sent1_mask = tf.cast(tf.sign(sentence1), dtype=tf.float32)
sent2_mask = tf.cast(tf.sign(sentence2), dtype=tf.float32)
sent1_len = tf.reduce_sum(sent1_mask, -1)
sent2_len = tf.reduce_sum(sent2_mask, -1)
antonym1 = tf.expand_dims(mnli.antonym1, -1)
antonym2 = tf.expand_dims(mnli.antonym2, -1)
exact1to2 = tf.expand_dims(mnli.exact1to2, -1)
exact2to1 = tf.expand_dims(mnli.exact2to1, -1)
synonym1 = tf.expand_dims(mnli.synonym1, -1)
synonym2 = tf.expand_dims(mnli.synonym2, -1)
sent1char = mnli.sent1char
sent2char = mnli.sent2char
with tf.variable_scope("word_embedding"):
glove_embedding = embedded(mnli.embedding)
embedding_pre = glove_embedding(sentence1)
embedding_hyp = glove_embedding(sentence2)
with tf.variable_scope("char_embedding"):
char_embedding = embedded(mnli.char_embedding, name="char")
char_embedding_pre = char_embedding(sent1char)
char_embedding_hyp = char_embedding(sent2char)
with tf.variable_scope("conv") as scope:
conv_pre = char_conv(char_embedding_pre, filter_size=filter_size)
scope.reuse_variables()
conv_hyp = char_conv(char_embedding_hyp, filter_size=filter_size)
embed_pre = tf.concat((embedding_pre, antonym1, exact1to2, synonym1, conv_pre),
-1)
sent1_mask = tf.cast(tf.sign(sentence1), dtype=tf.float32)
sent2_mask = tf.cast(tf.sign(sentence2), dtype=tf.float32)
sent1_len = tf.reduce_sum(sent1_mask, -1)
sent2_len = tf.reduce_sum(sent2_mask, -1)
antonym1 = tf.expand_dims(mnli.antonym1, -1)
antonym2 = tf.expand_dims(mnli.antonym2, -1)
exact1to2 = tf.expand_dims(mnli.exact1to2, -1)
exact2to1 = tf.expand_dims(mnli.exact2to1, -1)
synonym1 = tf.expand_dims(mnli.synonym1, -1)
synonym2 = tf.expand_dims(mnli.synonym2, -1)
sent1char = mnli.sent1char
sent2char = mnli.sent2char
with tf.variable_scope("word_embedding"):
glove_embedding = embedded(mnli.embedding)
embedding_pre = glove_embedding(sentence1)
embedding_hyp = glove_embedding(sentence2)
# with tf.variable_scope("char_embedding"):
# char_embedding = embedded(mnli.char_embedding, name="char")
# char_embedding_pre = char_embedding(sent1char)
# char_embedding_hyp = char_embedding(sent2char)
# with tf.variable_scope("conv") as scope:
# conv_pre = char_conv(char_embedding_pre)
# scope.reuse_variables()
# conv_hyp = char_conv(char_embedding_hyp)
# embed_pre = tf.concat((embedding_pre, antonym1, exact1to2, synonym1, conv_pre), -1)
# embed_hyp = tf.concat((embedding_hyp, antonym2, exact2to1, synonym2, conv_hyp), -1)