def _embedding_lookup_with_distributed_aggregation(params,
                                                   ids,
                                                   partition_strategy="mod",
                                                   name=None,
                                                   max_norm=None,
                                                   weights=None,
                                                   idx=None,
                                                   segment_ids=None):
  """Lookup helper for embedding_lookup_sparse_with_distributed_aggregation."""
  if params is None or params == []:  # pylint: disable=g-explicit-bool-comparison
    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]

  def maybe_normalize(x):
    if max_norm is not None:
      if x.get_shape().ndims is not None:
        ndims = x.get_shape().ndims
      else:
        ndims = array_ops.size(array_ops.shape(x))
      return clip_ops.clip_by_norm(x, max_norm, axes=list(range(1, ndims)))
    return x

  with ops.name_scope(name, "embedding_lookup_with_distributed_aggregation",
                      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")
    if np == 1:
      with ops.colocate_with(params[0]):
        ret = maybe_normalize(_do_gather(params[0], ids))
        ignore_weights = weights is None
        if not ignore_weights:
          if weights.dtype != ret.dtype:
            weights = math_ops.cast(weights, ret.dtype)
          # Reshape to allow broadcast
          ones = array_ops.fill(
              array_ops.expand_dims(array_ops.rank(ret) - 1, 0), 1)
          bcast_weights_shape = array_ops.concat(
              [array_ops.shape(weights), ones], 0)
          orig_weights_shape = weights.get_shape()
          weights = array_ops.reshape(weights, bcast_weights_shape)
          # Set weights shape after reshape
          if ret.get_shape().ndims is not None:
            weights.set_shape(
                orig_weights_shape.concatenate(
                    [1 for _ in range(ret.get_shape().ndims - 1)]))
          ret *= weights
          return math_ops.segment_sum(ret, segment_ids, name=name)
        else:
          return math_ops.sparse_segment_sum(ret, idx, segment_ids, name=name)
    else:
      ids = ops.convert_to_tensor(ids, name="ids")
      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 = params[0].get_shape()[0]
        for p in xrange(1, np):
          dim_0_size += 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):
            if params[p].get_shape()[0].value is not None:
              dim_0_sizes.append(params[p].get_shape()[0].value)
            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
        is_in_first_extras_partitions = math_ops.cast(p_assignments < extras,
                                                      flat_ids.dtype)
        new_ids = (is_in_first_extras_partitions * (flat_ids %
                                                    (ids_per_partition + 1)) +
                   (1 - is_in_first_extras_partitions) * (
                       (flat_ids - extras) % ids_per_partition))
      else:
        raise ValueError("Unrecognized partition strategy: " +
                         partition_strategy)

#.........这里部分代码省略.........

def scattered_embedding_lookup_sparse(params,
                                      sparse_values,
                                      dimension,
                                      combiner=None,
                                      default_value=None,
                                      name=None,
                                      hash_key=None):
  """Looks up embeddings of a sparse feature using parameter hashing.

  See `tf.contrib.layers.scattered_embedding_lookup` for embedding with hashing.

  Args:
    params: A `Tensor`, `list` of `Tensors`, or `PartitionedVariable`.
      Each tensor must be of rank 1 with fully-defined shape.
    sparse_values: A 2-D `SparseTensor` containing the values to be embedded.
      Some rows may be empty.
    dimension: Embedding dimension
    combiner: A string specifying how to combine embedding results for each
        entry. Currently "mean", "sqrtn" and "sum" are supported, with "mean"
        the default.
    default_value: The value to use for an entry with no features.
    name: An optional name for this op.
    hash_key: Specify the hash_key that will be used by the `FingerprintCat64`
      function to combine the crosses fingerprints on SparseFeatureCrossOp
      (optional).

  Returns:
     Dense tensor with shape [N, dimension] with N the number of rows in
       sparse_values.

  Raises:
    TypeError: If sparse_values is not a SparseTensor.
    ValueError: If combiner is not one of {"mean", "sqrtn", "sum"}.
  """
  if combiner is None:
    logging.warn("The default value of combiner will change from \"mean\" "
                 "to \"sqrtn\" after 2016/11/01.")
    combiner = "mean"
  if isinstance(params, variables.PartitionedVariable):
    params = list(params)
  if not isinstance(params, list):
    params = [params]
  if not isinstance(sparse_values, sparse_tensor.SparseTensor):
    raise TypeError("sparse_values must be SparseTensor")

  with ops.name_scope(name, "scattered_embedding_lookup_sparse",
                      params + [sparse_values]) as scope:
    # Fill in the empty rows.
    if default_value is None:
      # Random default values to reduce the risk of collision.
      if sparse_values.dtype == dtypes.string:
        default_value = "6ZxWzWOHxZ"
      else:
        default_value = 1288896567
    sparse_values, _ = sparse_ops.sparse_fill_empty_rows(
        sparse_values, default_value)

    segment_ids = sparse_values.indices[:, 0]
    if segment_ids.dtype != dtypes.int32:
      segment_ids = math_ops.cast(segment_ids, dtypes.int32)

    values = sparse_values.values
    values, idx = array_ops.unique(values)

    embeddings = scattered_embedding_lookup(
        params, values, dimension, hash_key=hash_key)

    if combiner == "sum":
      embeddings = math_ops.sparse_segment_sum(embeddings, idx, segment_ids,
                                               name=scope)
    elif combiner == "mean":
      embeddings = math_ops.sparse_segment_mean(embeddings, idx, segment_ids,
                                                name=scope)
    elif combiner == "sqrtn":
      embeddings = math_ops.sparse_segment_sqrt_n(embeddings, idx, segment_ids,
                                                  name=scope)
    else:
      raise ValueError("Combiner must be one of 'mean', 'sqrtn' or 'sum'.")

    return embeddings

def hashed_embedding_lookup_sparse(params,
                                   sparse_values,
                                   dimension,
                                   combiner="mean",
                                   default_value=None,
                                   name=None):
  """Looks up embeddings of a sparse feature using parameter hashing.

  See `tf.contrib.layers.hashed_embedding_lookup` for embedding with hashing.

  Args:
    params: A `Tensor` or `list` of `Tensors`.
      Each tensor must be of rank 1 with fully-defined shape.
    sparse_values: A 2-D `SparseTensor` containing the values to be embedded.
      Some rows may be empty.
    dimension: Embedding dimension
    combiner: A string specifying how to combine embedding results for each
        entry. Currently "mean", "sqrtn" and "sum" are supported, with "mean"
        the default.
    default_value: The value to use for an entry with no features.
    name: An optional name for this op.

  Returns:
     Dense tensor with shape [N, dimension] with N the number of rows in
       sparse_values.

  Raises:
    TypeError: If sparse_values is not a SparseTensor.
    ValueError: If combiner is not one of {"mean", "sqrtn", "sum"}.
  """

  if not isinstance(params, list):
    params = [params]
  if not isinstance(sparse_values, ops.SparseTensor):
    raise TypeError("sparse_values must be SparseTensor")

  with ops.name_scope(name, "hashed_sparse_embedding_lookup",
                      params + [sparse_values]) as scope:
    # Fill in the empty rows.
    if default_value is None:
      # Random default values to reduce the risk of collision.
      if sparse_values.dtype == dtypes.string:
        default_value = "6ZxWzWOHxZ"
      else:
        default_value = 1288896567
    sparse_values, _ = sparse_ops.sparse_fill_empty_rows(
        sparse_values, default_value)

    segment_ids = sparse_values.indices[:, 0]
    if segment_ids.dtype != dtypes.int32:
      segment_ids = math_ops.cast(segment_ids, dtypes.int32)

    values = sparse_values.values
    values, idx = array_ops.unique(values)

    embeddings = hashed_embedding_lookup(params, values, dimension)

    if combiner == "sum":
      embeddings = math_ops.sparse_segment_sum(embeddings, idx, segment_ids,
                                               name=scope)
    elif combiner == "mean":
      embeddings = math_ops.sparse_segment_mean(embeddings, idx, segment_ids,
                                                name=scope)
    elif combiner == "sqrtn":
      embeddings = math_ops.sparse_segment_sqrt_n(embeddings, idx, segment_ids,
                                                  name=scope)
    else:
      raise ValueError("Combiner must be one of 'mean', 'sqrtn' or 'sum'.")

    return embeddings

#.........这里部分代码省略.........

      output[0, :] = (params[1, :] * 2.0 + params[3, :] * 0.5) / (2.0 + 0.5)
      output[1, :] = params[0, :] * 1.0
      output[2, :] = params[1, :] * 3.0

  Raises:
    TypeError: If sp_ids is not a SparseTensor, or if sp_weights is neither
      None nor SparseTensor.
    ValueError: If combiner is not one of {"mean", "sqrtn", "sum"}.
  """
  if combiner is None:
    logging.warn("The default value of combiner will change from \"mean\" "
                 "to \"sqrtn\" after 2016/11/01.")
    combiner = "mean"
  if combiner not in ("mean", "sqrtn", "sum"):
    raise ValueError("combiner must be one of 'mean', 'sqrtn' or 'sum'")
  if isinstance(params, variables.PartitionedVariable):
    params = list(params)  # Iterate to get the underlying Variables.
  if not isinstance(params, list):
    params = [params]
  if not isinstance(sp_ids, sparse_tensor.SparseTensor):
    raise TypeError("sp_ids must be SparseTensor")
  ignore_weights = sp_weights is None
  if not ignore_weights:
    if not isinstance(sp_weights, sparse_tensor.SparseTensor):
      raise TypeError("sp_weights must be either None or SparseTensor")
    sp_ids.values.get_shape().assert_is_compatible_with(
        sp_weights.values.get_shape())
    sp_ids.indices.get_shape().assert_is_compatible_with(
        sp_weights.indices.get_shape())
    sp_ids.shape.get_shape().assert_is_compatible_with(
        sp_weights.shape.get_shape())
    # TODO(yleon): Add enhanced node assertions to verify that sp_ids and
    # sp_weights have equal indices and shapes.

  with ops.name_scope(name, "embedding_lookup_sparse",
                      params + [sp_ids]) as name:
    segment_ids = sp_ids.indices[:, 0]
    if segment_ids.dtype != dtypes.int32:
      segment_ids = math_ops.cast(segment_ids, dtypes.int32)

    ids = sp_ids.values
    if ignore_weights:
      ids, idx = array_ops.unique(ids)
    else:
      idx = None

    embeddings = embedding_lookup(
        params, ids, partition_strategy=partition_strategy, max_norm=max_norm)
    if not ignore_weights:
      weights = sp_weights.values
      if weights.dtype != embeddings.dtype:
        weights = math_ops.cast(weights, embeddings.dtype)

      # Reshape weights to allow broadcast
      ones = array_ops.fill(
          array_ops.expand_dims(array_ops.rank(embeddings) - 1, 0), 1)
      bcast_weights_shape = array_ops.concat(0, [
          array_ops.shape(weights), ones])

      orig_weights_shape = weights.get_shape()
      weights = array_ops.reshape(weights, bcast_weights_shape)

      # Set the weight shape, since after reshaping to bcast_weights_shape,
      # the shape becomes None.
      if embeddings.get_shape().ndims is not None:
        weights.set_shape(orig_weights_shape.concatenate(
            [1 for _ in range(embeddings.get_shape().ndims - 1)]))

      embeddings *= weights

      if combiner == "sum":
        embeddings = math_ops.segment_sum(embeddings, segment_ids, name=name)
      elif combiner == "mean":
        embeddings = math_ops.segment_sum(embeddings, segment_ids)
        weight_sum = math_ops.segment_sum(weights, segment_ids)
        embeddings = math_ops.div(embeddings, weight_sum, name=name)
      elif combiner == "sqrtn":
        embeddings = math_ops.segment_sum(embeddings, segment_ids)
        weights_squared = math_ops.pow(weights, 2)
        weight_sum = math_ops.segment_sum(weights_squared, segment_ids)
        weight_sum_sqrt = math_ops.sqrt(weight_sum)
        embeddings = math_ops.div(embeddings, weight_sum_sqrt, name=name)
      else:
        assert False, "Unrecognized combiner"
    else:
      assert idx is not None
      if combiner == "sum":
        embeddings = math_ops.sparse_segment_sum(embeddings, idx, segment_ids,
                                                 name=name)
      elif combiner == "mean":
        embeddings = math_ops.sparse_segment_mean(embeddings, idx, segment_ids,
                                                  name=name)
      elif combiner == "sqrtn":
        embeddings = math_ops.sparse_segment_sqrt_n(embeddings, idx,
                                                    segment_ids, name=name)
      else:
        assert False, "Unrecognized combiner"

    return embeddings

#.........这里部分代码省略.........
      partitioned by id % P, and we do separate lookups in params[p] for
      0 <= p < P, and then stitch the results back together into a single
      result tensor. The first dimension is allowed to vary as the vocab
      size is not necessarily a multiple of P.
    sp_ids: N x M SparseTensor of int64 ids (typically from FeatureValueToId),
      where N is typically batch size and M is arbitrary.
    sp_weights: either a SparseTensor of float / double weights, or None to
      indicate all weights should be taken to be 1. If specified, sp_weights
      must have exactly the same shape and indices as sp_ids.
    name: Optional name for the op.
    combiner: A string specifying the reduction op. Currently "mean" and "sum"
      are supported.
      "sum" computes the weighted sum of the embedding results for each row.
      "mean" is the weighted sum divided by the total weight.

  Returns:
    A dense tensor representing the combined embeddings for the
    sparse ids. For each row in the dense tensor represented by sp_ids, the op
    looks up the embeddings for all ids in that row, multiplies them by the
    corresponding weight, and combines these embeddings as specified.

    In other words, if
      shape(combined params) = [p0, p1, ..., pm]
    and
      shape(sp_ids) = shape(sp_weights) = [d0, d1, ..., dn]
    then
      shape(output) = [d0, d1, ..., dn-1, p1, ..., pm].

    For instance, if params is a 10x20 matrix, and sp_ids / sp_weights are

      [0, 0]: id 1, weight 2.0
      [0, 1]: id 3, weight 0.5
      [1, 0]: id 0, weight 1.0
      [2, 3]: id 1, weight 3.0

    with combiner="mean", then the output will be a 3x20 matrix where
      output[0, :] = (params[1, :] * 2.0 + params[3, :] * 0.5) / (2.0 + 0.5)
      output[1, :] = params[0, :] * 1.0
      output[2, :] = params[1, :] * 3.0

  Raises:
    TypeError: If sp_ids is not a SparseTensor, or if sp_weights is neither
      None nor SparseTensor.
    ValueError: If combiner is not one of {"mean", "sum"}.
  """
  if combiner not in ("mean", "sum"):
    raise ValueError("combiner must be one of 'mean' or 'sum'")
  if not isinstance(params, list):
    params = [params]
  if not isinstance(sp_ids, ops.SparseTensor):
    raise TypeError("sp_ids must be SparseTensor")
  ignore_weights = sp_weights is None
  if not ignore_weights and not isinstance(sp_weights, ops.SparseTensor):
    raise TypeError("sp_weights must be either None or SparseTensor")

  with ops.op_scope(params + [sp_ids], name, "embedding_lookup_sparse") as name:
    segment_ids = sp_ids.indices[:, 0]
    if segment_ids.dtype != types.int32:
      segment_ids = math_ops.cast(segment_ids, types.int32)

    ids = sp_ids.values
    if ignore_weights:
      ids, idx = array_ops.unique(ids)
    else:
      idx = None

    embeddings = embedding_lookup(params, ids)
    if not ignore_weights:
      weights = sp_weights.values
      if weights.dtype != embeddings.dtype:
        weights = math_ops.cast(weights, embeddings.dtype)

      # Reshape weights to allow broadcast
      ones = array_ops.fill(
          array_ops.expand_dims(array_ops.rank(embeddings) - 1, 0), 1)
      bcast_weights_shape = array_ops.concat(0, [
          array_ops.shape(weights), ones])
      weights = array_ops.reshape(weights, bcast_weights_shape)
      embeddings *= weights

      if combiner == "sum":
        embeddings = math_ops.segment_sum(embeddings, segment_ids, name=name)
      elif combiner == "mean":
        embeddings = math_ops.segment_sum(embeddings, segment_ids)
        weight_sum = math_ops.segment_sum(weights, segment_ids)
        embeddings = math_ops.div(embeddings, weight_sum, name=name)
      else:
        assert False, "Unrecognized combiner"
    else:
      assert idx is not None
      if combiner == "sum":
        embeddings = math_ops.sparse_segment_sum(embeddings, idx, segment_ids,
                                                 name=name)
      elif combiner == "mean":
        embeddings = math_ops.sparse_segment_mean(embeddings, idx, segment_ids,
                                                  name=name)
      else:
        assert False, "Unrecognized combiner"

    return embeddings