Suppose I have an input matrix of shape (batch_size ,channels ,h ,w)
in this case (1 ,2 ,3 ,3)
[[[[ 0., 1., 2.],
[ 3., 4., 5.],
[ 6., 7., 8.]],
[[ 9., 10., 11.],
[12., 13., 14.],
[15., 16., 17.]]]])
to do a convolution with it i unroll it to the shape of (batch_size ,channels * kernel_size * kernel_size ,out_h * out_w) which is:
[[[ 0., 1., 3., 4.],
[ 1., 2., 4., 5.],
[ 3., 4., 6., 7.],
[ 4., 5., 7., 8.],
[ 9., 10., 12., 13.],
[10., 11., 13., 14.],
[12., 13., 15., 16.],
[13., 14., 16., 17.]]]
now i want to get the unrolled matrix back to its original form which looks like this:
# for demonstration only the first and second column of the unrolled matrix
# the output should be the same shape as the initial matrix -> initialized to zeros
# current column -> [ 0., 1., 3., 4., 9., 10., 12., 13.]
[[[[0+0, 0+1, 0],
[0+3, 0+4, 0],
[0 , 0 , 0]],
[[0+9 , 0+10, 0],
[0+12, 0+13, 0],
[0 , 0 , 0]]]]
# for the next column it would be
# current column -> [ 1., 2., 4., 5., 10., 11., 13., 14.]
[[[[0 , 1+1, 0+2],
[3 , 4+4, 0+5],
[0 , 0 , 0 ]],
[[9 , 10+10, 0+11],
[12 , 13+13, 0+14],
[0 , 0 , 0 ]]]])
you basically put your unrolled elements back to its original place and sum the overlapping parts together.
But now to my question:
How could one implement this as fast as possible using numpy and as less loops as possible. I already just looped through it kernel by kernel but this aproach isnt feasible with larger inputs. I think this could be parallelized quite a bit but my numpy indexing and overall knowledge isnt good enough to figure out a good solution by myself.
thanks for reading and have a nice day :)