'Tensor' object does not support item assignment while need to slice and then do assignmnet in tensor array

Question

I want to find part of a tensor and then assign a constant number to that but I received this error. It seems that the assignment to a tensor is not allowed in tensorflow. with that in mind, anyone has any idea how can I accomplish this one?

So for example if the tensor is like this:

tf_a2 = tf.Variable(([[2, 5, 1, 4, 3],
                      [1, 6, 4, 2, 3],
                      [0, 0, 0, 6, 6],
                      [2, 1, 1, 3, 3],
                      [4, 4, 1, 2, 3]]))

I want to find the elements per row in which they have the same value then I replace any element with N except the first element from left.

For example in the example above, in row=3, three elements =0, so I kept the most left one the same as it is and then replace the right elements with the same value with N. in the same row there are two element with 6 value,I kept the most left element and replce all elements after that with the same value with N.

In row=4 ,1 is repeated two times, I again keep the most left one and replace any the right item which has the same value.

In row =5, 4 is repeated two times again. I keep the most left item, and then replace any item after that with the same value with N.

So for N= 9, the result would be:

    [[2 5 1 4 3]
     [1 6 4 2 3]
     [0 9 9 6 9]
     [2 1 9 3 9]
     [4 9 1 2 3]]

I have the correct code in numpy below:

  numpy code:  (a2[:,1:])[a2[:,1:]==a2[:,:-1]] = N

but I need to do it in tensorflow , I tried the below code though it still raises the same error:

tf.where(tf.equal(a2[:,1:], a2[:, :-1]),N,a2[:,1:])

The error:
    a2[:,1:][a2[:,1:]==a2[:,:-1]] = N
TypeError: 'Tensor' object does not support item assignment

I also looked at the links with the same error here, but they propose a solution exactly regarding their coding which did not match my code.

Thank you in advance

Please provide a Minimal, Reproducible Example that leads to the mentioned error. Please also include the full traceback. — a_guest
– a_guest, Commented Jun 11, 2019 at 22:12
@a_guest Thank you following with my question, I have updated it, please let me know if I need to add more data. — sariii
– sariii, Commented Jun 11, 2019 at 22:28
You code examples are still not self-contained. Please update with something that we can copy & paste in order to verify the behavior you describe. Furthermore the statement "I want to find rows which the left and right item are the same, then change it to a constant number like N." is not clear to me. Could you please elaborate a bit more? What means "left" and "right" element here? — a_guest
– a_guest, Commented Jun 12, 2019 at 13:20
@a_guest I have updated my question with a good example and explanations — sariii
– sariii, Commented Jun 12, 2019 at 14:47

KrisR89 · Accepted Answer · 2019-06-12 15:00:47Z

1

As the error message says, " 'Tensor' object does not support item assignment". But there are at least one workaround. One method is to element-wise multiply each element you want to change by zero (in the original matrix) and then create a new matrix (same shape) with all element zeros without the one you want to change. And then you can just add them together to get the desired matrix.

If i understand you're numpy solution correctly for this specific case, you always want to ignore the first column? If that is so, I think this tensorflow solution should work for you (tested and verified for tensorflow version 1.13.1).

import tensorflow as tf
tf.enable_eager_execution()

tf_a2 = tf.Variable(([[2, 5, 1, 4, 3],
                      [1, 6, 4, 2, 3],
                      [0, 0, 0, 6, 6],
                      [2, 1, 1, 3, 3],
                      [4, 4, 1, 2, 3]]))

N=9

first_col_change = tf.zeros([tf_a2.shape[0],1], dtype=tf.int32)
last_cols_change = tf.cast(tf.equal(tf_a2[:,1:], tf_a2[:, :-1]),tf.int32)
change_bool = tf.concat([first_col_change, last_cols_change],axis=-1)
not_change_bool = 1-change_bool
tf_a2_changed = tf_a2*not_change_bool + change_bool*N

print(tf_a2_changed)

which gives the output:

tf.Tensor(
[[2 5 1 4 3]
 [1 6 4 2 3]
 [0 9 9 6 9]
 [2 1 9 3 9]
 [4 9 1 2 3]], shape=(5, 5), dtype=int32)

edited Jun 12, 2019 at 15:00

answered Jun 12, 2019 at 14:28

KrisR89

1,5418 silver badges11 bronze badges

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

9 Comments

sariii Over a year ago

Thanks for following my question, I can not run your code successfully it raises this error: FailedPreconditionError (see above for traceback): Attempting to use uninitialized value Variable [[Node: Variable/read = Identity[T=DT_FLOAT, _device="/job:localhost/replica:0/task:0/device:GPU:0"](Variable)]]

sariii Over a year ago

I have also updated the question with better example

KrisR89 Over a year ago

What tensorflow version are you on? try "print("TensorFlow Ver: ", tf.__version__)". You can try "tf.enable_eager_execution()" it gives you ability to go line by line and print each tensor. great for error search. Just tested the code again here, and works for tensorflow 1.13.1

sariii Over a year ago

I have 1.10.0 tensorflow. Can you also, share your code in which you get its output? I just want to double check

KrisR89 Over a year ago

@sariii good that it worked! Tensorflow 1.13.1 is an "in-between" release for tensorflow 2.0. Therefore it works without sessions etc, which i don't think 1.10.0 support. I suggest you try out the new version as it will be more similar to the new 2.0 when it get out, and easier to prototype with.

|

Collectives™ on Stack Overflow

'Tensor' object does not support item assignment while need to slice and then do assignmnet in tensor array

1 Answer 1

9 Comments

Your Answer

Linked

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

9 Comments

Your Answer

Sign up or log in

Post as a guest

Linked

Related