gmr.m

gmm gmr gmm gmr

function [y, Sigma_y] = GMR(Priors, Mu, Sigma, x, in, out)
%
% This function performs Gaussian Mixture Regression (GMR), using the 
% parameters of a Gaussian Mixture Model (GMM). Given partial input data, 
% the algorithm computes the expected distribution for the resulting 
% dimensions. By providing temporal values as inputs, it thus outputs a 
% smooth generalized version of the data encoded in GMM, and associated 
% constraints expressed by covariance matrices.
%
% Inputs -----------------------------------------------------------------
%   o Priors:  1 x K array representing the prior probabilities of the K GMM 
%              components.
%   o Mu:      D x K array representing the centers of the K GMM components.
%   o Sigma:   D x D x K array representing the covariance matrices of the 
%              K GMM components.
%   o x:       P x N array representing N datapoints of P dimensions.
%   o in:      1 x P array representing the dimensions to consider as
%              inputs.
%   o out:     1 x Q array representing the dimensions to consider as
%              outputs (D=P+Q).
% Outputs ----------------------------------------------------------------
%   o y:       Q x N array representing the retrieved N datapoints of 
%              Q dimensions, i.e. expected means.
%   o Sigma_y: Q x Q x N array representing the N expected covariance 
%              matrices retrieved. 
%
% Copyright (c) 2006 Sylvain Calinon, LASA Lab, EPFL, CH-1015 Lausanne,
%               Switzerland, http://lasa.epfl.ch
%
% The program is free for non-commercial academic use. 
% Please contact the authors if you are interested in using the 
% software for commercial purposes. The software must not be modified or 
% distributed without prior permission of the authors.
% Please acknowledge the authors in any academic publications that have 
% made use of this code or part of it. Please use this BibTex reference: 
% 
% @article{Calinon06SMC,
%   title="On Learning, Representing and Generalizing a Task in a Humanoid 
%     Robot",
%   author="S. Calinon and F. Guenter and A. Billard",
%   journal="IEEE Transactions on Systems, Man and Cybernetics, Part B. 
%     Special issue on robot learning by observation, demonstration and 
%     imitation",
%   year="2006",
%   volume="36",
%   number="5"
% }

nbData = length(x);
nbVar = size(Mu,1);
nbStates = size(Sigma,3);

%% Fast matrix computation (see the commented code for a version involving 
%% one-by-one computation, which is easier to understand).
%%
%% Compute the influence of each GMM component, given input x
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
for i=1:nbStates
  Pxi(:,i) = Priors(i).*gaussPDF(x, Mu(in,i), Sigma(in,in,i));
end
beta = Pxi./repmat(sum(Pxi,2)+realmin,1,nbStates);
%% Compute expected means y, given input x
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
for j=1:nbStates
  y_tmp(:,:,j) = repmat(Mu(out,j),1,nbData) + Sigma(out,in,j)*inv(Sigma(in,in,j)) * (x-repmat(Mu(in,j),1,nbData));
end
beta_tmp = reshape(beta,[1 size(beta)]);
y_tmp2 = repmat(beta_tmp,[length(out) 1 1]) .* y_tmp;
y = sum(y_tmp2,3);
%% Compute expected covariance matrices Sigma_y, given input x
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
for j=1:nbStates
  Sigma_y_tmp(:,:,1,j) = Sigma(out,out,j) - (Sigma(out,in,j)*inv(Sigma(in,in,j))*Sigma(in,out,j));
end
beta_tmp = reshape(beta,[1 1 size(beta)]);
Sigma_y_tmp2 = repmat(beta_tmp.*beta_tmp, [length(out) length(out) 1 1]) .* repmat(Sigma_y_tmp,[1 1 nbData 1]);
Sigma_y = sum(Sigma_y_tmp2,4);


% %% Slow one-by-one computation (better suited to understand the algorithm) 
% %%
% %% Compute the influence of each GMM component, given input x
% %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% for i=1:nbStates
%   Pxi(:,i) = gaussPDF(x, Mu(in,i), Sigma(in,in,i));
% end
% beta = (Pxi./repmat(sum(Pxi,2)+realmin,1,nbStates))';
% %% Compute expected output distribution, given input x
% %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% y = zeros(length(out), nbData);
% Sigma_y = zeros(length(out), length(out), nbData);
% for i=1:nbData
%   % Compute expected means y, given input x
%   %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%   for j=1:nbStates
%     yj_tmp = Mu(out,j) + Sigma(out,in,j)*inv(Sigma(in,in,j)) * (x(:,i)-Mu(in,j));
%     y(:,i) = y(:,i) + beta(j,i).*yj_tmp;
%   end
%   % Compute expected covariance matrices Sigma_y, given input x
%   %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%   for j=1:nbStates
%     Sigmaj_y_tmp = Sigma(out,out,j) - (Sigma(out,in,j)*inv(Sigma(in,in,j))*Sigma(in,out,j));
%     Sigma_y(:,:,i) = Sigma_y(:,:,i) + beta(j,i)^2.* Sigmaj_y_tmp;
%   end
% end