def embed_text(tensors, embeddings):
"""Build embeddings using the word/char ids from `build_tensorize_text_fn`."""
wids = tensors["wids"]
cids = tensors["cids"]
embedding_weights = embeddings.get_initialized_params(trainable=False)
word_vecs = tf.nn.embedding_lookup(embedding_weights, wids)
char_emb = common_layers.character_cnn(cids)
return tf.concat([word_vecs, char_emb], -1)
def test_character_cnn(self):
with tf.Graph().as_default():
input_words = [["google", "lumiere"], [u"¯\\_(ツ)_/¯", u"(ᵔᴥᵔ)"],
[u"谷", u"歌"]]
char_ids = char_utils.batch_word_to_char_ids(
tf.constant(input_words), 10)
output_emb = common_layers.character_cnn(char_ids, num_filters=5)
with tf.Session() as sess:
sess.run(tf.compat.v1.global_variables_initializer())
actual_output_emb = sess.run(output_emb)
self.assertAllEqual(actual_output_emb.shape, [3, 2, 5])
def _embed(prefix):
"""Embed the input text based and word and character IDs."""
word_emb = tf.nn.embedding_lookup(embedding_weights,
features[prefix + "_wid"])
char_emb = common_layers.character_cnn(
char_ids=features[prefix + "_cid"],
emb_size=FLAGS.char_emb_size,
kernel_width=FLAGS.char_kernel_width,
num_filters=FLAGS.num_char_filters)
concat_emb = tf.concat([word_emb, char_emb], -1)
if mode == tf.estimator.ModeKeys.TRAIN:
concat_emb = tf.nn.dropout(concat_emb, 1.0 - FLAGS.dropout_ratio)
return concat_emb