
              gate_gradients=False,
              aggregation_method=None):
  """Constructs symbolic partial derivatives of `ys` w.r.t. x in `xs`.

  `ys` and `xs` are each a `Tensor` or a list of tensors.  `grad_ys`
  is a list of `Tensor`, holding the gradients received by the
  `ys`. The list must be the same length as `ys`.

  `gradients()` adds ops to the graph to output the partial
  derivatives of `ys` with respect to `xs`.  It returns a list of
  `Tensor` of length `len(xs)` where each tensor is the `sum(dy/dx)`
  for y in `ys`.

  `grad_ys` is a list of tensors of the same length as `ys` that holds
  the initial gradients for each y in `ys`.  When `grad_ys` is None,
  we fill in a tensor of '1's of the shape of y for each y in `ys`.  A
  user can provide their own initial 'grad_ys` to compute the
  derivatives using a different initial gradient for each y (e.g., if
  one wanted to weight the gradient differently for each value in
  each y).

  Args:
    ys: A `Tensor` or list of tensors to be differentiated.
    xs: A `Tensor` or list of tensors to be used for differentiation.
    grad_ys: Optional. A `Tensor` or list of tensors the same size as
      `ys` and holding the gradients computed for each y in `ys`.
    name: Optional name to use for grouping all the gradient ops together.
      defaults to 'gradients'.
    colocate_gradients_with_ops: If True, try colocating gradients with
      the corresponding op.
    gate_gradients: If True, add a tuple around the gradients returned
      for an operations.  This avoids some race conditions.
    aggregation_method: Specifies the method used to combine gradient terms.
      Accepted values are constants defined in the class `AggregationMethod`.

  Returns:
    A list of `sum(dy/dx)` for each x in `xs`.

  Raises:
    LookupError: if one of the operations between `x` and `y` does not
      have a registered gradient function.
    ValueError: if the arguments are invalid.

  """
  ys = _AsList(ys)
  xs = _AsList(xs)
  if grad_ys is None:
    grad_ys = [None] * len(ys)
  else:
    grad_ys = _AsList(grad_ys)
  with ops.op_scope(ys + xs + grad_ys, name, "gradients"):
    ys = ops.convert_n_to_tensor_or_indexed_slices(ys, name="y")
    xs = ops.convert_n_to_tensor_or_indexed_slices(xs, name="x")
    grad_ys = _DefaultGradYs(grad_ys, ys, colocate_gradients_with_ops)

    # The approach we take here is as follows: Create a list of all ops in the
    # subgraph between the ys and xs.  Visit these ops in reverse order of ids
    # to ensure that when we visit an op the gradients w.r.t its outputs have
    # been collected.  Then aggregate these gradients if needed, call the op's
    # gradient function, and add the generated gradients to the gradients for
    # its input.

    # Initialize the pending count for ops in the connected subgraph from ys
    # to the xs.
    to_ops = [t.op for t in ys]
    from_ops = [t.op for t in xs]
    pending_count, has_control_flow = _PendingCount(
        ops.get_default_graph(), to_ops, from_ops)

    # Iterate over the collected ops.
    #
    # grads: op => list of gradients received on each output endpoint of the
    # op.  The gradients for each endpoint are initially collected as a list.
    # When it is time to call the op's gradient function, for each endpoint we
    # aggregate the list of received gradients into a Add() Operation if there
    # is more than one.
    grads = {}

    # Add the initial gradients for the ys.
    for y, grad_y in zip(ys, grad_ys):
      _SetGrad(grads, y, grad_y)

    # Initialize queue with to_ops.
    queue = collections.deque()
    # Add the ops in 'to_ops' into the queue.
    to_ops_set = set()
    for op in to_ops:
      if op._id not in to_ops_set:
        to_ops_set.add(op._id)
        queue.append(op)
    # The set of 'from_ops'.
    stop_ops = _StopOps(from_ops, pending_count)
    while queue:
      # generate gradient subgraph for op.
      op = queue.popleft()
      with ops.device(_GetGradsDevice(op, colocate_gradients_with_ops)):
        if has_control_flow:
          control_flow_ops.EnterGradWhileContext(op)
#.........这里部分代码省略.........