Python tf.compat.v1.nn.dynamic_rnn用法及代码示例

创建由 RNNCell cell 指定的循环神经网络。 (已弃用)

用法

tf.compat.v1.nn.dynamic_rnn(
    cell, inputs, sequence_length=None, initial_state=None, dtype=None,
    parallel_iterations=None, swap_memory=False, time_major=False, scope=None
)

迁移到 TF2

警告：这个 API 是为 TensorFlow v1 设计的。继续阅读有关如何从该 API 迁移到本机 TensorFlow v2 等效项的详细信息。见TensorFlow v1 到 TensorFlow v2 迁移指南有关如何迁移其余代码的说明。

tf.compat.v1.nn.dynamic_rnn 与即刻执行和 tf.function 不兼容。请改用tf.keras.layers.RNN 进行 TF2 迁移。以 LSTM 为例，您可以使用 tf.keras.layers.LSTMCell 或直接通过 tf.keras.layers.LSTM 实例化 tf.keras.layers.RNN 层。创建 keras 层后，您可以通过调用具有输入和状态的层来获取输出和状态。有关 Keras RNN 的更多详细信息，请参阅本指南。您还可以在本文档中找到更多关于 Keras RNN 和 TF compat v1 rnn 的区别和比较的详细信息

到原生 TF2 的结构映射

前：

# create 2 LSTMCells
rnn_layers = [tf.compat.v1.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.compat.v1.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.nn.rnn_cell.LSTMStateTuple for each cell
outputs, state = tf.compat.v1.nn.dynamic_rnn(cell=multi_rnn_cell,
                                             inputs=data,
                                             dtype=tf.float32)

后：

# RNN layer can take a list of cells, which will then stack them together.
# By default, keras RNN will only return the last timestep output and will not
# return states. If you need whole time sequence output as well as the states,
# you can set `return_sequences` and `return_state` to True.
rnn_layer = tf.keras.layers.RNN([tf.keras.layers.LSTMCell(128),
                                 tf.keras.layers.LSTMCell(256)],
                                return_sequences=True,
                                return_state=True)
outputs, output_states = rnn_layer(inputs, states)

如何映射参数

TF1 参数名称	TF2 参数名称	注意
cell	cell	在RNN层构造函数中
inputs	inputs	在RNN层__call__
sequence_length	未使用	在 RNN 之前添加遮罩层：以达到相同的结果。
initial_state	initial_state	在RNN层__call__
dtype	dtype	在RNN层构造函数中
parallel_iterations	不支持
swap_memory	不支持
time_major	time_major	在RNN层构造函数中
scope	不支持

警告：此函数已弃用。它将在未来的版本中删除。更新说明：请使用keras.layers.RNN(cell)，相当于这个API

执行 inputs 的完全动态展开。

例子：

# create a BasicRNNCell
rnn_cell = tf.compat.v1.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.compat.v1.nn.dynamic_rnn(rnn_cell, input_data,
                                   initial_state=initial_state,
                                   dtype=tf.float32)

# create 2 LSTMCells
rnn_layers = [tf.compat.v1.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.compat.v1.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.nn.rnn_cell.LSTMStateTuple for each cell
outputs, state = tf.compat.v1.nn.dynamic_rnn(cell=multi_rnn_cell,
                                   inputs=data,
                                   dtype=tf.float32)