EmbeddingBagOffsetsSum#

Versioned name: EmbeddingBagOffsetsSum-3

Category: Sparse

Short description: Computes sums of “bags” of embeddings, without instantiating the intermediate embeddings.

Detailed description:

Operation EmbeddingBagOffsets is an implementation of torch.nn.EmbeddingBag with indices and offsets inputs being 1D tensors.

For each index in indices this operator gathers values from emb_table embedding table. Then values at indices in the range of the same bag (based on offset input) are reduced according to reduction attribute.

Values in offsets define starting index in indices tensor of each “bag”, e.g. offsets with value [0, 3, 4, 4, 6] define 5 “bags” containing [3, 1, 0, 2, num_indices-6] elements corresponding to [indices[0:3], indices[3:4], empty_bag, indices[4:6], indices[6:]] slices of indices per bag.

EmbeddingBagOffsetsSum is an equivalent to following NumPy snippet:

def embedding_bag_offsets(
    emb_table: np.ndarray,
    indices: np.ndarray,
    offsets: np.ndarray,
    default_index: Optional[int] = None,
    per_sample_weights: Optional[np.ndarray] = None,
):
    if per_sample_weights is None:
        per_sample_weights = np.ones_like(indices)
    embeddings = []
    for emb_idx, emb_weight in zip(indices, per_sample_weights):
        embeddings.append(emb_table[emb_idx] * emb_weight)
    previous_offset = offsets[0]
    bags = []
    offsets = np.append(offsets, len(indices))
    for bag_offset in offsets[1:]:
        bag_size = bag_offset - previous_offset
        if bag_size != 0:
            embedding_bag = embeddings[previous_offset:bag_offset]
            reduced_bag = np.add.reduce(embedding_bag)
            bags.append(reduced_bag)
        else:
            # Empty bag case
            if default_index is not None and default_index != -1:
                bags.append(emb_table[default_index])
            else:
                bags.append(np.zeros(emb_table.shape[1:]))
        previous_offset = bag_offset
    return np.stack(bags, axis=0)

Attributes: EmbeddingBagOffsetsSum operation has no attributes.

Inputs:

  • 1: emb_table tensor containing the embedding lookup table of the module of shape [num_emb, emb_dim1, emb_dim2, ...] and of type T. Required.

  • 2: indices tensor of shape [num_indices] and of type T_IND. Required.

  • 3: offsets tensor of shape [batch] and of type T_IND containing the starting index positions of each “bag” in indices. Maximum value of offsets cannot be greater than length of indices. Required.

  • 4: default_index scalar of type T_IND containing default index in embedding table to fill empty “bags”. If set to -1 or not provided, empty “bags” are filled with zeros. Reverse indexing using negative values is not supported. Optional.

  • 5: per_sample_weights tensor of the same shape as indices and of type T. Each value in this tensor are multiplied with each value pooled from embedding table for each index. Optional, default is tensor of ones. Optional.

Outputs:

  • 1: tensor of shape [batch, emb_dim1, emb_dim2, ...] and of type T containing embeddings for each bag.

Types

  • T: any numeric type.

  • T_IND: int32 or int64.

Example

Example 1: per_sample_weights are provided, default_index is set to 0 to fill empty bag with values gathered form emb_table on given index.

<layer ... type="EmbeddingBagOffsetsSum" ... >
    <input>
        <port id="0">     <!-- emb_table value is: [[-0.2, -0.6], [-0.1, -0.4], [-1.9, -1.8], [-1.,  1.5], [ 0.8, -0.7]] -->
            <dim>5</dim>
            <dim>2</dim>
        </port>
        <port id="1">     <!-- indices value is: [0, 2, 3, 4] -->
            <dim>4</dim>
        </port>
        <port id="2">     <!-- offsets value is: [0, 2, 2] - 3 "bags" containing [2,0,4-2] elements, second "bag" is empty -->
            <dim>3</dim>
        </port>
        <port id="3"/>    <!-- default_index value is: 0 -->
        <port id="4"/>    <!-- per_sample_weights value is: [0.5, 0.5, 0.5, 0.5] -->
            <dim>4</dim>
        </port>
    </input>
    <output>
        <port id="5">     <!-- output value is: [[-1.05, -1.2], [-0.2, -0.6], [-0.1, 0.4]] -->
            <dim>3</dim>
            <dim>2</dim>
        </port>
    </output>
</layer>

Example 2: per_sample_weights are provided, default_index is set to -1 to fill empty bag with 0.

<layer ... type="EmbeddingBagOffsets" ... >
    <input>
        <port id="0">     <!-- emb_table value is: [[-0.2, -0.6], [-0.1, -0.4], [-1.9, -1.8], [-1.,  1.5], [ 0.8, -0.7]] -->
            <dim>5</dim>
            <dim>2</dim>
        </port>
        <port id="1">     <!-- indices value is: [0, 2, 3, 4] -->
            <dim>4</dim>
        </port>
        <port id="2">     <!-- offsets value is: [0, 2, 2] - 3 "bags" containing [2,0,4-2] elements, second "bag" is empty -->
            <dim>3</dim>
        </port>
        <port id="3"/>    <!-- default_index value is: -1 - fill empty bag with 0-->
        <port id="4"/>    <!-- per_sample_weights value is: [0.5, 0.5, 0.5, 0.5] -->
            <dim>4</dim>
        </port>
    </input>
    <output>
        <port id="5">     <!-- output value is: [[-1.05, -1.2], [0., 0.], [-0.1, 0.4]] -->
            <dim>3</dim>
            <dim>2</dim>
        </port>
    </output>
</layer>