dragnn_ops.bulk_fixed_embeddings(handle, embedding_matrix, component=None, pad_to_batch=None, pad_to_steps=None, name=None)
dragnn_ops.bulk_fixed_embeddings(handle, embedding_matrix, component=None, pad_to_batch=None, pad_to_steps=None, name=None)
Defined in tensorflow/dragnn/core/ops/gen_dragnn_bulk_ops.py
.
This op is a more efficient version of BulkFixedFeatures to be run with large
batch sizes at inference time. The op takes a handle to ComputeSession and embedding matrices as tensor inputs, and directly outputs concatenated embedding vectors.
handle
: ATensor
of typestring
. handle to ComputeSession. embedding_matrix (num_channels matrices of float): embedding matrices, each shaped as vocab_dim[channel] x embedding_dim[channel].embedding_matrix
: A list of at least 1Tensor
objects of the same type. embedding matrices.component
: An optionalstring
. Defaults to""
.pad_to_batch
: An optionalint
. Defaults to-1
.pad_to_steps
: An optionalint
. Defaults to-1
.name
: A name for the operation (optional).
A tuple of Tensor
objects (output_handle, embedding_vectors, num_steps). *
output_handle
: A Tensor
of type string
. handle to the same
ComputeSession after advancement. embedding_vectors (matrix of float): output
concatenated embeddings, shaped as (batch * beam * token) x
sum_channel(embedding_dim[channel]). num_steps (int32 scalar): batch was
unrolled for these many steps. * embedding_vectors
: A Tensor
. Has the
same type as embedding_matrix
. * num_steps
: A Tensor
of type
int32
.