def predict(self, left_slots, right_slots):
"""
predict graph of this net
"""
left = left_slots[self.left_name]
right = right_slots[self.right_name]
left_emb = self.emb_layer.ops(left) # (N, len, D)
right_emb = self.emb_layer.ops(right) # (N, len, D)
## left
bi_left_outputs, _ = tf.nn.bidirectional_dynamic_rnn(self.fw_cell, self.bw_cell, left_emb,
sequence_length=utility.seq_length(left), dtype=tf.float32)
left_seq_encoder = tf.concat(bi_left_outputs, -1)
## right
bi_right_outputs, _ = tf.nn.bidirectional_dynamic_rnn(self.fw_cell, self.bw_cell, right_emb,
sequence_length=utility.seq_length(right), dtype=tf.float32)
right_seq_encoder = tf.concat(bi_right_outputs, -1)
cross = tf.matmul(left_seq_encoder, tf.transpose(right_seq_encoder, [0, 2, 1])) # (N, len, len)
if self.match_mask:
cross_mask = utility.get_cross_mask(left, right)
else:
cross_mask = None
conv_pool = self.cnn_layer.ops(cross, mask=cross_mask)
pool_relu = self.relu_layer.ops(conv_pool)
relu_hid1 = self.fc1_layer.ops(pool_relu)
hid1_tanh = self.tanh_layer.ops(relu_hid1)
pred = self.fc2_layer.ops(hid1_tanh)
return pred
def predict(self, left_slots, right_slots):
"""
predict graph of this net
"""
left = left_slots[self.left_name]
right = right_slots[self.right_name]
left_emb = self.emb_layer.ops(left) # (N, len, D)
right_emb = self.emb_layer.ops(right) # (N, len, D)
## left
left_length = seq_length(left) #得到实际的长度
left_encoder = self.rnn.ops(left_emb, left_length)
left_rep = self.extract.ops(left_encoder, left_length) #获得最后的隐藏层
right_length = seq_length(right)
right_encoder = self.rnn.ops(right_emb, right_length)
right_rep = self.extract.ops(right_encoder, right_length)
if self.task_mode == "pointwise":
rep_concat = tf.concat([left_rep, right_rep], -1)
hidden1 = self.fc1_layer.ops(rep_concat)
hidden1_relu = self.relu_layer.ops(hidden1)
pred = self.fc2_layer.ops(hidden1_relu)
elif self.task_mode == "pairwise":
left_hidden1 = self.fc1_layer.ops(left_rep)
right_hidden1 = self.fc1_layer.ops(right_rep)
pred = self.cos_layer.ops(left_hidden1, right_hidden1)
return pred
def predict(self, left_slots, right_slots):
"""
predict graph of this net
"""
left = left_slots[self.left_name]
right = right_slots[self.right_name]
left_emb = self.emb_layer.ops(left) # (N, len, D)
right_emb = self.emb_layer.ops(right) # (N, len, D)
## left
left_length = seq_length(left)
left_encoder = self.rnn.ops(left_emb, left_length)
left_rep = self.extract.ops(left_encoder, left_length)
right_length = seq_length(right)
right_encoder = self.rnn.ops(right_emb, right_length)
right_rep = self.extract.ops(right_encoder, right_length)
rep_concat = tf.concat([left_rep, right_rep], -1)
hidden1 = self.fc1_layer.ops(rep_concat)
pred = self.fc2_layer.ops(hidden1)
return pred
def predict(self, left_slots, right_slots):
"""
predict graph of this net
"""
left = left_slots[self.left_name]
right = right_slots[self.right_name]
left_emb = self.emb_layer.ops(left) # (N, len, D)
right_emb = self.emb_layer.ops(right) # (N, len, D)
## left
bi_left_outputs, _ = tf.nn.bidirectional_dynamic_rnn(
self.fw_cell,
self.bw_cell,
left_emb,
sequence_length=seq_length(left),
dtype=tf.float32)
left_seq_encoder = tf.concat(bi_left_outputs, -1)
## right
bi_right_outputs, _ = tf.nn.bidirectional_dynamic_rnn(
self.fw_cell,
self.bw_cell,
right_emb,
sequence_length=seq_length(right),
dtype=tf.float32)
right_seq_encoder = tf.concat(bi_right_outputs, -1)
#进行交互运算
cross = tf.matmul(left_seq_encoder,
tf.transpose(right_seq_encoder,
[0, 2, 1])) # (N, len, len)
#if self.match_mask:
# cross_mask = get_cross_mask(left, right)
# cross = tf.multiply(cross, tf.cast(cross_mask, tf.float32))
# #paddings = tf.ones_like(cross)*(-2**32+1)
# #cross = tf.where(tf.equal(cross_mask, 0), paddings, cross)
cross_reshape = tf.reshape(cross, [-1, self.seq_len1 * self.seq_len2])
#返回values，indices
k_max_match = tf.nn.top_k(cross_reshape, k=self.k_max_num,
sorted=True)[0]
#k_max_match = tf.Print(k_max_match, [k_max_match], "k_max_match")
pred = self.fc2_layer.ops(k_max_match)
return pred