Function group_embedding_lookup_sparse

tensorflow/python/ops/embedding_ops.py:1594–1930

This interface is designed for fused multiple embedding lookup. Args: params: list, tuple a list or tuple of trainable *Variable* or *EmbeddingVariable*. sp_ids: list, tuple a list or tuple of tf.SparseTensor or tf.RaggedTensor. btw Ragg

(params,
                                  sp_ids,
                                  combiners,
                                  sp_weights=None,
                                  partition_strategy='mod',
                                  is_sequence=False,
                                  params_num_per_group=sys.maxsize,
                                  name=None,
                                  )

@tf_export('nn.group_embedding_lookup_sparse')
def group_embedding_lookup_sparse(params,
                                  sp_ids,
                                  combiners,
                                  sp_weights=None,
                                  partition_strategy='mod',
                                  is_sequence=False,
                                  params_num_per_group=sys.maxsize,
                                  name=None,
                                  ):
    """
    This interface is designed for fused multiple embedding lookup.
    Args:
      params: list, tuple
              a list or tuple of trainable *Variable* or *EmbeddingVariable*.
      sp_ids: list, tuple
              a list or tuple of tf.SparseTensor or tf.RaggedTensor.
              btw RaggedTensor is preferred.
      combiners: list, tuple
              a list or tuple of string to specify the combiner of each embedding lookup, 
              supported args is *sum* or *mean*
      sp_weights: list, tuple
               a list or tuple of tf.SparseTensor used for embedding lookup.
      is_sequence: bool
                return list of `Tensor` of shape `[batch_size, D]` when is False
                return list of `Tensor` of shape `[batch_size, T, D]` when is True
      params_num_per_group: int
                The number of params in GroupEmbedding op.Function will schedule len(params) // params_num_per_group + 1
                GroupEmbedding Op. Default setting would launch one Op containing all params which is suitable for GPU scenarios
                to maximize the GPU utilization.On the contrast, you could set value to 1 when Op
                is placed on CPU so as to maximize inter parallelism.
      name: The operations name
    Returns
    -------
    emb_vec: list
            a list of tf.Tensor(the results of lookup).
  """

    if combiners is None:
        logging.warn('The default value of combiner will change from "mean" to "sqrtn" after 2016/11/01.'
                     )
        combiners = ['mean'] * len(params)
    if not isinstance(combiners, list):
        combiners = [combiners]
    for combiner in combiners:
        if combiner not in ('mean', 'sum'):
            raise ValueError("combiners must be one of 'mean', 'sum'")

    if params is None:
        raise ValueError('params must be specified')
    if not isinstance(params, list):
        params = [params]

    #Currently do not support PartitionedVariable.
    for index, param in enumerate(params):
      if isinstance(param, variables.PartitionedVariable):
        tmp_param = list(param)
        if len(tmp_param) != 1:
          raise TypeError("PartitionedVariable not support in 'group_embedding_lookup_sparse'. ")