如何使用 TensorFlow 中的指定索引访问 3D 张量的元素？

Question

I'm trying to get the rows of a 3D tensor in a specific order of indices. Here are the inputs:

import tensorflow as tf

matrix = tf.constant([
    [[0, 1], [2, 3], [4, 5], [6, 7]], 
    [[8, 9], [10, 11], [12, 13], [14, 15]], 
    [[16, 17], [18, 19], [20, 21], [22, 23]], 
    [[24, 25], [26, 27], [28, 29], [30, 31]], 
    [[32, 33], [34, 35], [36, 37], [38, 39]]
])

indx = tf.constant([[3,2,1,0], [0,1,2,3], [1,0,3,2], [0,3,1,2], [1,2,3,0]])

# required output tensor:
[[[6, 7], [4, 5], [2, 3], [0, 1]],
 [[8, 9], [10, 11], [12, 13], [14, 15]],
 [[18, 19], [16, 17], [22, 23], [20, 21]],
 [[24, 25], [30, 31], [26, 27], [28, 29]],
 [[34, 35], [36, 37], [38, 39], [32, 33]]]

I'm struggling with tf.gather_nd() . Any suggestion? I can see it's happening here but I'm not sure how to apply on entire matrix without using for loop or tf.map_fn

print(tf.gather_nd(matrix[0], tf.expand_dims(indx, -1)[0]).numpy().tolist())
print(tf.gather_nd(matrix[1], tf.expand_dims(indx, -1)[1]).numpy().tolist())
print(tf.gather_nd(matrix[2], tf.expand_dims(indx, -1)[2]).numpy().tolist())
print(tf.gather_nd(matrix[3], tf.expand_dims(indx, -1)[3]).numpy().tolist())
print(tf.gather_nd(matrix[4], tf.expand_dims(indx, -1)[4]).numpy().tolist())

"""
[[6, 7], [4, 5], [2, 3], [0, 1]]
[[8, 9], [10, 11], [12, 13], [14, 15]]
[[18, 19], [16, 17], [22, 23], [20, 21]]
[[24, 25], [30, 31], [26, 27], [28, 29]]
[[34, 35], [36, 37], [38, 39], [32, 33]]
"""

EDIT: I asked a similar question with respect to numpy. A clever indexing answer does solves the numpy version, but it's hard to apply it on Tensors. Feel free to take a look at the accepted answer here: How can I get elements from 3D matrix using specified indices in numpy?

Answer 1

Duh, that was stupid; There is already a very great function available that works on multi-dimensional array in tensorflow; tf.gather() Check out the batch_dims argument for more information.

>> tf.gather(matrix, indx, batch_dims=1).numpy().tolist()
[[[6, 7], [4, 5], [2, 3], [0, 1]],
 [[8, 9], [10, 11], [12, 13], [14, 15]],
 [[18, 19], [16, 17], [22, 23], [20, 21]],
 [[24, 25], [30, 31], [26, 27], [28, 29]],
 [[34, 35], [36, 37], [38, 39], [32, 33]]]