Reputation: 277
I am trying to "vectorize" this loop in Matlab for computational efficiency
for t=1:T
j=1;
for m=1:M
for n=1:N
y(t,j) = v{m,n} + data(t,:)*b{m,n} + data(t,:)*f{m,n}*data(t,:)';
j=j+1;
end
end
end
Where v
is a (M
x N
) cell of scalars. b
is a (M
x N
) cell of (K
x 1) vectors. f
is a (M
x N
) cell of (K
x K
) matrices. data
is a (T
x K
) array.
To give an example of what I mean the code I used to vectorize the same loop without the quadratic term is:
B = [reshape(cell2mat(v)',1,N*M);cell2mat(reshape(b'),1,M*N)];
X = [ones(T,1),data];
y = X*B;
Thanks!
Upvotes: 1
Views: 87
Reputation: 221614
Here's the most vectorized form targeted for performance -
% Extract as multi-dim arrays
vA = reshape([v{:}],M,N);
bA = reshape([b{:}],K,M,N);
fA = reshape([f{:}],K,K,M,N);
% Perform : data(t,:)*f{m,n} for all iterations
data_f_mult = reshape(data*reshape(fA,K,[]),T,K,M,N);
% Now there are three parts :
% v{m,n}
% data(t,:)*b{m,n}
% data(t,:)*f{m,n}*data(t,:)';
% Compute those parts one by one
parte1 = vA(:).';
parte2 = data*reshape(bA,[],M*N);
parte3 = zeros(T,M*N);
for t = 1:T
parte3(t,:) = data(t,:)*reshape(data_f_mult(t,:,:),K,[]);
end
% Finally sum those up and to present in desired format permute dims
sums = bsxfun(@plus, parte1, parte2 + parte3);
out = reshape(permute(reshape(sums,T,M,N),[1,3,2]),[],M*N);
Upvotes: 0
Reputation: 277
For those interested here was the solution I found
f = f';
tMat = blkdiag(f{:})+(blkdiag(f{:}))';
y2BB = [reshape(cell2mat(v)',1,N*M);...
cell2mat(reshape(b',1,M*N));...
reshape(diag(blkdiag(f{:})),K,N*M);...
reshape(tMat((tril(tMat,-1)~=0)),sum(1:K-1),M*N)];
y2YBar = [ones(T,1),data,data.^2];
jj=1;
kk=1;
ll=1;
for k=1:sum(1:K-1)
y2YBar = [y2YBar,data(:,jj).*data(:,kk+jj)];
if kk<(K-ll)
kk=kk+1;
else
kk=1;
jj=jj+1;
ll=ll+1;
end
end
y = y2YBar*y2BB;
Upvotes: 2