'Assign' 2D block slice in tensorflow?

问题

Does tensorflow provide a way to create a zero-tensor while replacing a general slice by another one? In particular, I need to assign 2D blocks in a matrix-tensor. Here's an example of what I need to achieve:

Given a tensor of variable shape, eg.

[1 2 3
 4 5 6],

and another tensor defining the slice, eg.

[0 0 0 0 0
 0 1 1 1 0
 0 1 1 1 0],

the new tensor should look as follows:

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

I know there's scatter_nd, but it seems like it can only 'replace' values along a full axis. Do I miss any operation or is there any workaround to achieve this?

回答1:

You can use assign on the appropriate slice of y:

import numpy as np
import tensorflow as tf

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

sess = tf.InteractiveSession()
sess.run(tf.global_variables_initializer())

z = y[1:,1:4].assign(x)
z.eval()

# returns
# array([[0, 0, 0, 0, 0],
#        [0, 1, 2, 3, 0],
#        [0, 4, 5, 6, 0]])

EDIT

Same thing dynamically (here for the position only)

import numpy as np
import tensorflow as tf

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

pos = tf.placeholder(tf.int32, shape=(2,))

sess = tf.InteractiveSession()
sess.run(tf.global_variables_initializer())

z = y[pos[0]:2+pos[0],pos[1]:3+pos[1]].assign(x)
z.eval({pos: [1,1]})

来源：https://stackoverflow.com/questions/44888587/assign-2d-block-slice-in-tensorflow