def init_embed(self, embeddings: Dict[str, TensorDef],
**kwargs) -> BaseLayer:
"""This method creates the "embedding" layer of the inputs, with an optional reduction
:param embeddings: A dictionary of embeddings
:Keyword Arguments: See below
* *embeddings_reduction* (defaults to `concat`) An operator to perform on a stack of embeddings
* *embeddings_dropout = float(kwargs.get('embeddings_dropout', 0.0))
:return: The output of the embedding stack followed by its reduction. This will typically be an output
with an additional dimension which is the hidden representation of the input
"""
reduction = kwargs.get(
'embeddings_reduction',
kwargs.get('embed_reduction_type', 'concat-subtract'))
reduction = create_embeddings_reduction(embed_reduction_type=reduction,
**kwargs)
embeddings_dropout = float(kwargs.get('embeddings_dropout', 0.0))
if len(embeddings) != 1:
raise Exception("Currently we only support a single embedding")
key_name = list(embeddings.keys())[0]
key1 = f"{key_name}[0]"
key2 = f"{key_name}[1]"
embeddings_dual = {
key1: embeddings[key_name],
key2: embeddings[key_name]
}
return EmbeddingsStack(embeddings_dual,
embeddings_dropout,
reduction=reduction)
def init_embed(self, embeddings: Dict[str, TensorDef],
**kwargs) -> BaseLayer:
"""This method creates the "embedding" layer of the inputs, with an optional reduction
:param embeddings: A dictionary of embeddings
:param kwargs: See below
:Keyword Arguments:
* *embeddings_reduction* (defaults to `concat`) An operator to perform on a stack of embeddings
* *embeddings_name* (``str``) Optional override to Keras default names
* *embeddings_dropout* (``float``) how much dropout post-reduction (defaults to 0.0)
:return: The output of the embedding stack followed by its reduction. This will typically be an output
with an additional dimension which is the hidden representation of the input
"""
name = kwargs.get('embeddings_name')
reduction = kwargs.get('embeddings_reduction',
kwargs.get('embed_reduction_type', 'concat'))
reduction = create_embeddings_reduction(embed_reduction_type=reduction,
**kwargs)
embeddings_dropout = float(kwargs.get('embeddings_dropout', 0.0))
return EmbeddingsStack(embeddings,
embeddings_dropout,
reduction=reduction,
name=name)