def testShapeInferenceKnownShape(self):
    with self.session(use_gpu=False):
      indices = array_ops.placeholder(dtypes.int64)

      shape = [4, 5, 6]
      output = sparse_ops.sparse_to_dense(indices, shape, 1, 0)
      self.assertEqual(output.get_shape(), [4, 5, 6])

      shape = array_ops.placeholder(dtypes.int64, shape=(3,))
      output = sparse_ops.sparse_to_dense(indices, shape, 1, 0)
      self.assertEqual(output.get_shape().as_list(), [None, None, None])

  def testSparseExpandDims(self):
    for rank in range(1, 4):
      # Create a dummy input. When rank=3, shape=[2, 4, 6].
      shape = np.arange(1, rank + 1) * 2
      before = np.arange(np.prod(shape)).reshape(shape)

      # Make entries sparse.
      before *= np.random.binomial(1, .2, before.shape)
      dense_shape = before.shape
      indices = np.array(np.where(before)).T
      values = before[before != 0]

      # Try every possible valid value of axis.
      for axis in range(-rank - 1, rank):
        expected_after = np.expand_dims(before, axis)

        for axis_as_tensor in [False, True]:
          dense_shape_t = constant_op.constant(dense_shape, dtype=dtypes.int64)
          indices_t = constant_op.constant(indices)
          values_t = constant_op.constant(values)
          before_t = sparse_tensor.SparseTensor(
              indices=indices_t, values=values_t, dense_shape=dense_shape_t)

          if axis_as_tensor:
            axis = constant_op.constant(axis)

          s = sparse_ops.sparse_expand_dims(before_t, axis)
          d = sparse_ops.sparse_to_dense(s.indices, s.dense_shape, s.values)
          self.assertAllEqual(self.evaluate(d), expected_after)

def one_hot_mask(labels, num_classes, scope=None):
  """Compute 1-hot encodings for masks.

  Given a label image, this computes the one hot encoding at
  each pixel.

  Args:
    labels: (batch_size, width, height, 1) tensor containing labels.
    num_classes: number of classes
    scope: optional scope name

  Returns:
    Tensor of shape (batch_size, width, height, num_classes) with
    a 1-hot encoding.
  """
  with ops.name_scope(scope, "OneHotMask", [labels]):
    height, width, depth = _shape(labels)
    assert depth == 1
    sparse_labels = math_ops.to_int32(array_ops.reshape(labels, [-1, 1]))
    sparse_size, _ = _shape(sparse_labels)
    indices = array_ops.reshape(math_ops.range(0, sparse_size, 1), [-1, 1])
    concated = array_ops.concat([indices, sparse_labels], 1)
    dense_result = sparse_ops.sparse_to_dense(concated,
                                              [sparse_size, num_classes], 1.0,
                                              0.0)
    result = array_ops.reshape(dense_result, [height, width, num_classes])
    return result

def _TopKGrad(op, grad, _):
  """Return the gradients for TopK.

  Args:
    op: The TopKOp for which we need to generate gradients.
    grad: Tensor. The gradients passed to the TopKOp.

  Returns:
    A list of two tensors, the first being the gradient w.r.t to the input and
    TopK, and the second being the gradient w.r.t. to the indices (all zero).
  """
  in_shape = array_ops.shape(op.inputs[0])
  ind_shape = array_ops.shape(op.outputs[1])

  ind_lastdim = array_ops.gather(ind_shape, array_ops.size(ind_shape) - 1)
  # Flatten indices to 2D.
  ind_2d = array_ops.reshape(op.outputs[1], array_ops.stack([-1, ind_lastdim]))

  in_lastdim = array_ops.gather(in_shape, array_ops.size(in_shape) - 1)
  outerdim = array_ops.shape(ind_2d)[0]
  # Compute linear indices (flattened to 1D).
  ind = array_ops.reshape(ind_2d + array_ops.expand_dims(
      math_ops.range(0, outerdim * in_lastdim, in_lastdim), -1), [-1])

  # Substitute grad to appropriate locations and fill the rest with zeros,
  # finally reshaping it to the original input shape.
  return [array_ops.reshape(
      sparse_ops.sparse_to_dense(ind,
                                 array_ops.reshape(
                                     math_ops.reduce_prod(in_shape), [1]),
                                 array_ops.reshape(grad, [-1]),
                                 validate_indices=False),
      in_shape), array_ops.zeros(
          [], dtype=dtypes.int32)]

  def _check(self, result_tensor, result_np, input_sp_t):
    self.assertAllEqual(input_sp_t.indices.eval(), result_tensor.indices.eval())
    self.assertAllEqual(input_sp_t.shape.eval(), result_tensor.shape.eval())

    res_densified = sparse_ops.sparse_to_dense(result_tensor.indices,
                                               result_tensor.shape,
                                               result_tensor.values).eval()
    self.assertAllEqual(res_densified, result_np)

 def test_one(n, m, as_tensors):
   expected = np.eye(n, m)
   if as_tensors:
     m = constant_op.constant(m)
     n = constant_op.constant(n)
   s = sparse_ops.sparse_eye(n, m)
   d = sparse_ops.sparse_to_dense(s.indices, s.dense_shape, s.values)
   self.assertAllEqual(self.evaluate(d), expected)

  def _check(self, result_tensor, result_np, input_sp_t):
    self.assertTrue(isinstance(result_tensor, sparse_tensor.SparseTensor))
    self.assertTrue(isinstance(input_sp_t, sparse_tensor.SparseTensor))
    self.assertAllEqual(input_sp_t.indices, result_tensor.indices)
    self.assertAllEqual(input_sp_t.dense_shape, result_tensor.dense_shape)

    res_densified = sparse_ops.sparse_to_dense(
        result_tensor.indices, result_tensor.dense_shape, result_tensor.values)
    self.assertAllEqual(result_np, res_densified)

def sufficient_statistics(x, axes, shift=None, keep_dims=False, name=None):
  """Calculate the sufficient statistics for the mean and variance of `x`.

  These sufficient statistics are computed using the one pass algorithm on
  an input that's optionally shifted. See:
  https://en.wikipedia.org/wiki/Algorithms_for_calculating_variance#Computing_shifted_data

  Args:
    x: A `Tensor`.
    axes: Array of ints. Axes along which to compute mean and variance.
    shift: A `Tensor` containing the value by which to shift the data for
      numerical stability, or `None` if no shift is to be performed. A shift
      close to the true mean provides the most numerically stable results.
    keep_dims: produce statistics with the same dimensionality as the input.
    name: Name used to scope the operations that compute the sufficient stats.

  Returns:
    Four `Tensor` objects of the same type as `x`:
    * the count (number of elements to average over).
    * the (possibly shifted) sum of the elements in the array.
    * the (possibly shifted) sum of squares of the elements in the array.
    * the shift by which the mean must be corrected or None if `shift` is None.
  """
  with ops.op_scope([x, axes, shift], name, "sufficient_statistics"):
    x = ops.convert_to_tensor(x, name="x")
    x_shape = x.get_shape()
    if x_shape.is_fully_defined():
      counts = 1
      m_shape = []
      for d in xrange(x_shape.ndims):
        dim = x_shape[d].value
        if d in set(axes):
          counts *= dim
          dim = 1
        m_shape.append(dim)
      counts = constant_op.constant(counts, dtype=x.dtype)
    else:  # shape needs to be inferred at runtime.
      x_shape = array_ops.shape(x)
      select_axes = sparse_ops.sparse_to_dense(axes, array_ops.shape(x_shape),
                                               True, False)
      m_shape = math_ops.select(select_axes, array_ops.ones_like(x_shape),
                                x_shape)
      counts = math_ops.cast(
          math_ops.reduce_prod(x_shape / m_shape),
          x.dtype,
          name="count")
    if shift is not None:
      shift = ops.convert_to_tensor(shift, name="shift")
      m_ss = math_ops.sub(x, shift)
      v_ss = math_ops.squared_difference(x, shift)
    else:  # no shift.
      m_ss = x
      v_ss = math_ops.square(x)
    m_ss = math_ops.reduce_sum(m_ss, axes, keep_dims=keep_dims, name="mean_ss")
    v_ss = math_ops.reduce_sum(v_ss, axes, keep_dims=keep_dims, name="var_ss")
  return counts, m_ss, v_ss, shift

def sequence_loss_by_example(logits, targets, weights, num_decoder_symbols,
                             average_across_timesteps=True,
                             softmax_loss_function=None, name=None):
  """Weighted cross-entropy loss for a sequence of logits (per example).

  Args:
    logits: list of 2D Tensors of shape [batch_size x num_decoder_symbols].
    targets: list of 1D batch-sized int32 Tensors of the same length as logits.
    weights: list of 1D batch-sized float-Tensors of the same length as logits.
    num_decoder_symbols: integer, number of decoder symbols (output classes).
    average_across_timesteps: If set, divide the returned cost by the total
      label weight.
    softmax_loss_function: function (inputs-batch, labels-batch) -> loss-batch
      to be used instead of the standard softmax (the default if this is None).
    name: optional name for this operation, default: "sequence_loss_by_example".

  Returns:
    1D batch-sized float Tensor: the log-perplexity for each sequence.

  Raises:
    ValueError: if len(logits) is different from len(targets) or len(weights).
  """
  if len(targets) != len(logits) or len(weights) != len(logits):
    raise ValueError("Lengths of logits, weights, and targets must be the same "
                     "%d, %d, %d." % (len(logits), len(weights), len(targets)))
  with ops.op_scope(logits + targets + weights, name,
                    "sequence_loss_by_example"):
    batch_size = array_ops.shape(targets[0])[0]
    log_perp_list = []
    length = batch_size * num_decoder_symbols
    for i in xrange(len(logits)):
      if softmax_loss_function is None:
        # TODO(lukaszkaiser): There is no SparseCrossEntropy in TensorFlow, so
        # we need to first cast targets into a dense representation, and as
        # SparseToDense does not accept batched inputs, we need to do this by
        # re-indexing and re-sizing. When TensorFlow adds SparseCrossEntropy,
        # rewrite this method.
        indices = targets[i] + num_decoder_symbols * math_ops.range(batch_size)
        with ops.device("/cpu:0"):  # Sparse-to-dense must be on CPU for now.
          dense = sparse_ops.sparse_to_dense(
              indices, array_ops.expand_dims(length, 0), 1.0,
              0.0)
        target = array_ops.reshape(dense, [-1, num_decoder_symbols])
        crossent = nn_ops.softmax_cross_entropy_with_logits(
            logits[i], target, name="SequenceLoss/CrossEntropy{0}".format(i))
      else:
        crossent = softmax_loss_function(logits[i], targets[i])
      log_perp_list.append(crossent * weights[i])
    log_perps = math_ops.add_n(log_perp_list)
    if average_across_timesteps:
      total_size = math_ops.add_n(weights)
      total_size += 1e-12  # Just to avoid division by 0 for all-0 weights.
      log_perps /= total_size
  return log_perps

def _SparseToDense(sparse_indices,
                   output_size,
                   sparse_values,
                   default_value,
                   validate_indices=True):
  return sparse_ops.sparse_to_dense(
      sparse_indices,
      output_size,
      sparse_values,
      default_value=default_value,
      validate_indices=validate_indices)

def _SparseToDense(sparse_indices,
                   output_size,
                   sparse_values,
                   default_value,
                   validate_indices=True):
  feed_sparse_indices = array_ops.placeholder(dtypes.int32)
  feed_dict = {feed_sparse_indices: sparse_indices}
  return sparse_ops.sparse_to_dense(
      feed_sparse_indices,
      output_size,
      sparse_values,
      default_value=default_value,
      validate_indices=validate_indices).eval(feed_dict=feed_dict)

def _sparse_vs_dense_xent_benchmark_dense(labels, logits):
    labels = tf.identity(labels)
    logits = tf.identity(logits)
    with tf.device("/cpu:0"):  # Sparse-to-dense must be on CPU
        batch_size = tf.shape(logits)[0]
        num_entries = tf.shape(logits)[1]
        length = batch_size * num_entries
        labels += num_entries * tf.range(batch_size)
        target = sparse_ops.sparse_to_dense(labels, tf.pack([length]), 1.0, 0.0)
    target = tf.reshape(target, tf.pack([-1, num_entries]))
    crossent = tf.nn.softmax_cross_entropy_with_logits(logits, target, name="SequenceLoss/CrossEntropy")
    crossent_sum = tf.reduce_sum(crossent)
    grads = tf.gradients([crossent_sum], [logits])[0]

    return (crossent_sum, grads)

  def _apply_transform(self, input_tensors):
    """Applies the transformation to the `transform_input`.

    Args:
        input_tensors: a list of Tensors representing the input to
        the Transform.

    Returns:
        A namedtuple of Tensors representing the transformed output.
    """
    s = input_tensors[0]

     # pylint: disable=not-callable
    return self.return_type(sparse_ops.sparse_to_dense(
        s.indices, s.shape, s.values, default_value=self.default_value))

def _zero_out_float_grad(op, grad):
    """The gradients for `zero_out_float`.

  Args:
    op: The `zero_out_float` `Operation` that we are differentiating, which we can use
      to find the inputs and outputs of the original op.
    grad: Gradient with respect to the output of the `zero_out_float` op.

  Returns:
    Gradients with respect to the input of `zero_out_float`.
  """
    to_zero = op.inputs[0]
    shape = array_ops.shape(to_zero)
    index = array_ops.zeros_like(shape)
    first_grad = array_ops.reshape(grad, [-1])[0]
    to_zero_grad = sparse_ops.sparse_to_dense([index], shape, first_grad, 0)
    return [to_zero_grad]  # List of one Tensor, since we have one input

#.........这里部分代码省略.........
      labels = math_ops.cast(labels, dtypes.int64)
    labels_flat = array_ops.reshape(labels, [-1])

    # Sample the negative labels.
    #   sampled shape: [num_sampled] tensor
    #   true_expected_count shape = [batch_size, 1] tensor
    #   sampled_expected_count shape = [num_sampled] tensor
    if sampled_values is None:
      sampled_values = candidate_sampling_ops.log_uniform_candidate_sampler(
          true_classes=labels,
          num_true=num_true,
          num_sampled=num_sampled,
          unique=True,
          range_max=num_classes)
    # NOTE: pylint cannot tell that 'sampled_values' is a sequence
    # pylint: disable=unpacking-non-sequence
    sampled, true_expected_count, sampled_expected_count = sampled_values
    # pylint: enable=unpacking-non-sequence

    # labels_flat is a [batch_size * num_true] tensor
    # sampled is a [num_sampled] int tensor
    all_ids = array_ops.concat(0, [labels_flat, sampled])

    # weights shape is [num_classes, dim]
    all_w = embedding_ops.embedding_lookup(
        weights, all_ids, partition_strategy=partition_strategy)
    all_b = embedding_ops.embedding_lookup(biases, all_ids)
    # true_w shape is [batch_size * num_true, dim]
    # true_b is a [batch_size * num_true] tensor
    true_w = array_ops.slice(
        all_w, [0, 0], array_ops.pack([array_ops.shape(labels_flat)[0], -1]))
    true_b = array_ops.slice(all_b, [0], array_ops.shape(labels_flat))

    # inputs shape is [batch_size, dim]
    # true_w shape is [batch_size * num_true, dim]
    # row_wise_dots is [batch_size, num_true, dim]
    dim = array_ops.shape(true_w)[1:2]
    new_true_w_shape = array_ops.concat(0, [[-1, num_true], dim])
    row_wise_dots = math_ops.mul(
        array_ops.expand_dims(inputs, 1),
        array_ops.reshape(true_w, new_true_w_shape))
    # We want the row-wise dot plus biases which yields a
    # [batch_size, num_true] tensor of true_logits.
    dots_as_matrix = array_ops.reshape(row_wise_dots,
                                       array_ops.concat(0, [[-1], dim]))
    true_logits = array_ops.reshape(_sum_rows(dots_as_matrix), [-1, num_true])
    true_b = array_ops.reshape(true_b, [-1, num_true])
    true_logits += true_b

    # Lookup weights and biases for sampled labels.
    #   sampled_w shape is [num_sampled, dim]
    #   sampled_b is a [num_sampled] float tensor
    sampled_w = array_ops.slice(
        all_w, array_ops.pack([array_ops.shape(labels_flat)[0], 0]), [-1, -1])
    sampled_b = array_ops.slice(all_b, array_ops.shape(labels_flat), [-1])

    # inputs has shape [batch_size, dim]
    # sampled_w has shape [num_sampled, dim]
    # sampled_b has shape [num_sampled]
    # Apply X*W'+B, which yields [batch_size, num_sampled]
    sampled_logits = math_ops.matmul(inputs,
                                     sampled_w,
                                     transpose_b=True) + sampled_b

    if remove_accidental_hits:
      acc_hits = candidate_sampling_ops.compute_accidental_hits(
          labels, sampled, num_true=num_true)
      acc_indices, acc_ids, acc_weights = acc_hits

      # This is how SparseToDense expects the indices.
      acc_indices_2d = array_ops.reshape(acc_indices, [-1, 1])
      acc_ids_2d_int32 = array_ops.reshape(math_ops.cast(
          acc_ids, dtypes.int32), [-1, 1])
      sparse_indices = array_ops.concat(
          1, [acc_indices_2d, acc_ids_2d_int32], "sparse_indices")
      # Create sampled_logits_shape = [batch_size, num_sampled]
      sampled_logits_shape = array_ops.concat(
          0,
          [array_ops.shape(labels)[:1], array_ops.expand_dims(num_sampled, 0)])
      if sampled_logits.dtype != acc_weights.dtype:
        acc_weights = math_ops.cast(acc_weights, sampled_logits.dtype)
      sampled_logits += sparse_ops.sparse_to_dense(
          sparse_indices, sampled_logits_shape, acc_weights,
          default_value=0.0, validate_indices=False)

    if subtract_log_q:
      # Subtract log of Q(l), prior probability that l appears in sampled.
      true_logits -= math_ops.log(true_expected_count)
      sampled_logits -= math_ops.log(sampled_expected_count)

    # Construct output logits and labels. The true labels/logits start at col 0.
    out_logits = array_ops.concat(1, [true_logits, sampled_logits])
    # true_logits is a float tensor, ones_like(true_logits) is a float tensor
    # of ones. We then divide by num_true to ensure the per-example labels sum
    # to 1.0, i.e. form a proper probability distribution.
    out_labels = array_ops.concat(
        1, [array_ops.ones_like(true_logits) / num_true,
            array_ops.zeros_like(sampled_logits)])

  return out_logits, out_labels

 def _flat_map_fn(x):
   return dataset_ops.Dataset.from_tensor_slices(
       sparse_ops.sparse_to_dense(x.indices, x.dense_shape, x.values))

def sufficient_statistics(x, axes, shift=False, keep_dims=False, name=None):
  """Calculate the sufficient statistics for the mean and variance of `x`.

  These sufficient statistics are computed using the one pass algorithm on
  an input that's optionally shifted using the value of the 1st element in `x`.
  See:
  https://en.wikipedia.org/wiki/Algorithms_for_calculating_variance#Computing_shifted_data
  Unfortunately, in some cases using a random individual sample as the shift
  value leads experimentally to very poor numerical stability, so it is disabled
  by default. The one-pass approach might have to be revised accordingly.

  Args:
    x: A `Tensor`.
    axes: Array of ints. Axes along which to compute mean and variance.
    shift: If true, shift the data to provide more numerically stable results.
    keep_dims: produce statistics with the same dimensionality as the input.
    name: Name used to scope the operations that compute the sufficient stats.

  Returns:
    Four `Tensor` objects of the same type as `x`:
    * the count (number of elements to average over).
    * the (possibly shifted) sum of the elements in the array.
    * the (possibly shifted) sum of squares of the elements in the array.
    * the shift by which the mean must be corrected or None if `shift` is False.
  """
  with ops.op_scope([x, axes], name, "sufficient_statistics"):
    x = ops.convert_to_tensor(x, name="x")
    x_shape = x.get_shape()
    if x_shape.is_fully_defined():
      counts = 1
      m_shape = []
      for d in xrange(x_shape.ndims):
        dim = x_shape[d].value
        if d in set(axes):
          counts *= dim
          dim = 1
        m_shape.append(dim)
      counts = constant_op.constant(counts, dtype=x.dtype)
    else:  # shape needs to be inferred at runtime.
      x_shape = array_ops.shape(x)
      select_axes = sparse_ops.sparse_to_dense(axes, array_ops.shape(x_shape),
                                               True, False)
      m_shape = math_ops.select(select_axes, array_ops.ones_like(x_shape),
                                x_shape)
      counts = math_ops.cast(
          math_ops.reduce_prod(x_shape / m_shape),
          x.dtype,
          name="count")
    if shift:
      shift_value = array_ops.slice(x, array_ops.zeros_like(m_shape), m_shape)
      m_ss = math_ops.sub(x, shift_value)
      v_ss = math_ops.squared_difference(x, shift_value)
      if keep_dims:
        shift_value = array_ops.identity(shift_value, name="shift")
      else:
        shift_value = array_ops.squeeze(shift_value,
                                        squeeze_dims=axes,
                                        name="shift")
    else:  # not shift.
      m_ss = x
      v_ss = math_ops.square(x)
      shift_value = None
    m_ss = math_ops.reduce_sum(m_ss, axes, keep_dims=keep_dims, name="mean_ss")
    v_ss = math_ops.reduce_sum(v_ss, axes, keep_dims=keep_dims, name="var_ss")
  return counts, m_ss, v_ss, shift_value

#.........这里部分代码省略.........
  Returns:
    out_logits, out_labels: tensors with shape [batch_size, num_true +
        num_sampled] for passing to either SigmoidCrossEntropyWithLogits (NCE)
        or SoftmaxCrossEntropyWithLogits (sampled softmax).

  """

  with ops.op_scope(
      [weights, biases, inputs, labels], name, "compute_sampled_logits"):
    if labels.dtype != types.int64:
      labels = math_ops.cast(labels, types.int64)
    labels_flat = array_ops.reshape(labels, [-1])

    # Sample the negative labels.
    #   sampled shape: num_sampled vector
    #   true_expected_count shape = [batch_size, 1]
    #   sampled_expected_count shape = num_sampled vector
    if sampled_values is None:
      sampled_values = candidate_sampling_ops.log_uniform_candidate_sampler(
          true_classes=labels,
          num_true=num_true,
          num_sampled=num_sampled,
          unique=True,
          range_max=num_classes)
    # NOTE: pylint cannot tell that 'sampled_values' is a sequence
    # pylint: disable=unpacking-non-sequence
    sampled, true_expected_count, sampled_expected_count = sampled_values
    # pylint: enable=unpacking-non-sequence

    # weights shape is [num_classes, dim]
    # labels_flat is a [batch_size * num_true] vector
    # true_w shape is [batch_size * num_true, dim]
    # true_b is a [batch_size * num_true] vector
    true_w = embedding_ops.embedding_lookup(weights, labels_flat)
    true_b = embedding_ops.embedding_lookup(biases, labels_flat)

    # inputs shape is [batch_size, dim]
    # true_w shape is [batch_size * num_true, dim]
    # row_wise_dots is [batch_size, num_true, dim]
    dim = array_ops.shape(true_w)[1:2]
    new_true_w_shape = array_ops.concat(0, [[-1, num_true], dim])
    row_wise_dots = math_ops.mul(
        array_ops.expand_dims(inputs, 1),
        array_ops.reshape(true_w, new_true_w_shape))
    # We want the row-wise dot plus biases which yields a
    # [batch_size, num_true] tensor of true_logits.
    dots_as_matrix = array_ops.reshape(row_wise_dots,
                                       array_ops.concat(0, [[-1], dim]))
    true_logits = array_ops.reshape(_sum_rows(dots_as_matrix), [-1, num_true])
    true_b = array_ops.reshape(true_b, [-1, num_true])
    true_logits += true_b

    # Lookup weights and biases for sampled labels.
    #   sampled is a num_sampled int vector
    #   sampled_w shape is [num_sampled, dim]
    #   sampled_b is a num_sampled float vector
    sampled_w = embedding_ops.embedding_lookup(weights, sampled)
    sampled_b = embedding_ops.embedding_lookup(biases, sampled)

    # inputs has shape [batch_size, dim]
    # sampled_w has shape [num_sampled, dim]
    # sampled_b has shape [num_sampled]
    # Apply X*W'+B, which yields [batch_size, num_sampled]
    sampled_logits = math_ops.matmul(inputs,
                                     sampled_w,
                                     transpose_b=True) + sampled_b

    if remove_accidental_hits:
      acc_hits = candidate_sampling_ops.compute_accidental_hits(
          labels, sampled, num_true=num_true)
      acc_indices, acc_ids, acc_weights = acc_hits

      # This is how SparseToDense expects the indices.
      acc_indices_2d = array_ops.reshape(acc_indices, [-1, 1])
      acc_ids_2d_int32 = array_ops.reshape(math_ops.cast(
          acc_ids, types.int32), [-1, 1])
      sparse_indices = array_ops.concat(
          1, [acc_indices_2d, acc_ids_2d_int32], "sparse_indices")
      # Create sampled_logits_shape = [batch_size, num_sampled]
      sampled_logits_shape = array_ops.concat(
          0,
          [array_ops.shape(labels)[:1], array_ops.expand_dims(num_sampled, 0)])
      sampled_logits += sparse_ops.sparse_to_dense(
          sparse_indices, sampled_logits_shape, acc_weights, 0.0)

    if subtract_log_q:
      # Subtract log of Q(l), prior probability that l appears in sampled.
      true_logits -= math_ops.log(true_expected_count)
      sampled_logits -= math_ops.log(sampled_expected_count)

    # Construct output logits and labels. The true labels/logits start at col 0.
    out_logits = array_ops.concat(1, [true_logits, sampled_logits])
    # true_logits is a float tensor, ones_like(true_logits) is a float tensor
    # of ones. We then divide by num_true to ensure the per-example labels sum
    # to 1.0, i.e. form a proper probability distribution.
    out_labels = array_ops.concat(
        1, [array_ops.ones_like(true_logits) / num_true,
            array_ops.zeros_like(sampled_logits)])

  return out_logits, out_labels

 def testZeroDefault(self):
   with self.cached_session():
     x = sparse_ops.sparse_to_dense(2, [4], 7).eval()
     self.assertAllEqual(x, [0, 0, 7, 0])

 def testShapeInferenceUnknownShape(self):
   with self.session(use_gpu=False):
     indices = array_ops.placeholder(dtypes.int64)
     shape = array_ops.placeholder(dtypes.int64)
     output = sparse_ops.sparse_to_dense(indices, shape, 1, 0)
     self.assertEqual(output.get_shape().ndims, None)