def get_column_hidden_vectors(self):
#vector representations for the column names
self.column_hidden_vectors = tf.reduce_sum(
nn_utils.get_embedding(self.batch_number_column_names,
self.utility, self.params), 2)
self.word_column_hidden_vectors = tf.reduce_sum(
nn_utils.get_embedding(self.batch_word_column_names, self.utility,
self.params), 2)
def get_column_hidden_vectors(self):
#vector representations for the column names
self.column_hidden_vectors = tf.reduce_sum(
nn_utils.get_embedding(self.batch_number_column_names, self.utility,
self.params), 2)
self.word_column_hidden_vectors = tf.reduce_sum(
nn_utils.get_embedding(self.batch_word_column_names, self.utility,
self.params), 2)
def LSTM_question_embedding(self, sentence, sentence_length):
#LSTM processes the input question
lstm_params = "question_lstm"
hidden_vectors = []
sentence = self.batch_question
question_hidden = tf.zeros(
[self.batch_size, self.utility.FLAGS.embedding_dims],
self.data_type)
question_c_hidden = tf.zeros(
[self.batch_size, self.utility.FLAGS.embedding_dims],
self.data_type)
if (self.utility.FLAGS.rnn_dropout > 0.0):
if (self.mode == "train"):
rnn_dropout_mask = tf.cast(
tf.random_uniform(
tf.shape(question_hidden), minval=0.0, maxval=1.0) <
self.utility.FLAGS.rnn_dropout,
self.data_type) / self.utility.FLAGS.rnn_dropout
else:
rnn_dropout_mask = tf.ones_like(question_hidden)
for question_iterator in range(self.question_length):
curr_word = sentence[:, question_iterator]
question_vector = nn_utils.apply_dropout(
nn_utils.get_embedding(curr_word, self.utility, self.params),
self.utility.FLAGS.dropout, self.mode)
question_hidden, question_c_hidden = nn_utils.LSTMCell(
question_vector, question_hidden, question_c_hidden,
lstm_params, self.params)
if (self.utility.FLAGS.rnn_dropout > 0.0):
question_hidden = question_hidden * rnn_dropout_mask
hidden_vectors.append(tf.expand_dims(question_hidden, 0))
hidden_vectors = tf.concat(axis=0, values=hidden_vectors)
return question_hidden, hidden_vectors
def LSTM_question_embedding(self, sentence, sentence_length):
#LSTM processes the input question
lstm_params = "question_lstm"
hidden_vectors = []
sentence = self.batch_question
question_hidden = tf.zeros(
[self.batch_size, self.utility.FLAGS.embedding_dims], self.data_type)
question_c_hidden = tf.zeros(
[self.batch_size, self.utility.FLAGS.embedding_dims], self.data_type)
if (self.utility.FLAGS.rnn_dropout > 0.0):
if (self.mode == "train"):
rnn_dropout_mask = tf.cast(
tf.random_uniform(
tf.shape(question_hidden), minval=0.0, maxval=1.0) <
self.utility.FLAGS.rnn_dropout,
self.data_type) / self.utility.FLAGS.rnn_dropout
else:
rnn_dropout_mask = tf.ones_like(question_hidden)
for question_iterator in range(self.question_length):
curr_word = sentence[:, question_iterator]
question_vector = nn_utils.apply_dropout(
nn_utils.get_embedding(curr_word, self.utility, self.params),
self.utility.FLAGS.dropout, self.mode)
question_hidden, question_c_hidden = nn_utils.LSTMCell(
question_vector, question_hidden, question_c_hidden, lstm_params,
self.params)
if (self.utility.FLAGS.rnn_dropout > 0.0):
question_hidden = question_hidden * rnn_dropout_mask
hidden_vectors.append(tf.expand_dims(question_hidden, 0))
hidden_vectors = tf.concat(axis=0, values=hidden_vectors)
return question_hidden, hidden_vectors
def get_column_hidden_vectors(self):
#vector representations for the column names
number_column_lengths = tf.reshape(
self.batch_number_column_name_lengths,
[self.batch_size, self.num_cols, 1])
word_column_lengths = tf.reshape(
self.batch_word_column_name_lengths,
[self.batch_size, self.num_word_cols, 1])
number_column_embeddings = nn_utils.get_embedding(
self.batch_number_column_names, self.utility, self.params,
self.batch_number_column_name_mask)
word_column_embeddings = nn_utils.get_embedding(
self.batch_word_column_names, self.utility, self.params,
self.batch_word_column_name_mask)
number_column_description_lengths = tf.reshape(
self.batch_number_column_description_lengths,
[self.batch_size, self.num_cols, 1])
word_column_description_lengths = tf.reshape(
self.batch_word_column_description_lengths,
[self.batch_size, self.num_word_cols, 1])
number_column_description_embeddings = nn_utils.get_embedding(
self.batch_number_column_descriptions, self.utility, self.params,
self.batch_number_column_description_mask)
word_column_description_embeddings = nn_utils.get_embedding(
self.batch_word_column_descriptions, self.utility, self.params,
self.batch_word_column_description_mask)
number_column_sum = tf.reduce_sum(number_column_embeddings, 2)
word_column_sum = tf.reduce_sum(word_column_embeddings, 2)
number_column_description_sum = tf.reduce_sum(
number_column_description_embeddings, 2)
word_column_description_sum = tf.reduce_sum(
word_column_description_embeddings, 2)
self.column_hidden_vectors = tf.divide(
number_column_sum, number_column_lengths) + tf.divide(
number_column_description_sum,
number_column_description_lengths)
self.word_column_hidden_vectors = tf.divide(
word_column_sum, word_column_lengths) + tf.divide(
word_column_description_sum, word_column_description_lengths)
def create_summary_embeddings(self):
#embeddings for each text entry in the table using parameters of the question module
self.summary_text_entry_embeddings = tf.reduce_sum(
tf.expand_dims(self.batch_exact_match, 3) * tf.expand_dims(
tf.expand_dims(
tf.expand_dims(
nn_utils.get_embedding(
self.utility.entry_match_token_id, self.utility,
self.params), 0), 1), 2), 2)
def create_summary_embeddings(self):
#embeddings for each text entry in the table using parameters of the question module
self.summary_text_entry_embeddings = tf.reduce_sum(
tf.expand_dims(self.batch_exact_match, 3) * tf.expand_dims(
tf.expand_dims(
tf.expand_dims(
nn_utils.get_embedding(self.utility.entry_match_token_id,
self.utility, self.params), 0), 1),
2), 2)
