def _embed_word_indices(self, indices, embeddings):
"""Embed all indexes using the given embeddings.
Parameters
----------
indices : ndarray
indices of the items to embed using the given embeddings matrix
Note: indices are flattened
embeddings : ndarray (vocab_size x emb_size)
embeddings matrix
Returns
-------
ndarray (len(indices) x emb_size)
embedded indices
"""
concatenated_input = indices.flatten()
# Rami's fix
if config.device == 'gpu':
embedded = embeddings[concatenated_input]
else:
embedded = theano.sparse_grad(embeddings[concatenated_input])
return embedded
def link(self, inputs):
""" Input should be a matrix with the rows representing examples.
We need embeddings for all indices in a matrix, that's why we need to
flatten the matrix first, get all embeddings for the indices and than
reshape it again.
Parameters
----------
inputs
0) indexes of the examples to get the embeddings of
Returns
-------
[]
0) embeddings of the input
"""
self.inputs = inputs
input = self.inputs[0]
concatenated_input = input.flatten()
# Rami's fix
if config.device == 'gpu':
indexed_rows = self.weights[concatenated_input]
else:
indexed_rows = theano.sparse_grad(self.weights[concatenated_input])
concatenated_rows = indexed_rows.flatten()
num_examples = input.shape[0]
width = concatenated_rows.size // num_examples
self.outputs = [concatenated_rows.reshape((num_examples, width))]
return self.outputs
def _embed_word_indices(self, indices, embeddings):
"""Embed all indexes using the given embeddings.
Parameters
----------
indices : ndarray
indices of the items to embed using the given embeddings matrix
Note: indices are flattened
embeddings : ndarray (vocab_size x emb_size)
embeddings matrix
Returns
-------
ndarray (len(indices) x emb_size)
embedded indices
"""
concatenated_input = indices.flatten()
# Rami's fix
if config.device == 'gpu':
embedded = embeddings[concatenated_input]
else:
embedded = theano.sparse_grad(embeddings[concatenated_input])
return embedded
def __init__(self, input, w_values,embedding_dim,mini_batch_size,width, name="EmbeddingLayer", params=None):
self.n_in = 1.0
self.w_values=w_values
super(EmbeddingLayer, self).__init__(name,params)
concatenated_input = input.flatten()
if theano.config.device == 'cpu':
indexed_rows = theano.sparse_grad(self.weights[concatenated_input])
else:
indexed_rows = self.weights[concatenated_input]
concatenated_rows = indexed_rows.flatten()
#mini_batch_size = input.shape[0]
#width = input.shape[1]
self.output = concatenated_rows.reshape((mini_batch_size, width*embedding_dim))
self.params = [self.weights]