def bilstm_layer(self, lstm_inputs):
"""
BiLSTM层
:param lstm_inputs: [batch_size, num_steps, emb_size]
:return: [batch_size, num_steps, 2*lstm_dim]
"""
lstm_inputs = tf.layers.dense(lstm_inputs,
self.lstm_dim,
use_bias=True)
lstm_inputs = tf.nn.relu(lstm_inputs)
with tf.variable_scope("char_BiLSTM"):
lstm_cell = {}
for direction in ["forward", "backward"]:
with tf.variable_scope(direction):
lstm_cell[direction] = rnn.CoupledInputForgetGateLSTMCell(
self.lstm_dim, # 每个LSTM cell内部的神经元数量（即隐层参数维度）
use_peepholes=True,
initializer=self.initializer,
state_is_tuple=True)
outputs, final_states = tf.nn.bidirectional_dynamic_rnn(
lstm_cell["forward"], # 前向传播cell
lstm_cell["backward"], # 后向传播cell
lstm_inputs,
dtype=tf.float32,
sequence_length=self.lengths,
time_major=False)
output_forward, output_backward = tf.unstack(outputs, axis=0)
output_h = tf.add(output_forward, output_backward)
mix_input = tf.add(tf.nn.relu(output_h), lstm_inputs)
with tf.variable_scope("char_BiLSTM_2"):
lstm_cell = {}
for direction in ["forward", "backward"]:
with tf.variable_scope(direction):
lstm_cell[direction] = rnn.CoupledInputForgetGateLSTMCell(
self.lstm_dim, # 每个LSTM cell内部的神经元数量（即隐层参数维度）
use_peepholes=True,
initializer=self.initializer,
state_is_tuple=True)
outputs, final_states = tf.nn.bidirectional_dynamic_rnn(
lstm_cell["forward"], # 前向传播cell
lstm_cell["backward"], # 后向传播cell
mix_input,
dtype=tf.float32,
sequence_length=self.lengths,
time_major=False)
output_forward, output_backward = tf.unstack(outputs, axis=0)
output_h = tf.add(output_forward, output_backward)
mix_input = tf.add(tf.nn.relu(mix_input), tf.nn.relu(output_h))
return tf.concat([output_h, mix_input], axis=2)
def bi_lstm_layer(lstm_inputs, lstm_dim):
"""
该函数负责将句子用双向LSTM编码，取最后的句子编码
lstm_inputs: [batch_size, num_steps, emb_size]
lstm_dim 由于输出结果维度中最后一维是2*lstm_dim，而我们需要的是
[batch_size, emb_size]
那么根据这个关系我们可以确定我们需要的lstm_dim会是embedding_size/2
return: [batch_size, 2*lstm_dim]
"""
with tf.variable_scope("bi_lstm"):
lstm_cell = {}
for direction in ["forward", "backward"]:
with tf.variable_scope(direction):
lstm_cell[direction] = rnn.CoupledInputForgetGateLSTMCell(
lstm_dim,
use_peepholes=True,
initializer=tf.random_normal_initializer(stddev=0.1),
state_is_tuple=True)
outputs, final_states = tf.nn.bidirectional_dynamic_rnn(
lstm_cell["forward"],
lstm_cell["backward"],
lstm_inputs,
dtype=tf.float32)
# 取出output最后一个时刻句子的embedding结果作为后续输入
out = tf.concat(outputs, axis=2)
res = out[:, -1, :]
return res
def biLSTM_layer(self, model_inputs, lstm_dim, lengths, name=None):
"""
:param model_inputs: [batch_size, num_steps, emb_size]
:param lengths:
:param name:
:return: [batch_size, num_step, 2*lstm_dim]
"""
with tf.variable_scope("char_BiLSTM" if not name else name):
lstm_cell = {}
for direction in ["forward", "backward"]:
with tf.variable_scope(direction):
lstm_cell[direction] = rnn.CoupledInputForgetGateLSTMCell(
lstm_dim,
use_peepholes=True,
initializer=self.initializer,
state_is_tuple=True
)
outputs, final_statues = tf.nn.bidirectional_dynamic_rnn(
lstm_cell["forward"],
lstm_cell["backward"],
model_inputs,
dtype=tf.float32,
sequence_length=lengths
)
return tf.concat(outputs, axis=2)
def biLSTM_layer(self, lstm_inputs, lstm_dim, lengths, name=None):
with tf.variable_scope('char_lstm' if not name else name):
lstm_cell = {}
for direction in ['forward', 'backward']:
with tf.variable_scope(direction):
lstm_cell[direction] = rnn.CoupledInputForgetGateLSTMCell(
lstm_dim,
use_peepholes=True,
initializer=self.initializer,
state_is_tuple=True)
outputs, final_states = tf.nn.bidirectional_dynamic_rnn(
lstm_cell['forward'],
lstm_cell('backward'),
lstm_inputs,
dtype=tf.float32,
sequence_length=lengths
)
return tf.concat(outputs, axis=2)
def bilstm_layer(self, lstm_inputs):
"""
BiLSTM层
:param lstm_inputs: [batch_size, num_steps, emb_size]
:return: [batch_size, num_steps, 2*lstm_dim]
"""
with tf.variable_scope("char_BiLSTM"):
lstm_cell = {}
for direction in ["forward", "backward"]:
with tf.variable_scope(direction):
lstm_cell[direction] = rnn.CoupledInputForgetGateLSTMCell(
self.lstm_dim, # 每个LSTM cell内部的神经元数量（即隐层参数维度）
use_peepholes=True, initializer=self.initializer, state_is_tuple=True)
outputs, final_states = tf.nn.bidirectional_dynamic_rnn(
lstm_cell["forward"], # 前向传播cell
lstm_cell["backward"], # 后向传播cell
lstm_inputs,
dtype=tf.float32,
sequence_length=self.lengths)
return tf.concat(outputs, axis=2)
def biLSTMLayer(self, lstmInputs, lstmDim, lengths, name=None):
"""
:param lstmInputs: [batchSize, numSteps, emb_size]
:return: [batchSize, numSteps, 2*lstmDim]
"""
with tf.variable_scope("char_BiLSTM" if not name else name):
lstmCell = {}
for direction in ["forward", "backward"]:
with tf.variable_scope(direction):
lstmCell[direction] = rnn.CoupledInputForgetGateLSTMCell(
lstmDim,
usePeepholes=True,
initializer=self.initializer,
stateIsTuple=True)
outputs, finalStates = tf.nn.bidirectional_dynamic_rnn(
lstmCell["forward"],
lstmCell["backward"],
lstmInputs,
dtype=tf.float32,
sequence_length=lengths)
return tf.concat(outputs, axis=2)
def biLSTM_layer(self, lstm_inputs, lstm_dim, lengths, name=None):
"""
:param lstm_inputs: [batch_size, num_steps, emb_size] num_steps = MaxLenSentence
:return: [batch_size, num_steps, 2*lstm_dim]
"""
with tf.variable_scope("char_BiLSTM" if not name else name):
lstm_cell = {} # 定义LSTM单元
for direction in ["forward", "backward"]:
with tf.variable_scope(direction):
lstm_cell[direction] = rnn.CoupledInputForgetGateLSTMCell(
lstm_dim, # 输出结果维度
use_peepholes=True,
initializer=self.initializer,
state_is_tuple=True) # 若为真则输出和状态分2列输出，若为假则输出和状态在一列中输出
outputs, final_states = tf.nn.bidirectional_dynamic_rnn( # 定义Bi-LSTM网络
lstm_cell["forward"],
lstm_cell["backward"],
lstm_inputs,
dtype=tf.float32,
sequence_length=lengths) # sequence_length为BatchSize
return tf.concat(outputs, axis=2)
def outadaptation_layer(self, lstm_inputs, lstm_dim, lengths, name=None):
"""
:param lstm_inputs: [batch_size, num_steps, emb_size]
:return: [batch_size, num_steps, 2*lstm_dim]
"""
with tf.variable_scope("outadaptation" if not name else name) as scope:
lstm_cell = {}
for direction in ["forward", "backward"]:
with tf.variable_scope(direction, reuse=True):
lstm_cell[direction] = rnn.CoupledInputForgetGateLSTMCell(
lstm_dim,
use_peepholes=True,
initializer=self.initializer,
state_is_tuple=True)
outputs, final_states = tf.nn.bidirectional_dynamic_rnn(
lstm_cell["forward"],
lstm_cell["backward"],
lstm_inputs,
dtype=tf.float32,
sequence_length=lengths)
hidden_lstm = tf.concat(outputs, axis=2)
return hidden_lstm
def biLSTM_layer_2(inputs, lstm_dim, lengths, name=None):
"""
:param lstm_inputs: [batch_size, num_steps, emb_size]
:return: [batch_size, num_steps, 2*lstm_dim]
"""
initializer = initializers.xavier_initializer()
# sess = tf.Session()
# sess.run(tf.global_variables_initializer())
with tf.variable_scope("BiLSTM" if not name else name):
lstm_cell = {}
for direction in ["forward", "backward"]:
with tf.variable_scope(direction):
lstm_cell[direction] = rnn.CoupledInputForgetGateLSTMCell(
lstm_dim,
use_peepholes=True,
initializer=initializer,
state_is_tuple=True)
outputs, final_states = tf.nn.bidirectional_dynamic_rnn(
lstm_cell["forward"],
lstm_cell["backward"],
inputs,
dtype=tf.float32,
sequence_length=lengths)
return tf.concat(outputs, axis=2)