alternative for tensorprod that is coder compatible

Hi,

I have implemented a custom 2D convolution layer. Since nested loops were too slow I rearranged my input U and my filter weights V to be of dimensions (filter_width *filter_height) x input channels x (output_width *output_height) and (filter_width *filter_height) x input channels x number of filters and then used the tensorprod function and reshaped the output to output_height x output_width x number of filters.

To my actual problem: I have two 3-D tensors for which I want to compute a tensor product, which contracts the first two dimensions of each tensor with each other.

U: a x b x c

V: a x b x d

W = tensorprod(U,V,[ 1 2])

W will then be of dimensions c x d.

I cannot use the Matlab tensorprod function because it is not Matlab coder compatible. Is there another fast solution to compute W, which is also compatible with Matlab coder?

[Matthias Kreuzer]

I solved it myself

function result = mytensorprod(A,B,dims)

% performs the tensor product for matrices A and B

% A and B are contracted along the dimensions specified in dims

% First A and B are permuted to perform the inner product

% Same functionality as the matlab function tensorprod

sizeA = size(A); sizeB = size(B);

dimAouter = 1:length(sizeA); dimAouter(dims)=[];

dimBouter = 1:length(sizeB); dimBouter(dims)=[];

Ap = permute(A,[dimAouter,dims]);

Apr = reshape(Ap,[prod(sizeA(dimAouter)),prod(sizeA(dims))]);

Bp = permute(B,[dims,dimBouter]);

Bpr = reshape(Bp,[prod(sizeB(dims)),prod(sizeB(dimBouter))]);

result = Apr*Bpr;

result = reshape(result,[sizeA(dimAouter),sizeB(dimBouter)]);

end