def doc_embbedding(self, sent_topic_emb, seqLen, scope=None):
with tf.variable_scope(scope or "sent_topic_embedding"):
if self.cfg.doc_encode == "bigru":
birnn_sent = biGRU(sent_topic_emb, self.cfg.sent_hidden, seqLen)
elif self.cfg.doc_encode == 'bilstm':
birnn_sent = biLSTM(sent_topic_emb, self.cfg.sent_hidden, seqLen)
else:
raise ValueError("no such encoder %s" %(self.cfg.doc_encode))
doc_emb = mask_attention(birnn_sent, self.cfg.atten_size, seqLen) # (b_sz, sent_hidden * 2)
return doc_emb
def sent_encode(self, user_word_emb, scope=None):
with tf.variable_scope(scope or "sent_encode"):
if self.cfg.lstm_sent_encode == 'bigru':
sent_emb = biGRU(user_word_emb, self.cfg.word_hidden, self.ph_sNum)
sent_emb = avg_pooling(sent_emb, self.cfg.maxseq, self.cfg.word_hidden * 2)
elif self.cfg.lstm_sent_encode == 'bilstm':
sent_emb = biLSTM(user_word_emb, self.cfg.word_hidden, self.ph_sNum)
sent_emb = avg_pooling(sent_emb, self.cfg.maxseq, self.cfg.word_hidden * 2)
else:
raise ValueError("no such sent encode %s" % (self.cfg.lstm_sent_encode))
return sent_emb
def topic_embedding(self, topic_word_emb, scope=None):
# topic_word_emb = tf.expand_dims(topic_word_emb, axis=0)
with tf.variable_scope(scope or "topic_encode"):
if self.cfg.topic_encode == "bigru":
topic_seqLen = get_length(topic_word_emb)
topic_seqLen = tf.reshape(topic_seqLen, [-1, ])
birnn_x = biGRU(topic_word_emb, self.cfg.word_hidden, topic_seqLen)
elif self.cfg.topic_encode == 'bilstm':
topic_seqLen = get_length(topic_word_emb)
topic_seqLen = tf.reshape(topic_seqLen, [-1, ])
birnn_x = biLSTM(topic_word_emb, self.cfg.word_hidden, topic_seqLen)
else:
raise ValueError("no such encoder %s" %(self.cfg.rnn_encode))
topic_emb = avg_pooling(birnn_x, self.cfg.topic_num, self.cfg.word_hidden * 2) # shape: (1, topic, word_hidden)
# topic_emb = mask_attention(birnn_x, self.cfg.topic_num, self.cfg.word_hidden * 2, self.cfg.atten_size, topic_seqLen) # sequence_length, dim, attn_size, seqLen
topic_emb = tf.reshape(topic_emb, [self.cfg.topic_num, self.cfg.word_hidden * 2]) # shape: (topic_num, word_hidden)
return topic_emb
def sent_encode(self, user_word_emb, scope=None):
# b_sz, ststp, wtstp, emb_sz = tf.unstack(tf.shape(user_word_emb))
with tf.variable_scope(scope or "sent_encode"):
sent_word = tf.reshape(user_word_emb, [-1, self.cfg.maxword, self.cfg.emb_size])
sent_wNUm = tf.reshape(self.ph_wNum, [-1, ])
if self.cfg.ute_sent_encode == "cnn":
sent_emb = cnn_layer(sent_word, self.cfg.emb_size, filter_sizes=self.cfg.filter_sizes, num_filters=self.cfg.num_filters)
elif self.cfg.ute_sent_encode == 'bigru':
sent_emb = biGRU(sent_word,self.cfg.word_hidden, sent_wNUm)
sent_emb = mask_attention(sent_emb, self.cfg.maxword, self.cfg.word_hidden * 2, self.cfg.atten_size, sent_wNUm)
elif self.cfg.ute_sent_encode == 'bilstm':
sent_emb = biLSTM(sent_word, self.cfg.word_hidden, sent_wNUm)
sent_emb = mask_attention(sent_emb, self.cfg.maxword, self.cfg.word_hidden * 2, self.cfg.atten_size, sent_wNUm)
else:
raise ValueError("no such sent encode %s" % (self.cfg.ute_sent_encode))
return sent_emb
def sent_encode(self, user_word_emb, scope=None):
with tf.variable_scope(scope or "sent_encode"):
sent_word = tf.reshape(user_word_emb,
[-1, self.cfg.maxword, self.cfg.emb_size])
sent_wNUm = tf.reshape(self.ph_wNum, [
-1,
])
if self.cfg.han_sent_encode == 'bigru':
sent_emb = biGRU(sent_word, self.cfg.word_hidden, sent_wNUm)
elif self.cfg.han_sent_encode == 'bilstm':
sent_emb = biLSTM(sent_word, self.cfg.word_hidden, sent_wNUm)
else:
raise ValueError("no such sent encode %s" %
(self.cfg.han_sent_encode))
sent_emb = mask_attention(sent_emb, self.cfg.maxword,
self.cfg.word_hidden * 2,
self.cfg.atten_size,
sent_wNUm) # attention mechanism
return sent_emb