Re: K-L Distance

Hi,

I wrote some code a while back in Matlab to calculate the KL for
Gaussian Mixture Models.

It's pretty crude but it works.

mix1 and mix2 are the mixture model distributions.
gmmprob returns the density at x.

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

divergence = quad('kl_function', a, b, [], [], mix1, mix2);

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

function y = kl_function(x, mix1, mix2)
%KL_FUNCTION - The interior of the KL integral

len = length(x);
y = zeros(1,len);

for i = 1:len
temp = gmmprob(mix1, x(i)).*log(gmmprob(mix1, x(i)) ./ gmmprob(mix2,
x(i)));;
if(isnan(temp) | isinf(temp))
warning('Integration Problem.')
y(i) = 0;
else
y(i) = temp;
end
end

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

Hope this helps.

M

.

Quantcast