Let we have a tensor with shape $n\times d\times h\times w\times p\times p$ and we want to concat innet matrix with shape $p\time p$, such that we made a matrix with shape $n\times d\times ph\times pw$. How can I do it?
array([[[[[[ 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]]]]]])
after concat
array([[[[0, 1, 2, 9, 10, 11],
[3, 4, 5, 12, 13, 14],
[6, 7, 8, 15, 16, 17],
[18, 19, 20, 27, 28, 29],
[21, 22, 23, 30, 31, 32],
[24, 25, 26, 33, 34, 35]]]])
I have done a lot of experiments using reshape, but without success. One of my experiments
a.reshape(n, d, p*h, p*w)
I can do it using for loop, but I think that without this, too, it is possible. Please help me. Code using for loop
p = 3
arr = np.arange(1*1*2*2*p*p).reshape(1, 1, 2, 2, p, p)
answer = np.zeros(shape=(1, 1, 2*p, 2*p))
for (n, d, h, w) in np.ndindex(*arr.shape[:4]):
answer[n, d, h:h+p, w:w+p] = arr[n, d, h, w]
In [15]: arr=np.arange(0,36).reshape(2,2,3,3)
reshape cannot reorder the elements of the array. I started with [0,1,...35], and reshape retains that:
In [18]: arr.reshape(2,3,6)
Out[18]:
array([[[ 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]]])
We have to somehow reorder the elements, putting the [9,10,11] block adjacent to the [0,1,2]. transpose is one such tool:
In [19]: arr.transpose(0,2,1,3)
Out[19]:
array([[[[ 0, 1, 2],
[ 9, 10, 11]],
[[ 3, 4, 5],
[12, 13, 14]],
[[ 6, 7, 8],
[15, 16, 17]]],
[[[18, 19, 20],
[27, 28, 29]],
[[21, 22, 23],
[30, 31, 32]],
[[24, 25, 26],
[33, 34, 35]]]])
In [20]: arr.transpose(0,2,1,3).reshape(6,6)
Out[20]:
array([[ 0, 1, 2, 9, 10, 11],
[ 3, 4, 5, 12, 13, 14],
[ 6, 7, 8, 15, 16, 17],
[18, 19, 20, 27, 28, 29],
[21, 22, 23, 30, 31, 32],
[24, 25, 26, 33, 34, 35]])
To do the same by assignment to a 'blank' we need something like:
In [32]: res=np.zeros((6,6),int)
In [33]: res[:,:3] = arr[:,::2,:].reshape(6,3)
In [34]: res
Out[34]:
array([[ 0, 1, 2, 0, 0, 0],
[ 3, 4, 5, 0, 0, 0],
[ 6, 7, 8, 0, 0, 0],
[18, 19, 20, 0, 0, 0],
[21, 22, 23, 0, 0, 0],
[24, 25, 26, 0, 0, 0]])
In [35]: res[:,3:] = arr[:,1::2,:].reshape(6,3)
In [36]: res
Out[36]:
array([[ 0, 1, 2, 9, 10, 11],
[ 3, 4, 5, 12, 13, 14],
[ 6, 7, 8, 15, 16, 17],
[18, 19, 20, 27, 28, 29],
[21, 22, 23, 30, 31, 32],
[24, 25, 26, 33, 34, 35]])
A concatenate version of that same block join:
In [41]: np.concatenate((arr[:,::2], arr[:,1::2]), axis=3)
Out[41]:
array([[[[ 0, 1, 2, 9, 10, 11],
[ 3, 4, 5, 12, 13, 14],
[ 6, 7, 8, 15, 16, 17]]],
[[[18, 19, 20, 27, 28, 29],
[21, 22, 23, 30, 31, 32],
[24, 25, 26, 33, 34, 35]]]])