tf.contrib.rnn.BasicLSTMCell 是单个 LSTM 单元还是 LSTM 层? [英] Is tf.contrib.rnn.BasicLSTMCell a single LSTM unit or a LSTM layer?

问题描述

在张量流中，有一个名为 BasicLSTMCell 的 lstm 实现，它位于 tf.contrib.rnn.BasicLSTMCell.它有一个参数 num_units 表示 LSTM 单元中的单元数.但我不知道那是什么意思.

in the tensorflow, there is a lstm implementation called BasicLSTMCell which at tf.contrib.rnn.BasicLSTMCell. And it has a parameter num_units which means the number of units in the LSTM cell. But I do not know what that means.

如果我像这样定义一个 lstm 单元格:

If I define a lstm cell like this:

lstm_cell = tf.contrib.rnn.BasicLSTMCell(512).

lstm_cell 是什么样的?是lstm节点还是512节点的lstm层??谁能告诉我这个?

what does the lstm_cell look like? It is a lstm node or a lstm layer with 512 node??Who can tell me about this ?

推荐答案

它是一个 LSTM 层，有 512 个单元.

It is an LSTM layer with` 512 units.

BasicLSTMCell 实现抽象类 RNNCell.来自文档:

BasicLSTMCell implements the abstract class RNNCell. From the documentation:

代表 RNN 单元格的抽象对象.

Abstract object representing an RNN cell.

每个 RNNCell 都必须具有以下属性，并使用签名 (output, next_state) = call(input, state) 实现 call.

Every RNNCell must have the properties below and implement call with the signature (output, next_state) = call(input, state).

[...]

这个单元格的定义与文献中使用的定义不同.在文献中，单元格"是指具有单个标量输出的对象.此定义指的是此类单元的水平阵列.

This definition of cell differs from the definition used in the literature. In the literature, 'cell' refers to an object with a single scalar output. This definition refers to a horizontal array of such units.

创建 LSTM 层并展开反向传播波谷时间的一种常用方法如下:

A common way of creating the LSTM layer together with the unrolling for Back Propagation Trough Time is the following one:

lstm_cell = tf.contrib.rnn.BasicLSTMCell(512)
outputs, final_state = tf.nn.static_rnn(cell=lstm_cell,
                           dtype=tf.float32,
                           inputs=some_input_sequence)

哪里:

• some_input_sequence 是 num_steps 个大小为 [batch_size, input_size] 的张量列表
• outputs 将包含 some_input_sequence 的每个元素之后的层的输出.所以它又是一个 num _steps 元素的列表，其大小为 [batch_size, 512](其中 512 是单元格的单位数)
• final_state 将包含处理整个序列后的状态.特别是，对于 LSTM，它是一个具有两个元素的命名元组，c 和 h(LSTM 的两个状态).

• some_input_sequence is a list of num_steps tensors of size [batch_size, input_size]
• outputs will contain the output of the layer after each of the elements of some_input_sequence. So it is again a list of num _steps elements of size [batch_size, 512] (where 512 was the number of units of your cell)
• final_state will contain the state after the entire sequence has been processed. In particular, for LSTM, it is a named tuple with two elements, c and h (the two states of a LSTM).

这篇关于tf.contrib.rnn.BasicLSTMCell 是单个 LSTM 单元还是 LSTM 层?的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！