Reading a TensorArray in Tensorflow always returns zeroes

Question

I am using Tensorflow v1.15. I have a very basic implementation of the TensorArray given in the following example:

import tensorflow as tf

an_array = tf.TensorArray(dtype=tf.float32, size=5, dynamic_size=True, clear_after_read=False, element_shape=(16, 7, 2))
for i in range(5):
    val = tf.random.normal(shape=(16, 7, 2))
    an_array.write(i, val)
    print(tf.Session().run(val))
tensors = [an_array.read(j) for j in range(5)]
print(tf.Session().run(tensors))

The print in the for loop does not print all zeros, while the last print statement does. Why is this happening? Thanks.

Answer 1

tf.TensorArray.write returns a new tf.TensorArray where the writing operation has taken place. In general, the output of this function should replace the previous reference to the array:

import tensorflow as tf

an_array = tf.TensorArray(dtype=tf.float32, size=5, dynamic_size=True,
                          clear_after_read=False, element_shape=(16, 7, 2))
for i in range(5):
    val = tf.random.normal(shape=(16, 7, 2))
    # Replace tensor array reference with the written one
    an_array = an_array.write(i, val)
    print(tf.Session().run(val))
tensors = [an_array.read(j) for j in range(5)]
print(tf.Session().run(tensors))