def biLSTM_layer(self, lstm_inputs, lstm_dim, lengths, name=None):
"""
:param lstm_inputs: [batch_size, num_steps, emb_size]
:return: [batch_size, 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,
(encoder_fw_final_state,
encoder_bw_final_state)) = tf.nn.bidirectional_dynamic_rnn(
lstm_cell["forward"],
lstm_cell["backward"],
lstm_inputs,
dtype=tf.float32,
sequence_length=lengths)
final_state = tf.concat(
(encoder_fw_final_state.h, encoder_bw_final_state.h), -1)
return final_state
def biLSTM_layer(self, lstm_inputs, lstm_dim, lengths, name=None):
"""
:param lstm_inputs: [batch_size, num_steps, emb_size]
:param lstm_dim:
:param name:
:return: [batch_size, num_steps, 2*lstm_dim]
"""
with tf.variable_scope("word_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_status = 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, lstm_dim, lengths, name=None):
"""
:param lstm_inputs: [batch_size, num_steps, emb_size]
:param lstm_dim:
:param name:
:return: [batch_size, num_steps, 2*lstm_dim]
为何返回是2*lstm_dim，因为其是双向的lstm。每个方向的输出为lstm_dim
"""
with tf.variable_scope("word_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_status = tf.nn.bidirectional_dynamic_rnn(
lstm_cell['forward'],
lstm_cell['backward'],
lstm_inputs,
dtype=tf.float32,
sequence_length=lengths)
# 因为单向的lstm输出的格式为[batch_size, num_steps,lstm_dim]。
# 2表示在lstm_dim这个维度进行拼接。
# 个人觉得outputs的输出格式为[[batch_size, num_steps,lstm_dim],[batch_size, num_steps,lstm_dim]]
# 即是一个list。list里面的每一个元素是单向的lstm的输出
return tf.concat(outputs, axis=2)
def MultiRNN(self, x, weight, bias, initializer, activation, num_layers):
x = tf.unstack(tf.reshape(
x, [tf.shape(x)[0], self.STEPS,
int(self.X_DIM / self.STEPS)]),
self.STEPS,
axis=1)
lstm_cell = rnn.CoupledInputForgetGateLSTMCell(self.CELL_SIZE,
forget_bias=10.0,
use_peepholes=True,
proj_clip=15.0,
initializer=initializer,
activation=activation)
lstm_cell = rnn.MultiRNNCell([lstm_cell] * num_layers)
outputs, _ = rnn.static_rnn(lstm_cell, x, dtype=tf.float32)
return tf.matmul(outputs[-1], weight) + bias, outputs
def lstm_layer():
# B-LSTM 构建双向一层的LSTM
# 定义两个LSTM网络
with tf.variable_scope('lstm_layer'):
lstm_cell = {}
for direction in ["forward", "backward"]:
with tf.variable_scope(direction):
lstm_cell[direction] = rnn.CoupledInputForgetGateLSTMCell(
FLAGS.lstm_size,
use_peepholes=True,
initializer=initializers.xavier_initializer(),
state_is_tuple=True)
outputs, final_states = tf.nn.bidirectional_dynamic_rnn(
lstm_cell["forward"],
lstm_cell["backward"],
embedding,
dtype=tf.float32,
sequence_length=lengths)
return tf.concat(outputs, axis=2)
def biLSTM_layer(self):
"""
:param lstm_inputs: [batch_size, num_steps, emb_size]
:return: [batch_size, num_steps, 2*lstm_dim]
"""
with tf.variable_scope("BiLSTM"):
lstm_cell = {}
for direction in ["forward", "backward"]:
with tf.variable_scope(direction):
lstm_cell[direction] = rnn.CoupledInputForgetGateLSTMCell(
self.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"],
self.embedding,
dtype=tf.float32,
sequence_length=self.lengths)
return tf.concat(outputs, axis=2)
def biLSTM_layer(self, lstm_inputs, lstm_dim, lengths, name=None):
with tf.variable_scope("char_BiLSTM", reuse=tf.AUTO_REUSE):
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,
(encoder_fw_final_state,
encoder_bw_final_state)) = tf.nn.bidirectional_dynamic_rnn(
lstm_cell["forward"],
lstm_cell["backward"],
lstm_inputs,
dtype=tf.float32,
sequence_length=lengths)
final_state = tf.concat(
(encoder_fw_final_state.h, encoder_bw_final_state.h), -1)
return tf.concat(outputs, axis=2), final_state
def biLSTM_layer(self, lstm_inputs, lstm_dim, lengths, name=None):
with tf.variable_scope('char_BiLSTM', reuse=tf.AUTO_REUSE):
lstm_cell = {}
for direction in ["forward", "backward"]:
with tf.variable_scope(direction):
lstm_cell[direction] = rnn.CoupledInputForgetGateLSTMCell(
lstm_dim,
use_peepholes=True,
initializer=self.initiaizer,
state_is_tuple=True)
(outputs,
(encoder_fw_final_state,
encoder_bw_final_state)) = tf.nn.bidirectional_dynamic_rnn(
lstm_cell["forward"],
lstm_cell["backward"],
lstm_inputs,
dtype=tf.float32,
sequence_length=lengths)
# 每句话经过当前cell后会得到一个state，
# 经过多少个cell，就有多少个LSTMStateTuple，即每个cell都会输出一个 tuple(c, h)
# state中的h跟output 的最后一个时刻的输出是一样的，即：output[:, -1, :] = state[0].h
final_state = tf.concat(
(encoder_fw_final_state.h, encoder_bw_final_state.h), -1)
return tf.concat(outputs, -1), final_state
def basic_cell(hidden_size, initializer):
_cell = rnn.CoupledInputForgetGateLSTMCell(
hidden_size,
use_peepholes=True,
initializer=initializer)
return _cell
