def _get_embedding_layer(self, embedding_file=None):
if self.embedding_layer is None:
word_vocab_size = self.data_processor.get_vocab_size(onto_aware=False)
synset_vocab_size = self.data_processor.get_vocab_size(onto_aware=True)
if embedding_file is None:
if not self.tune_embedding:
print >>sys.stderr, "Pretrained embedding is not given. Setting tune_embedding to True."
self.tune_embedding = True
embedding_weights = None
else:
# TODO: Other sources for prior initialization
embedding = self.data_processor.get_embedding_matrix(embedding_file, onto_aware=True)
# Put the embedding in a list for Keras to treat it as weights of the embedding layer.
embedding_weights = [embedding]
if self.set_sense_priors:
initial_sense_prior_parameters = numpy.random.uniform(low=0.01, high=0.99,
size=(word_vocab_size, 1))
# While setting weights, Keras wants trainable weights first, and then the non trainable
# weights. If we are not tuning the embedding, we need to keep the sense priors first.
if not self.tune_embedding:
embedding_weights = [initial_sense_prior_parameters] + embedding_weights
else:
embedding_weights.append(initial_sense_prior_parameters)
self.embedding_layer = OntoAwareEmbedding(word_vocab_size, synset_vocab_size, self.embed_dim,
weights=embedding_weights, mask_zero=True,
set_sense_priors=self.set_sense_priors,
tune_embedding=self.tune_embedding,
name="embedding")
return self.embedding_layer
评论列表
文章目录
