tensorflow：如何交错两个张量的列（例如使用 tf.scatter_nd）？

Question

I've read the tf.scatter_nd documentation and run the example code for 1D and 3D tensors... and now I'm trying to do it for a 2D tensor. I want to 'interleave' the columns of two tensors. For 1D tensors, one can do this via

'''
We want to interleave elements of 1D tensors arr1 and arr2, where
arr1 = [10, 11, 12]
arr2 = [1, 2, 3, 4, 5, 6]
such that
desired result = [1, 2, 10, 3, 4, 11, 5, 6, 12]
'''

import tensorflow as tf

with tf.Session() as sess:

    updates1 = tf.constant([1,2,3,4,5,6])
    indices1 = tf.constant([[0], [1], [3], [4], [6], [7]])
    shape = tf.constant([9])
    scatter1 = tf.scatter_nd(indices1, updates1, shape)

    updates2 = tf.constant([10,11,12])
    indices2 = tf.constant([[2], [5], [8]])
    scatter2 = tf.scatter_nd(indices2, updates2, shape)

    result = scatter1 + scatter2

    print(sess.run(result))

(aside: is there a better way to do this? I'm all ears.)

This gives the output

[ 1 2 10 3 4 11 5 6 12]

Yay! that worked!

Now lets' try to extend this to 2D.

    '''
    We want to interleave the *columns* (not rows; rows would be easy!) of

    arr1 = [[1,2,3,4,5,6],[1,2,3,4,5,6],[1,2,3,4,5,6]]
    arr2 = [[10 11 12], [10 11 12], [10 11 12]]
    such that
    desired result = [[1,2,10,3,4,11,5,6,12],[1,2,10,3,4,11,5,6,12],[1,2,10,3,4,11,5,6,12]]
    '''

    updates1 = tf.constant([[1,2,3,4,5,6],[1,2,3,4,5,6],[1,2,3,4,5,6]])
    indices1 = tf.constant([[0], [1], [3], [4], [6], [7]])
    shape = tf.constant([3, 9])
    scatter1 = tf.scatter_nd(indices1, updates1, shape)

This gives the error ValueError: The outer 1 dimensions of indices.shape=[6,1] must match the outer 1 dimensions of updates.shape=[3,6]: Dimension 0 in both shapes must be equal, but are 6 and 3. Shapes are [6] and [3]. for 'ScatterNd_2' (op: 'ScatterNd') with input shapes: [6,1], [3,6], [2]. ValueError: The outer 1 dimensions of indices.shape=[6,1] must match the outer 1 dimensions of updates.shape=[3,6]: Dimension 0 in both shapes must be equal, but are 6 and 3. Shapes are [6] and [3]. for 'ScatterNd_2' (op: 'ScatterNd') with input shapes: [6,1], [3,6], [2].

Seems like my indices is specifying row indices instead of column indices, and given the way that arrays are "connected" in numpy and tensorflow (ie row-major order), does that mean I need to explicitly specify every single pair of indices for every element in updates1 ? Or is there some kind of 'wildcard' specification I can use for the rows? (Note indices1 = tf.constant([[:,0], [:,1], [:,3], [:,4], [:,6], [:,7]]) gives syntax errors, as it probably should.)

Would it be easier to just do a transpose, interleave the rows, then transpose back? Because I tried that...

scatter1 = tf.scatter_nd(indices1, tf.transpose(updates1), tf.transpose(shape))
print(sess.run(tf.transpose(scatter1)))

...and got a much longer error message, that I don't feel like posting unless someone requests it.

PS- I searched to make sure this isn't a duplicate -- I find it hard to imagine that someone else hasn't asked this before -- but turned up nothing.

Answer 1

This is pure slicing but I didn't know that syntax like arr1[0:,:][:,:2] actually works. It seems it does but not sure if it is better.

This may be the wildcard slicing mechanism you are looking for.

arr1 = tf.constant([[1,2,3,4,5,6],[1,2,3,4,5,7],[1,2,3,4,5,8]])
arr2 = tf.constant([[10, 11, 12], [10, 11, 12], [10, 11, 12]])

with tf.Session() as sess :
    sess.run( tf.global_variables_initializer() )
    print(sess.run(tf.concat([arr1[0:,:][:,:2], arr2[0:,:] [:,:1],
                              arr1[0:,:][:,2:4],arr2[0:, :][:, 1:2],
                              arr1[0:,:][:,4:6],arr2[0:, :][:, 2:3]],axis=1)))

Output is

[[ 1  2 10  3  4 11  5  6 12]
 [ 1  2 10  3  4 11  5  7 12]
 [ 1  2 10  3  4 11  5  8 12]]

So, for example,

arr1[0:,:] returns

[[1 2 3 4 5 6]
 [1 2 3 4 5 7]
 [1 2 3 4 5 8]]

and arr1[0:,:][:,:2] returns the first two columns

[[1 2]
 [1 2]
 [1 2]]

axis is 1.

Answer 2

Some moderators might have regarded my question as a duplicate of this one , not because the questions are the same, but only because the answers contain parts one can use to answer this question -- ie specifying every index combination by hand.

A totally different method would be to multiply by a permutation matrix as shown in the last answer to this question . Since my original question was about scatter_nd, I'm going to post this solution but wait to see what other answers come in... (Alternatively, I or someone could edit the question to make it about reordering columns, not specific to scatter_nd --EDIT: I have just edited the question title to reflect this ).

Here, we concatenate the two different arrays/tensors...

import numpy as np
import tensorflow as tf

sess = tf.Session()

# the ultimate application is for merging variables which should be in groups,
#   e.g. in this example, [1,2,10] is a group of 3, and there are 3 groups of 3
n_groups = 3
vars_per_group = 3    # once the single value from arr2 (below) is included

arr1 = 10+tf.range(n_groups, dtype=float)
arr1 = tf.stack((arr1,arr1,arr1),0)
arr2 = 1+tf.range(n_groups * (vars_per_group-1), dtype=float)
arr2 = tf.stack((arr2,arr2,arr2),0)

catted = tf.concat((arr1,arr2),1)        # concatenate the two arrays together
print("arr1 = \n",sess.run(arr1))
print("arr2 = \n",sess.run(arr2))
print("catted = \n",sess.run(catted))

Which gives output

arr1 = 
 [[10. 11. 12.]
 [10. 11. 12.]
 [10. 11. 12.]]
arr2 = 
 [[1. 2. 3. 4. 5. 6.]
 [1. 2. 3. 4. 5. 6.]
 [1. 2. 3. 4. 5. 6.]]
catted = 
 [[10. 11. 12.  1.  2.  3.  4.  5.  6.]
 [10. 11. 12.  1.  2.  3.  4.  5.  6.]
 [10. 11. 12.  1.  2.  3.  4.  5.  6.]]

Now we build the permutation matrix and multiply...

start_index = 2               # location of where the interleaving begins

# cml = "column map list" is the list of where each column will get mapped to
cml = [start_index + x*(vars_per_group) for x in range(n_groups)]  # first array  
for i in range(n_groups):                                       # second array 
    cml += [x + i*(vars_per_group) for x in range(start_index)] # vars before start_index
    cml += [1 + x + i*(vars_per_group) + start_index \
        for x in range(vars_per_group-start_index-1)]           # vars after start_index
print("\n cml = ",cml,"\n")

# Create a permutation matrix using p
np_perm_mat = np.zeros((len(cml), len(cml)))
for idx, i in enumerate(cml):
    np_perm_mat[idx, i] = 1
perm_mat = tf.constant(np_perm_mat,dtype=float)

result = tf.matmul(catted, perm_mat)
print("result = \n",sess.run(result))

Which gives output

cml =  [2, 5, 8, 0, 1, 3, 4, 6, 7] 

result = 
 [[ 1.  2. 10.  3.  4. 11.  5.  6. 12.]
 [ 1.  2. 10.  3.  4. 11.  5.  6. 12.]
 [ 1.  2. 10.  3.  4. 11.  5.  6. 12.]]

Even though this doesn't use scatter_nd as the original question asked, one thing I like about this is, you can allocate the perm_mat once in some __init__() method, and hang on to it, and after that initial overhead it's just matrix-matrix multiplication by a sparse, constant matrix, which should be pretty fast. (?)

Still happy to wait and see what other answers might come in.