## Issue

I am trying to convert the following code in Numpy to Tensorflow, so that I can use that code in my neural network. Sadly, I have failed to so yet, due to the error, that Tensorflow does not support item assignment, which I would need for this code to work.

```
# Create zeros matrices
l_bbox = np.zeros((6,6), np.float32)
l_score = np.zeros((1,6,6), np.float32)
# Set coordinates
x1, y1, x2, y2 = 1, 1, 5, 4
# Assign coordinates to the zeros matrix
l_score[:, y1:y2, x1:x2] = 1.
l_score
"""
array([[[0., 0., 0., 0., 0., 0.],
[0., 1., 1., 1., 1., 0.],
[0., 1., 1., 1., 1., 0.],
[0., 1., 1., 1., 1., 0.],
[0., 0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0., 0.]]], dtype=float32)
"""
# Assign varying values to l_bbox
for yy in range(y1, y2):
l_bbox[yy, x1:x2] = yy - y1
l_bbox
"""
array([[0., 0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0., 0.],
[0., 1., 1., 1., 1., 0.],
[0., 2., 2., 2., 2., 0.],
[0., 0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0., 0.]], dtype=float32)
"""
```

I have tried to do this first in a loop and then stack the lists up using `tf.stack`

, but that seems like a really non-tensorflow way (and it does not work on a GPU). What I would like to have is a way to do the aforementioned transformation for several batch sizes in Tensorflow. This answer suggests to use `tf.where`

but I also have no clue how to leverage this in that specific case. I would appreciate any help.

## Solution

You could slice and directly assign like this assuming the logic you have is simple as shown. But there are also other functions to do this.

```
import numpy as np
import tensorflow as tf
tfl_bbox = tf.Variable(tf.zeros([6, 6]))
tfl_score = tf.Variable(tf.zeros([1, 6, 6]))
# Set coordinates
x1, y1, x2, y2 = 1, 1, 5, 4
# Assign coordinates to the zeros matrix
tfl_score[:, y1:y2, x1:x2].assign(tf.ones_like(tf.constant(1,shape=(1,y2-y1,x2-x1)), dtype=tf.float32))
# Create zeros matrices
l_bbox = np.zeros((6,6), np.float32)
l_score = np.zeros((1,6,6), np.float32)
# Set coordinates
x1, y1, x2, y2 = 1, 1, 5, 4
# Assign coordinates to the zeros matrix
l_score[:, y1:y2, x1:x2] = 1.
print(l_score)
print(tfl_score)
"""
array([[[0., 0., 0., 0., 0., 0.],
[0., 1., 1., 1., 1., 0.],
[0., 1., 1., 1., 1., 0.],
[0., 1., 1., 1., 1., 0.],
[0., 0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0., 0.]]], dtype=float32)
"""
# Assign varying values to l_bbox
for yy in range(y1, y2):
l_bbox[yy, x1:x2] = yy - y1
tfl_bbox[yy, x1:x2].assign(tf.constant(yy - y1,shape=(x2-x1), dtype=tf.float32))
print(l_bbox)
print(tfl_bbox)
"""
array([[0., 0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0., 0.],
[0., 1., 1., 1., 1., 0.],
[0., 2., 2., 2., 2., 0.],
[0., 0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0., 0.]], dtype=float32)
"""
```

Answered By – Mohan Radhakrishnan

Answer Checked By – Jay B. (BugsFixing Admin)