Best Practice to Convert a Tensor to TensorArray in TensorFlow – TensorFlow Tutorial

In this tutorial, we will introduce the way to convert a tensor to a tensorarray object in tensorflow, which is very useful when you are bulding a custom lstm or bilstm.

Create a tensor

We create a tensor first.

import tensorflow as tf
import numpy as np
x = tf.Variable(np.array(range(36)), dtype = np.float32, name = 'x')
x = tf.reshape(x, [3, 3, 4])

The tensor x will be:

[[[ 0.  1.  2.  3.]
  [ 4.  5.  6.  7.]
  [ 8.  9. 10. 11.]]

 [[12. 13. 14. 15.]
  [16. 17. 18. 19.]
  [20. 21. 22. 23.]]

 [[24. 25. 26. 27.]
  [28. 29. 30. 31.]
  [32. 33. 34. 35.]]]

Create a tensorarray to save tensors

We will create a tensorarray object gen_o with size = 3

gen_o = tf.TensorArray(dtype=tf.float32, size=3,
                                             dynamic_size=False, infer_shape=True)

To understand how to create or use tensorarray, you can read:

Understand TensorFlow TensorArray: A Beginner Tutorial

Convert a tensor to tensorarray

We will use two methods to convert a tensor to tensorarray.

Method 1: use tensorarray.unstack()

Here is an example:

gen_o = gen_o.unstack(x)
z0 = gen_o.read(0)
z1 = gen_o.read(1)
z2 = gen_o.read(2)

Then print z0, z1 and z2.

with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    print(sess.run(x))
    print("z0=")
    print(sess.run(z0))
    print("z1=")
    print(sess.run(z1))
    print("z2=")
    print(sess.run(z2))

You will get the result.

z0=
[[ 0.  1.  2.  3.]
 [ 4.  5.  6.  7.]
 [ 8.  9. 10. 11.]]
z1=
[[12. 13. 14. 15.]
 [16. 17. 18. 19.]
 [20. 21. 22. 23.]]
z2=
[[24. 25. 26. 27.]
 [28. 29. 30. 31.]
 [32. 33. 34. 35.]]

We can find the shape of z0, z1 and z2 is (3, 4)

Method 2: use tensorarray.split() function

Here is an example:

gen_o = gen_o.split(x, lengths = [1,1,1])
z0 = gen_o.read(0)
z1 = gen_o.read(1)
z2 = gen_o.read(2)

You should notice lengths = [1,1,1], which means there are only len(lengths) tensors in tensorarray object.

Print z0, z1 and z2, you will get the result:

z0=
[[[ 0.  1.  2.  3.]
  [ 4.  5.  6.  7.]
  [ 8.  9. 10. 11.]]]
z1=
[[[12. 13. 14. 15.]
  [16. 17. 18. 19.]
  [20. 21. 22. 23.]]]
z2=
[[[24. 25. 26. 27.]
  [28. 29. 30. 31.]
  [32. 33. 34. 35.]]]

You will find the shape of z0, z1 and z2 is (1, 3, 4), which is different from gen_o.unstack(x).