How to use im2col and col2im together with GEMM to archive a conv2d operation?

[tangjiasheng]

Can you provide some sample codes? No idea about how to manipulate 5 dims (or 6 dims).

[akhauriyash]

Check out the link here:
https://leonardoaraujosantos.gitbooks.io/artificial-inteligence/content/making_faster.html

Now, for better understanding:
open im2col.py and add the code below at the bottom. Then run python im2col.py and observe the output:

init = Variable(torch.randn((1, 3, 4, 4)).cuda())
z  = 0
for i in range(init.size(1)):
    for j in range(init.size(2)):
        for k in range(init.size(3)):
            init[0, i, j, k] = z + 1
            z+=1

unrolled = im2col_batch(init, 2, 1, 0)
print(init.size())
print(unrolled.size())
unr_reshape = torch.reshape(unrolled, (1, unrolled.size(1)*unrolled.size(2)*unrolled.size(3), unrolled.size(4)*unrolled.size(5)))
print(unr_reshape)

You should be able to replicate the blog image with this code. I hope this helps in understanding how to manipulate the dimensions!

Cheers!

[akhauriyash]

Alternatively, just change this function to:

def im2col_batch(input, kernel_size, stride, padding):
    if input.dim() == 3:
        return _im2col(input, kernel_size, stride, padding)
    elif input.dim() == 4:
        shape = (input.size(0),) + im2col_shape(input.size()[1:], kernel_size, stride, padding)
        out = input.new(*shape)
        for x, o in zip(input, out):
             _im2col(x, kernel_size, stride, padding, out=o)
        out = torch.reshape(out, (out.size(0), out.size(1)*out.size(2)*out.size(3), out.size(4)*out.size(5)))
        return out