how to create a encode_raw tensorflow function?

Question

I am trying to perform the opposite of what tf.decode_raw does.

An example would be given a tensor of dtype=tf.float32, I would like to have a function encode_raw() that takes in a float tensor and returns a Tensor of type string.

This is useful because then I can use tf.write_file to write the file.

Does anyone know how to create such a function in Tensorflow using existing functions?

Answer 1

I would recommend writing numbers as text with tf.as_string. If you really want to write them as a binary string, however, it turns out to be possible:

import tensorflow as tf

with tf.Graph().as_default():
  character_lookup = tf.constant([chr(i) for i in range(256)])
  starting_dtype = tf.float32
  starting_tensor = tf.random_normal(shape=[10, 10], stddev=1e5, 
                                     dtype=starting_dtype)
  as_string = tf.reduce_join(
      tf.gather(character_lookup,
                tf.cast(tf.bitcast(starting_tensor, tf.uint8), tf.int32)))
  back_to_tensor = tf.reshape(tf.decode_raw(as_string, starting_dtype),
                              [10, 10]) # Shape information is lost
  with tf.Session() as session:
    before, after = session.run([starting_tensor, back_to_tensor])
    print(before - after)

This for me prints an array of all zeros.

Answer 2

For those working with Python 3:

chr() has a different behavior in Python 3 that changes the byte output obtained with the code from previous answer. Replacing this code line

character_lookup = tf.constant([chr(i) for i in range(256)])

with

character_lookup = tf.constant([i.tobytes() for i in np.arange(256, dtype=np.uint8)])

fixes this issue.