CODEWITHSUNDEEP
matlabnormalizationvectorization
Tim
Tim

Reputation: 2050

Vectorizing matlab column normalization

I want to normalize a matrix per column. I now have this code which works fine:

A = randn(10,3)

maxA = max(A,[],1)
minA = min(A,[],1)

for i=1:size(A,2)
    A(:,i) = (A(:,i) - minA(i) ./ (maxA(i) - minA(i))
end

However, since my matrix will be much bigger, about 10k by 60k looping will take forever. How could I vectorize my code? I have thought of using Matlab's normc but that does not do the same as my code.

Upvotes: 3

Views: 2462

Answers (2)

Serg
Serg

Reputation: 14108

In general it could be:

A = (A - ones(size(A)) * diag(minA)) / diag(maxA - minA);

or 

A = (A - ones(size(A)) * diag(minA)) * diag(1 ./ (maxA - minA));

but taking into account the size:

m = repmat(minA, size(A, 1), 1);
n = repmat(maxA - minA, size(A, 1), 1);
A = (A - m) ./ n;

Upvotes: 0

H.Muster
H.Muster

Reputation: 9317

You can use bsxfun

A = randn(10,3)

maxA = max(A,[],1)
minA = min(A,[],1)

bsxfun(@minus, A, minA ./ abs(maxA - minA))

However, I do not quite understand your normalizing. Wouldn't one rather use 

(A(:,i) - minA(i)) ./ (maxA(i) - minA(i))

to normalize? If so, the bsxfun statement should read: 

bsxfun(@times, bsxfun(@minus, A, minA), 1./abs(maxA - minA))

Upvotes: 5

Related Questions