Files
OpenFace/model_training/CCNF/CCRF/lib/gradientCCRF.m
2018-05-05 11:21:09 +01:00

92 lines
3.3 KiB
Matlab

function [ logGradientAlphas, logGradientBetas, SigmaInv, ChDecomp ] = gradientCCRF( alphas, betas, lambda_a, lambda_b, precalcQ2withoutBeta, xq, yq, mask, precalcYQ, useIndicator, PrecalcQ2Flat)
%GRADIENTPRF Summary of this function goes here
% Detailed explanation goes here
% Calculate the Sigma inverse now
% [SigmaInv2] = CalcSigmaCCRF(alphas, betas, precalcQ2withoutBeta, mask);
% This is an optimised version as it does not use the whole matrix but
% a lower diagonal part due to symmetry
numElemsInSeq = size(precalcQ2withoutBeta{1}, 1);
[SigmaInv] = CalcSigmaCCRFflat(alphas, betas, numElemsInSeq, PrecalcQ2Flat, mask, useIndicator);
% Get the actual sigma from out SigmaInv
% Sigma = inv(SigmaInv);
% Below is an optimised version of the above using Cholesky decomposition
% which decomposes a matrix into a upper triangular (R) and its
% conjugate transpose R'; A = R'*R for real numbers, thus
% inv(A) = inv(R)inv(R')
ChDecomp=chol(SigmaInv);
I=eye(size(SigmaInv));
% Rinv = (R\I);
% Sigma = Rinv*Rinv';
% This is a very slightly faster version of the above
Sigma=ChDecomp\(ChDecomp'\I);
b = CalcbCCRF(alphas, xq, mask, useIndicator);
% mu = SigmaInv \ b = Sigma * b;
% as we've calculate Sigma already, this is equivalent of the above
mu = Sigma * b;
logGradientAlphas = zeros(size(alphas));
logGradientBetas = zeros(size(betas));
K1 = numel(alphas);
K2 = numel(betas);
% calculating the derivative of L with respect to alpha_k
for k = 1:K1
if(useIndicator)
dQ1da = diag(mask(:,k));
dbda = xq(:,k).*mask(:,k);
gaussGradient = -yq'*dQ1da*yq +2*yq'*dbda -2 * dbda' * mu + mu'*dQ1da*mu;
zGradient = Sigma(:)'*dQ1da(:);
else
% if we don't use the masks here's a speedup
gaussGradient = -yq'*yq +2*yq'*xq(:,k) -2 * xq(:,k)' * mu + sum(mu.^2);
% simplification as trace(Sigma * I) = trace(Sigma)
zGradient = trace(Sigma);
end
% add the Z derivative now
dLda = zGradient + gaussGradient;
% add regularisation
dLda = dLda - lambda_a * alphas(k);
logGradientAlphas(k) = alphas(k) * dLda;
end
% This was done for gradient checking
% [alphasG, betaG] = gradientAnalytical(nFrames, S, alphas, beta, xq, yq, mask);
% calculating the derivative of log(L) with respect to the betas
for k=1:K2
% Bs = Bs(:,:,k);
% dSdb = q2./betas(k); we precalculate this, as it does not change
% over the course of optimisation (dSdb - dSigma/dbeta)
dSdb = precalcQ2withoutBeta{k};
% -yq'*dSdb*yq can be precalculated as they don't change through
% iterations (precalcQ2withoutBeta is dSdb
% gaussGradient = -yq'*dSdb*yq + mu'*dSdb*mu;
% this does the above line
gaussGradient = precalcYQ(k) + mu'*dSdb*mu;
% zGradient = trace(Sigma*dSdb);
zGradient = Sigma(:)'*dSdb(:); % equivalent but faster to the above line
dLdb = gaussGradient + zGradient;
% add regularisation term
dLdb = dLdb - lambda_b * betas(k);
logGradientBetas(k) = betas(k) * dLdb;
end
end