Tensorflow: Stringify tensor as a whole (without creating tensor of strings)

Question

I have the following array,

a = tf.random.uniform((5,2), 0, 10)

<tf.Tensor: shape=(5, 2), dtype=float32, numpy=
array([[3.8656425 , 6.7514324 ],
       [0.49138665, 3.5968459 ],
       [4.435692  , 4.7223845 ],
       [7.3588967 , 0.31867146],
       [1.6837907 , 3.2266355 ]], dtype=float32)>

What I want is array of stringified arrays like below this would return a numpy array, but I want to do tensorflow ops to return a tensor:

list(map(str, a.numpy()))

['[3.8656425 6.7514324]',
 '[0.49138665 3.5968459 ]',
 '[4.435692  4.7223845]',
 '[7.3588967  0.31867146]',
 '[1.6837907 3.2266355]']

When I use tf.as_string()

tf.as_string(a)

<tf.Tensor: shape=(5, 2), dtype=string, numpy=
array([[b'3.865643', b'6.751432'],
       [b'0.491387', b'3.596846'],
       [b'4.435692', b'4.722384'],
       [b'7.358897', b'0.318671'],
       [b'1.683791', b'3.226635']], dtype=object)>

I also tried using

tf.map_fn(tf.as_string, a, dtype=tf.string)

# Same output as above

tf.as_string() converts float/int tensors to string tensors of same shape. Is there any tensorflow op that stringifies tensors as a whole?

What version of TensorFlow are you using? And do you need the result as a tensor, and using only TensorFlow operations, or is it okay to get a NumPy array and then convert that to string? — javidcf
– javidcf, Commented Mar 3, 2020 at 11:05
I'm using tensorflow 2; I would need a tensor result, I don't want to use the numpy way mentioned in the question (or a similar method) and then convert it to a tensor. I'd like an op that returns a tensor, thanks — Saravanabalagi Ramachandran
– Saravanabalagi Ramachandran, Commented Mar 3, 2020 at 12:09

javidcf · Accepted Answer · 2020-03-03 11:42:38Z

2

You can use tf.strings.format:

import tensorflow as tf

tf.random.set_seed(0)
a = tf.random.uniform((5,2), 0, 10)
b = tf.map_fn(lambda r: tf.strings.format('{}', r, summarize=-1), a, tf.string)
print(b)
# tf.Tensor(
# [b'[2.91975141 2.06566453]' b'[5.35390759 5.61257458]'
#  b'[4.16674519 8.0782795]' b'[4.93225098 9.98129272]'
#  b'[6.96735144 1.25373602]'], shape=(5,), dtype=string)

edited Mar 3, 2020 at 11:42

answered Mar 3, 2020 at 11:10

javidcf

59.9k7 gold badges87 silver badges134 bronze badges

Sign up to request clarification or add additional context in comments.

Comments

Saravanabalagi Ramachandran · Accepted Answer · 2020-03-03 10:58:46Z

0

As a workaround, I'm joining individual strings

b = tf.map_fn(lambda x: tf.strings.join(x, separator=" "), tf.as_string(a))
b = tf.map_fn(lambda x: tf.strings.join(['[', x, ']']), b)

<tf.Tensor: shape=(5,), dtype=string, numpy=
array([b'[3.865643 6.751432]', b'[0.491387 3.596846]',
       b'[4.435692 4.722384]', b'[7.358897 0.318671]',
       b'[1.683791 3.226635]'], dtype=object)>

2 Comments

Frederik Bode Over a year ago

this is not a workaround, this the way to do it. Your misconception is that you think each row of your matrix in your tf.Tensor is one tensor, or it belongs together or something. You might think that, but tensorflow doesn't. You can see that in <tf.Tensor: shape=(5, 2), ..., it's just a 5x2 matrix. It might be printed as a list of lists, but it's not, it's a tf.Tensor. So, you'll need some extra steps to get what you want. As your entire implementation of the extra steps is within tensorflow, this LGTM.

Saravanabalagi Ramachandran Over a year ago

My objective was to simply Stringify tensor as a whole, and in the example I have given in the question, I am trying to do the same, except that to each element in a tensor (i.e inside a map_fn). I was expecting an inbuilt tf fn that does that. That said I agree that tensorflow might be doing something similar to do the formatting using tf.strings.format, with more fancy options though.

Collectives™ on Stack Overflow

Tensorflow: Stringify tensor as a whole (without creating tensor of strings)

2 Answers 2

Comments

2 Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

2 Answers 2

Comments

2 Comments

Your Answer

Sign up or log in

Post as a guest

Related