Tensorflow - add 2x2 matrix to 4x4 fractal at starting index-CodePudding

Consider a n x n tensor, such as:

board = tf.zeros([9,9], dtype=tf.int32)

How do I add another tensor (matrix) to it? If this isn't possible, I want to know how to update the value of the (original) tensor (beginning at location (x, y)) with the values in the smaller to_add tensor. Thanks.

to_add = tf.constant([[1,2,3],[4,5,6],[7,8,9]], dtype=tf.int32)

For example, how do I add the following matrices, starting at (1,0)

0 0 0       1 0        0 1 0 
0 0 0       1 0    =   0 1 0 
0 0 0                  0 0 0

CodePudding user response：

import tensorflow as tf

board = tf.zeros([9,9], dtype=tf.int32)
to_add = tf.constant([[1,2,3],[4,5,6],[7,8,9]], dtype=tf.int32)
start = (2, 3)

def add_t_to_t(t_large, t_small, start):
    i, j = start
    m, n = t_small.shape
    large = t_large.numpy()
    large[i:i   m, j:j   n]  = t_small.numpy()
    return tf.convert_to_tensor(large)

board = add_t_to_t(board, to_add, start)
print(board)

prints

tf.Tensor(
[[0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0]
 [0 0 0 1 2 3 0 0 0]
 [0 0 0 4 5 6 0 0 0]
 [0 0 0 7 8 9 0 0 0]
 [0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0]], shape=(9, 9), dtype=int32)

Wouldn't know how to do this with pure tensorflow, but I am sure there is a (more cumbersome) way to do so.

CodePudding user response：

Add follows NumPy broadcasting rules, so you need to add tensors that have the same shape. What you could use is tf.pad() function to add padding to your to_add and make it the same shape as board

Here is an example:

import tensorflow as tf
board = tf.zeros([9,9], dtype=tf.int32)
to_add = tf.constant([[1,2,3],[4,5,6],[7,8,9]], dtype=tf.int32)

padding = tf.constant([[0,6],[1,5]], dtype=tf.int32)
padded = tf.pad(to_add, padding)
print('Pad:')
print(padded)

print('New Board:')
new_board = tf.add(board, padded)
print(new_board)

Output:

Pad:
tf.Tensor(
[[0 1 2 3 0 0 0 0 0]
 [0 4 5 6 0 0 0 0 0]
 [0 7 8 9 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0]], shape=(9, 9), dtype=int32)
New Board:
tf.Tensor(
[[0 1 2 3 0 0 0 0 0]
 [0 4 5 6 0 0 0 0 0]
 [0 7 8 9 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0]], shape=(9, 9), dtype=int32)