[TensorFlow笔记] BasicRNNCell源码浅析

BasicRNNCell是抽象类RNNCell的一个最简单的实现。

class BasicRNNCell(RNNCell):

    def __init__(self, num_units, activation=None, reuse=None):
        super(BasicRNNCell, self).__init__(_reuse=reuse)
        self._num_units = num_units
        self._activation = activation or math_ops.tanh
        self._linear = None

    @property
    def state_size(self):
        return self._num_units

    @property
    def output_size(self):
        return self._num_units

    def call(self, inputs, state):
        if self._linear is None:
            self._linear = _Linear([inputs, state], self._num_units, True)

        output = self._activation(self._linear([inputs, state]))
        return output, output

实现了下面的运算：

用公式表示就是：

$$h_t=tanh(W_k[ x_t, h_{t-1}]+b)$$

有时这个公式会写成下面这个形式（W叫Kernel，U叫Recurrent Kernel，图形可参考这里）：

$$h_t=tanh(Wx_t+Uh_{t-1}+b)$$

结果是一样的。

从源代码里可以看到，state_size和output_size都跟num_units都是同一个数字，call函数返回两个一模一样的向量。

简单调用例子：

import tensorflow as tf
import numpy as np

batch_size = 3
input_dim = 2
output_dim = 4

inputs = tf.placeholder(dtype=tf.float32, shape=(batch_size, input_dim))
previous_state = tf.random_normal(shape=(batch_size, output_dim))

cell = tf.contrib.rnn.BasicRNNCell(num_units=output_dim)
output, state = cell(inputs, previous_state)

X = np.ones(shape=(batch_size, input_dim))

with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    o, s = sess.run([output, state], feed_dict={inputs: X})

    print(X)
    print(previous_state.eval())
    print(o)
    print(s)

Out：
# Input:
[[ 1.  1.]
 [ 1.  1.]
 [ 1.  1.]]

# Previous State:
[[ 0.29562142  1.88447475 -0.71091568 -1.03161728]
 [-0.32763469 -0.4521957  -0.33536151  0.06760707]
 [-0.04627729  0.04288582 -0.62693876  0.70083541]]

# Output = State
[[ 0.1948889   0.77429289  0.41136274  0.42551333]
 [-0.99327117 -0.68583459  0.97010344 -0.56064779]
 [-0.98540735 -0.51250875  0.92181391 -0.98040372]]
[[ 0.1948889   0.77429289  0.41136274  0.42551333]
 [-0.99327117 -0.68583459  0.97010344 -0.56064779]
 [-0.98540735 -0.51250875  0.92181391 -0.98040372]]

额外

1. 想要查看内部$W_k$和 b b <script type="math/tex" id="MathJax-Element-4">b</script>的值可参考这里。