def save_word2vec_format(self, dest, source):
with tf.variable_scope('model', reuse=None):
if source is VocabType.Token:
vocab_size = self.word_vocab_size
embedding_size = self.config.EMBEDDINGS_SIZE
index = self.index_to_word
var_name = 'WORDS_VOCAB'
elif source is VocabType.Target:
vocab_size = self.target_word_vocab_size
embedding_size = self.config.EMBEDDINGS_SIZE * 3
index = self.index_to_target_word
var_name = 'TARGET_WORDS_VOCAB'
else:
raise ValueError(
'vocab type should be VocabType.Token or VocabType.Target.'
)
embeddings = tf.get_variable(var_name,
shape=(vocab_size + 1,
embedding_size),
dtype=tf.float32,
trainable=False)
self.saver = tf.train.Saver()
self.load_model(self.sess)
np_embeddings = self.sess.run(embeddings)
with open(dest, 'w') as words_file:
common.save_word2vec_file(words_file, vocab_size, embedding_size,
index, np_embeddings)
def save_word2vec_format(self, dest_save_path: str, vocab_type: VocabType):
if vocab_type not in VocabType:
raise ValueError('`vocab_type` should be `VocabType.Token`, `VocabType.Target` or `VocabType.Path`.')
vocab_embedding_matrix = self._get_vocab_embedding_as_np_array(vocab_type)
index_to_word = self.vocabs.get(vocab_type).index_to_word
with open(dest_save_path, 'w') as words_file:
common.save_word2vec_file(words_file, index_to_word, vocab_embedding_matrix)