RNN网络搭建

#rnn网络单元
tf.nn.rnn_cell
tf.keras.layers.RNN

tf.nn.rnn_cell.BasicRNNCell#最基本的rnn单元
(
    num_units,
    activation=None,#默认tanh，使用tf.nn.relu效果更好
    reuse=None,
    name=None,
    dtype=None,
    **kwargs
)
tf.nn.rnn_cell.RNNCell#抽象rnn单元
(
    trainable=True,
    name=None,
    dtype=None,
    **kwargs
)
tf.nn.rnn_cell.LSTMCell#LSTM单元
(
    num_units,
    use_peepholes=False,#是否使用窥视孔
    cell_clip=None,#浮点值，对输出进行修剪
    initializer=None,
    num_proj=None,#整型，投影矩阵的输出维数
    proj_clip=None,
    num_unit_shards=None,#已被移除
    num_proj_shards=None,#已被移除
    forget_bias=1.0,
    state_is_tuple=True,
    activation=None,
    reuse=None,
    name=None,
    dtype=None,
    **kwargs
)
tf.nn.rnn_cell.GRUCell#GRU单元
(
    num_units,
    activation=None,
    reuse=None,
    kernel_initializer=None,
    bias_initializer=None,
    name=None,
    dtype=None,
    **kwargs
)

#多个单元
layers = [tf.nn.rnn_cell.GRUCell(size) for size in hidden_sizes]
cells = tf.nn.rnn_cell.MultiRNNCell(layers)

#搭建rnn
f.nn.dynamic_rnn
tf.nn.bidirectional_dynamic_rnn

#整体过程
hidden_sizes=[]
layers = [tf.nn.rnn_cell.GRUCell(size) for size in hidden_sizes]
cells = tf.nn.rnn_cell.MultiRNNCell(layers)
output, out_state = tf.nn.dynamic_rnn(cell, seq, length, initial_state)

#代码
import tensorflow as tf

batch_size=64
hidden_size=64
max_time=50
input_data=tf.placeholder(tf.float32,shape=(batch_size,max_time,64))
# 创建单元
rnn_cell = tf.nn.rnn_cell.BasicRNNCell(hidden_size)

# 初始化
initial_state = rnn_cell.zero_state(batch_size, dtype=tf.float32)

# 'state'：[batch_size, cell_state_size]
# ‘outputs’：[batch_size, max_time, cell_state_size]
outputs, state = tf.nn.dynamic_rnn(rnn_cell, input_data,
                                   initial_state=initial_state,
                                   dtype=tf.float32)