堆叠多层bi-lstm的方法
自己在搭建多层双向lstm网络的时候，看了很多的资料，一开始的时候搭建都是不成功的，后来看了官方的资料还有一些博客，使用了3种方法搭建成功了！
方法1：
n_hidden_units=50 # 隐藏层神经元数目
num_layers=3 # 双向lstm神经网络的层数
n_steps=15
n_inputs=32
X=tf.placeholder(dtype=tf.float32,[None, n_steps,n_inputs]
def lstm_cell_fw(): # 前向的lstm网络
return tf.contrib.rnn.BasicLSTMCell(n_hidden_units,forget_bias=1.0,state_is_tuple=True)
def lstm_cell_bw(): # 反向的lstm网络
return tf.contrib.rnn.BasicLSTMCell(n_hidden_units,forget_bias=1.0,state_is_tuple=True)
# 堆叠网络
stacked_lstm_fw=tf.contrib.rnn.MultiRNNCell([lstm_cell_fw() for _ in range(num_layers)])
stacked_lstm_bw=tf.contrib.rnn.MultiRNNCell([lstm_cell_bw() for _ in range(num_layers)])
# 输出
outputs,_=tf.nn.bidirectional_dynamic_rnn(stacked_lstm_fw,stacked_lstm_bw,X,dtyep=tf.float32,time_major=False)
方法2：
单独的定义每一层中的网络，同样还是以3层双向lstm为例，假设每一层双向lstm中隐层的单元状态为50：`
hidden_units=[50,50,50]
num_layers=3
n_steps=15
n_inputs=32
X=tf.placeholder(dtype=tf.float32,[None, n_steps,n_inputs]
# 前向网络
single_cell_fw=[tf.nn.rnn_cell.LSTMCell(size) for size in hidden_units]
cell_fw=tf.nn.rnn_cell.MultiRNNCell(single_cell_fw)
# 反向网络
single_cell_bw=[tf.nn.rnn_cell.LSTMCell(size) for size in hidden_units]
cell_bw=tf.nn.rnn_cell.MultiRNNCell(single_cell_bw)
outputs,_=tf.nn.bidirectional_dynamic_rnn(cell_fw,cell_bw,X,dtype=tf.float32,time_major=False)
方法3：
n_hidden_units=50 # 隐藏层神经元数目
num_layers=3 # 双向lstm神经网络的层数
n_steps=15
n_inputs=32
X=tf.placeholder(dtype=tf.float32,[None, n_steps,n_inputs]
# 定义前向网络
lstm_cell_fw=tf.contrib.rnn.BasicLSTMCell(n_hidden_units,forget_bias=1.0,state_is_tuple=True)
lstm_cell_fw1=tf.contrib.rnn.BasicLSTMCell(n_hidden_units,forget_bias=1.0,state_is_tuple=True)
lstm_cell_bw=tf.contrib.rnn.BasicLSTMCell(n_hidden_units,forget_bias=1.0,state_is_tuple=True)
lstm_cell_bw1=tf.contrib.rnn.BasicLSTMCell(n_hidden_units,forget_bias=1.0,state_is_tuple=True)
fw_stack_rnn=[lstm_cell_fw1]
bw_stack_rnn=[lstm_cell_bw1]
for i in range(num_layers):
fw_stack_rnn.append(lstm_cell_fw)
bw_stack_rnn.append(lstm_cell_bw)
# 输出
outputs,_=tf.nn.bidirectional_dyanmic_rnn(fw_stack_rnn,bw_stack_rnn,X,dtype=tf.float32,time_major=False)
注意一点：在自己使用cell_fw=tf.nn.rnn_cell.MultiRNNCell([lstm_cell_fw() * num_layers,state_is_tuple=True)的时候，编译器总是报cell_fw不是一个instance RNNCell的错误，要求输入的得是instance RNNCell，对于此种情况,我使用了上述的3中堆叠方式，网络都搭建成功了！