python - BasicRNNCell 中的内部变量

Question

我有以下示例代码来测试 BasicRNNCell。我想得到它的内部矩阵，以便我可以使用我自己的代码计算 output_res、newstate_res 的值，以确保我可以重现 的值>output_res, newstate_res.

在 tensorflow 源代码中，它说 output = new_state = act(W * input + U * state + B)。有人知道我怎样才能得到 W 和 U 吗？ (我尝试访问 cell._kernel，但它不可用。)

$ cat ./main.py
#!/usr/bin/env python
# vim: set noexpandtab tabstop=2 shiftwidth=2 softtabstop=-1 fileencoding=utf-8:

import tensorflow as tf
import numpy as np

batch_size = 4
vector_size = 3

inputs = tf.placeholder(
        tf.float32
        , [batch_size, vector_size]
        )

num_units = 2
state = tf.zeros([batch_size, num_units], tf.float32)

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

X = np.zeros([batch_size, vector_size])
#X = np.ones([batch_size, vector_size])
with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())

    output_res, newstate_res = sess.run([output, newstate], feed_dict = {inputs: X})
    print(output_res)
    print(newstate_res)
sess.close()

$ ./main.py
[[ 0.  0.]
 [ 0.  0.]
 [ 0.  0.]
 [ 0.  0.]]
[[ 0.  0.]
 [ 0.  0.]
 [ 0.  0.]
 [ 0.  0.]]

Answer 1

简短回答:您知道您在寻找cell._kernel。下面是一些使用 variables 属性获取内核(和偏差)的代码，该属性在大多数 TensorFlow RNN 中:

import tensorflow as tf
import numpy as np

batch_size = 4
vector_size = 3
inputs = tf.placeholder(tf.float32, [batch_size, vector_size])

num_units = 2
state = tf.zeros([batch_size, num_units], tf.float32)

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

print("Output of cell.variables is a list of Tensors:")
print(cell.variables)
kernel, bias = cell.variables

X = np.zeros([batch_size, vector_size])
with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())

    output_, newstate_, k_, b_ = sess.run(
        [output, newstate, kernel, bias], feed_dict = {inputs: X})
    print("Output:")
    print(output_)
    print("New State == Output:")
    print(newstate_)
    print("\nKernel:")
    print(k_)
    print("\nBias:")
    print(b_)

输出

Output of cell.variables is a list of Tensors:
[<tf.Variable 'basic_rnn_cell/kernel:0' shape=(5, 2) dtype=float32_ref>, 
<tf.Variable 'basic_rnn_cell/bias:0' shape=(2,) dtype=float32_ref>]
Output:
[[ 0.  0.]
 [ 0.  0.]
 [ 0.  0.]
 [ 0.  0.]]
New State == Output:
[[ 0.  0.]
 [ 0.  0.]
 [ 0.  0.]
 [ 0.  0.]]

Kernel:
[[ 0.41417515 -0.64997244]
 [-0.40868729 -0.90995187]
 [ 0.62134564 -0.88962835]
 [-0.35878009 -0.25680023]
 [ 0.35606658 -0.83596271]]

Bias:
[ 0.  0.]

长答:你还问了怎么得到W和U，我复制call的实现，讨论一下W和U在哪。

def call(self, inputs, state):
     """Most basic RNN: output = new_state = act(W * input + U * state + B)."""

    gate_inputs = math_ops.matmul(
        array_ops.concat([inputs, state], 1), self._kernel)
    gate_inputs = nn_ops.bias_add(gate_inputs, self._bias)
    output = self._activation(gate_inputs)
    return output, output

看起来不像有一个 W 和一个 U，但它们确实存在。本质上，内核的第一个 vector_size 行是 W，内核的下一个 num_units 行是 U。也许在 LaTeX 中查看逐元素数学会有所帮助:

我使用 m 作为通用批量索引，v 作为 vector_size，n 作为 num_units，b 为 batch_size。还有 [ ; ] 表示串联。由于 TensorFlow 是批量处理的，因此实现通常使用右乘矩阵。

因为这是一个非常基本的 RNN，output == new_state。下一次迭代的“历史”只是当前迭代的输出。