Tensorflow - dynamic_rnn 学习

函数目的就是为了构建一个RNN网络，前面少不了定义cell的类型，如LSTM与GRU等

API 里面 dynamic_rnn 函数的参数如下：

tf.nn.dynamic_rnn(
    cell,
    inputs,
    sequence_length=None,
    initial_state=None,
    dtype=None,
    parallel_iterations=None,
    swap_memory=False,
    time_major=False,
    scope=None
)

例子：

# create a BasicRNNCell
rnn_cell = tf.nn.rnn_cell.BasicRNNCell(hidden_size)

# 'outputs' is a tensor of shape [batch_size, max_time, cell_state_size]

# defining initial state
initial_state = rnn_cell.zero_state(batch_size, dtype=tf.float32)

# 'state' is a tensor of shape [batch_size, cell_state_size]
outputs, state = tf.nn.dynamic_rnn(rnn_cell, input_data,
                                   initial_state=initial_state,
                                   dtype=tf.float32)

 多层网络：

这个定义的方法值得学习一个，因为这样定义可以避免维度错误，主要原因是TensorFlow的版本问题，可以构建一个多层网络，而且embedding size 和 hidden size的维度可以不同。

# create 2 LSTMCells
rnn_layers = [tf.nn.rnn_cell.LSTMCell(size) for size in [128, 256]]

# create a RNN cell composed sequentially of a number of RNNCells
multi_rnn_cell = tf.nn.rnn_cell.MultiRNNCell(rnn_layers)

# 'outputs' is a tensor of shape [batch_size, max_time, 256]
# 'state' is a N-tuple where N is the number of LSTMCells containing a
# tf.contrib.rnn.LSTMStateTuple for each cell
outputs, state = tf.nn.dynamic_rnn(cell=multi_rnn_cell,
                                   inputs=data,
                                   dtype=tf.float32)

参数 ：

cell : RNN单元的一个实例如GRU LSTM等

inputs：[batch_size,max_time] 这两个参数比如有，后面可以是嵌套的数据

sequence_length :可选，int32 或 int64

initial_state:可选，初始状态

其余的省略了 一般也不用

参考https://tensorflow.google.cn/api_docs/python/tf/nn/dynamic_rnn?hl=zh-cn

返回值

outouts [batch_size, max_time, cell.output_size] 这样的形状，output_size = hidden size

state 末态，形状为[batch_size, cell.state_size]