def with_same_shape(expected_tensor, tensor):
  """Assert tensors are the same shape, from the same graph.

  Args:
    expected_tensor: Tensor with expected shape.
    tensor: Tensor of actual values.
  Returns:
    The original tensor argument, possibly with assert ops added.
  """
  with ops.name_scope('%s/' % tensor.op.name, values=[expected_tensor, tensor]):
    tensor_shape = expected_tensor.get_shape()
    expected_shape = (
        tensor_shape.as_list() if tensor_shape.is_fully_defined()
        else array_ops.shape(expected_tensor, name='expected_shape'))
    return with_shape(expected_shape, tensor)

def _AggregatedGrads(grads, op, loop_state, aggregation_method=None):
  """Get the aggregated gradients for op.

  Args:
    grads: The map of memoized gradients.
    op: The op to get gradients for.
    loop_state: An object for maintaining the state of the while loops in the
                graph. It is of type ControlFlowState. None if the graph
                contains no while loops.
    aggregation_method: Specifies the method used to combine gradient terms.
      Accepted values are constants defined in the class `AggregationMethod`.

  Returns:
    A list of gradients, one per each output of `op`. If the gradients
      for a particular output is a list, this function aggregates it
      before returning.

  Raises:
    TypeError: if the incoming grads are not Tensors or IndexedSlices.
    ValueError: if the arguments are invalid.

  """
  if aggregation_method is None:
    aggregation_method = AggregationMethod.DEFAULT
  if aggregation_method not in [
      AggregationMethod.ADD_N, AggregationMethod.EXPERIMENTAL_TREE,
      AggregationMethod.EXPERIMENTAL_ACCUMULATE_N
  ]:
    raise ValueError("Invalid aggregation_method specified %s." %
                     aggregation_method)
  out_grads = _GetGrads(grads, op)
  for i, out_grad in enumerate(out_grads):
    if loop_state:
      if isinstance(out_grad, (ops.Tensor, ops.IndexedSlices)):
        assert control_flow_ops.IsLoopSwitch(op)
        continue
    # Grads have to be Tensors or IndexedSlices
    if (isinstance(out_grad, collections.Sequence) and not all([
        isinstance(g, (ops.Tensor, ops.IndexedSlices)) for g in out_grad
        if g is not None
    ])):
      raise TypeError("gradients have to be either all Tensors "
                      "or all IndexedSlices")
    # Aggregate multiple gradients, and convert [] to None.
    if out_grad:
      if len(out_grad) < 2:
        used = "nop"
        out_grads[i] = out_grad[0]
      elif all([isinstance(g, ops.Tensor) for g in out_grad if g is not None]):
        tensor_shape = _AccumulatorShape(out_grad)
        if (aggregation_method == AggregationMethod.EXPERIMENTAL_ACCUMULATE_N
            and len(out_grad) > 2 and tensor_shape.is_fully_defined()):
          # The benefit of using AccumulateN is that its inputs can be combined
          # in any order and this can allow the expression to be evaluated with
          # a smaller memory footprint.  When used with gpu_allocator_retry,
          # it is possible to compute a sum of terms which are much larger than
          # total GPU memory.
          # AccumulateN can currently only be used if we know the shape for
          # an accumulator variable.  If this is not known, or if we only have
          # 2 grads then we fall through to the "tree" case below.
          used = "accumulate_n"
          out_grads[i] = math_ops.accumulate_n(out_grad)
        elif aggregation_method in [
            AggregationMethod.EXPERIMENTAL_TREE,
            AggregationMethod.EXPERIMENTAL_ACCUMULATE_N
        ]:
          # Aggregate all gradients by doing pairwise sums: this may
          # reduce performance, but it can improve memory because the
          # gradients can be released earlier.
          #
          # TODO(vrv): Consider replacing this with a version of
          # tf.AddN() that eagerly frees its inputs as soon as they are
          # ready, so the order of this tree does not become a problem.
          used = "tree"
          with ops.name_scope(op.name + "_gradient_sum"):
            running_sum = out_grad[0]
            for grad in out_grad[1:]:
              running_sum = math_ops.add_n([running_sum, grad])
            out_grads[i] = running_sum
        else:
          used = "add_n"
          out_grads[i] = _MultiDeviceAddN(out_grad)
        logging.vlog(2, "  _AggregatedGrads %d x %s using %s",
                     len(out_grad), tensor_shape, used)
      else:
        out_grad = math_ops._as_indexed_slices_list(
            [g for g in out_grad if g is not None])
        out_grad = [_HandleNestedIndexedSlices(x) for x in out_grad]
        # Form IndexedSlices out of the concatenated values and
        # indices.
        out_grads[i] = ops.IndexedSlices(
            array_ops.concat_v2([x.values for x in out_grad], 0),
            array_ops.concat_v2([x.indices for x in out_grad], 0),
            out_grad[0].dense_shape)
    else:
      out_grads[i] = []
  return out_grads