#include "utils.h"
#include <stdlib.h>
+#include <math.h>
#include <gsl/gsl_linalg.h>
// TODO: don't recompute indexes ai(...) and mi(...) when possible
Real* phi, // parametre de moyenne renormalisé, calculé par l'EM
Real* rho, // parametre de variance renormalisé, calculé par l'EM
Real* pi, // parametre des proportions renormalisé, calculé par l'EM
- Real* LLF, // log vraisemblance associée à cet échantillon, pour les valeurs estimées des paramètres
+ Real* llh, // (derniere) log vraisemblance associée à cet échantillon,
+ // pour les valeurs estimées des paramètres
Real* S,
+ int* affec,
// additional size parameters
int n, // nombre d'echantillons
int p, // nombre de covariables
copyArray(phiInit, phi, p*m*k);
copyArray(rhoInit, rho, m*m*k);
copyArray(piInit, pi, k);
- zeroArray(LLF, maxi);
//S is already allocated, and doesn't need to be 'zeroed'
//Other local variables: same as in R
Real* b = (Real*)malloc(k*sizeof(Real));
Real* X2 = (Real*)malloc(n*p*k*sizeof(Real));
Real* Y2 = (Real*)malloc(n*m*k*sizeof(Real));
- Real dist = 0.;
- Real dist2 = 0.;
- int ite = 0;
+ *llh = -INFINITY;
Real* pi2 = (Real*)malloc(k*sizeof(Real));
- Real* ps = (Real*)malloc(m*k*sizeof(Real));
- Real* nY2 = (Real*)malloc(m*k*sizeof(Real));
- Real* ps1 = (Real*)malloc(n*m*k*sizeof(Real));
- Real* Gam = (Real*)malloc(n*k*sizeof(Real));
const Real EPS = 1e-15;
// Additional (not at this place, in R file)
Real* gam2 = (Real*)malloc(k*sizeof(Real));
Real* sqNorm2 = (Real*)malloc(k*sizeof(Real));
+ Real* detRho = (Real*)malloc(k*sizeof(Real));
gsl_matrix* matrix = gsl_matrix_alloc(m, m);
gsl_permutation* permutation = gsl_permutation_alloc(m);
Real* YiRhoR = (Real*)malloc(m*sizeof(Real));
Real* Rho = (Real*)malloc(m*m*k*sizeof(Real));
Real* Pi = (Real*)malloc(k*sizeof(Real));
- while (ite < mini || (ite < maxi && (dist >= tau || dist2 >= sqrt(tau))))
+ for (int ite=1; ite<=maxi; ite++)
{
copyArray(phi, Phi, p*m*k);
copyArray(rho, Rho, m*m*k);
{
//Y2[,mm,r] = sqrt(gam[,r]) * Y[,mm]
for (int u=0; u<n; u++)
- Y2[ai(u,mm,r,n,m,k)] = sqrt(gam[mi(u,r,n,k)]) * Y[mi(u,mm,m,n)];
+ Y2[ai(u,mm,r,n,m,k)] = sqrt(gam[mi(u,r,n,k)]) * Y[mi(u,mm,n,m)];
}
for (int i=0; i<n; i++)
{
}
//tant que les proportions sont negatives
- int kk = 0;
- int pi2AllPositive = 0;
+ int kk = 0,
+ pi2AllPositive = 0;
Real invN = 1./n;
while (!pi2AllPositive)
{
kk++;
}
- //(pi.^gamma)*b
+ //sum(pi^gamma * b)
Real piPowGammaDotB = 0.;
for (int v=0; v<k; v++)
piPowGammaDotB += pow(pi[v],gamma) * b[v];
- //(pi2.^gamma)*b
+ //sum(pi2^gamma * b)
Real pi2PowGammaDotB = 0.;
for (int v=0; v<k; v++)
pi2PowGammaDotB += pow(pi2[v],gamma) * b[v];
- //transpose(gam2)*log(pi2)
- Real prodGam2logPi2 = 0.;
+ //sum(gam2 * log(pi2))
+ Real gam2DotLogPi2 = 0.;
for (int v=0; v<k; v++)
- prodGam2logPi2 += gam2[v] * log(pi2[v]);
+ gam2DotLogPi2 += gam2[v] * log(pi2[v]);
+
//t(m) la plus grande valeur dans la grille O.1^k tel que ce soit décroissante ou constante
- while (-invN*a + lambda*piPowGammaDotB < -invN*prodGam2logPi2 + lambda*pi2PowGammaDotB
+ while (-invN*a + lambda*piPowGammaDotB < -invN*gam2DotLogPi2 + lambda*pi2PowGammaDotB
&& kk<1000)
{
Real pow_01_kk = pow(0.1,kk);
//pi2 = pi + 0.1^kk * (1/n*gam2 - pi)
for (int v=0; v<k; v++)
pi2[v] = pi[v] + pow_01_kk * (invN*gam2[v] - pi[v]);
- //pi2 was updated, so we recompute pi2PowGammaDotB and prodGam2logPi2
+ //pi2 was updated, so we recompute pi2PowGammaDotB and gam2DotLogPi2
pi2PowGammaDotB = 0.;
for (int v=0; v<k; v++)
pi2PowGammaDotB += pow(pi2[v],gamma) * b[v];
- prodGam2logPi2 = 0.;
+ gam2DotLogPi2 = 0.;
for (int v=0; v<k; v++)
- prodGam2logPi2 += gam2[v] * log(pi2[v]);
+ gam2DotLogPi2 += gam2[v] * log(pi2[v]);
kk++;
}
Real t = pow(0.1,kk);
{
for (int mm=0; mm<m; mm++)
{
+ Real ps = 0.,
+ nY2 = 0.;
+ // Compute ps, and nY2 = sum(Y2[,mm,r]^2)
for (int i=0; i<n; i++)
{
- //< X2(i,:,r) , phi(:,mm,r) >
+ //< X2[i,,r] , phi[,mm,r] >
Real dotProduct = 0.;
for (int u=0; u<p; u++)
dotProduct += X2[ai(i,u,r,n,p,k)] * phi[ai(u,mm,r,p,m,k)];
- //ps1[i,mm,r] = Y2[i,mm,r] * sum(X2[i,,r] * phi[,mm,r])
- ps1[ai(i,mm,r,n,m,k)] = Y2[ai(i,mm,r,n,m,k)] * dotProduct;
+ //ps = ps + Y2[i,mm,r] * sum(X2[i,,r] * phi[,mm,r])
+ ps += Y2[ai(i,mm,r,n,m,k)] * dotProduct;
+ nY2 += Y2[ai(i,mm,r,n,m,k)] * Y2[ai(i,mm,r,n,m,k)];
}
- //ps[mm,r] = sum(ps1[,mm,r])
- Real sumPs1 = 0.;
- for (int u=0; u<n; u++)
- sumPs1 += ps1[ai(u,mm,r,n,m,k)];
- ps[mi(mm,r,m,k)] = sumPs1;
- //nY2[mm,r] = sum(Y2[,mm,r])
- Real sumNy2 = 0.;
- for (int u=0; u<n; u++)
- sumNy2 += Y2[ai(u,mm,r,n,m,k)];
- nY2[mi(mm,r,m,k)] = sumNy2;
- //rho[mm,mm,r] = (ps[mm,r]+sqrt(ps[mm,r]^2+4*nY2[mm,r]*(gam2[r]))) / (2*nY2[mm,r])
- rho[ai(mm,mm,r,m,m,k)] = ( ps[mi(mm,r,m,k)] + sqrt( ps[mi(mm,r,m,k)]*ps[mi(mm,r,m,k)]
- + 4*nY2[mi(mm,r,m,k)] * gam2[r] ) ) / (2*nY2[mi(mm,r,m,k)]);
+ //rho[mm,mm,r] = (ps+sqrt(ps^2+4*nY2*gam2[r])) / (2*nY2)
+ rho[ai(mm,mm,r,m,m,k)] = (ps + sqrt(ps*ps + 4*nY2 * gam2[r])) / (2*nY2);
}
}
+
for (int r=0; r<k; r++)
{
for (int j=0; j<p; j++)
{
for (int mm=0; mm<m; mm++)
{
- //sum(phi[-j,mm,r] * Gram2[j, setdiff(1:p,j),r])
- Real dotPhiGram2 = 0.0;
+ //sum(phi[-j,mm,r] * Gram2[j,-j,r])
+ Real phiDotGram2 = 0.;
for (int u=0; u<p; u++)
{
if (u != j)
- dotPhiGram2 += phi[ai(u,mm,r,p,m,k)] * Gram2[ai(j,u,r,p,p,k)];
+ phiDotGram2 += phi[ai(u,mm,r,p,m,k)] * Gram2[ai(j,u,r,p,p,k)];
}
- //S[j,mm,r] = -rho[mm,mm,r]*ps2[j,mm,r] + sum(phi[-j,mm,r] * Gram2[j, setdiff(1:p,j),r])
- S[ai(j,mm,r,p,m,k)] = -rho[ai(mm,mm,r,m,m,k)] * ps2[ai(j,mm,r,p,m,k)] + dotPhiGram2;
- Real pow_pir_gamma = pow(pi[r],gamma);
- if (fabs(S[ai(j,mm,r,p,m,k)]) <= n*lambda*pow_pir_gamma)
- phi[ai(j,mm,r,p,m,k)] = 0;
- else if (S[ai(j,mm,r,p,m,k)] > n*lambda*pow_pir_gamma)
+ //S[j,mm,r] = -rho[mm,mm,r]*ps2[j,mm,r] + sum(phi[-j,mm,r] * Gram2[j,-j,r])
+ S[ai(j,mm,r,p,m,k)] = -rho[ai(mm,mm,r,m,m,k)] * ps2[ai(j,mm,r,p,m,k)]
+ + phiDotGram2;
+ Real pirPowGamma = pow(pi[r],gamma);
+ if (fabs(S[ai(j,mm,r,p,m,k)]) <= n*lambda*pirPowGamma)
+ phi[ai(j,mm,r,p,m,k)] = 0.;
+ else if (S[ai(j,mm,r,p,m,k)] > n*lambda*pirPowGamma)
{
- phi[ai(j,mm,r,p,m,k)] = (n*lambda*pow_pir_gamma - S[ai(j,mm,r,p,m,k)])
+ phi[ai(j,mm,r,p,m,k)] = (n*lambda*pirPowGamma - S[ai(j,mm,r,p,m,k)])
/ Gram2[ai(j,j,r,p,p,k)];
}
else
{
- phi[ai(j,mm,r,p,m,k)] = -(n*lambda*pow_pir_gamma + S[ai(j,mm,r,p,m,k)])
+ phi[ai(j,mm,r,p,m,k)] = -(n*lambda*pirPowGamma + S[ai(j,mm,r,p,m,k)])
/ Gram2[ai(j,j,r,p,p,k)];
}
}
// Etape E //
/////////////
+ // Precompute det(rho[,,r]) for r in 1...k
int signum;
- Real sumLogLLF2 = 0.;
+ for (int r=0; r<k; r++)
+ {
+ for (int u=0; u<m; u++)
+ {
+ for (int v=0; v<m; v++)
+ matrix->data[u*m+v] = rho[ai(u,v,r,m,m,k)];
+ }
+ gsl_linalg_LU_decomp(matrix, permutation, &signum);
+ detRho[r] = gsl_linalg_LU_det(matrix, signum);
+ }
+
+ Real sumLogLLH = 0.;
for (int i=0; i<n; i++)
{
for (int r=0; r<k; r++)
YiRhoR[u] += Y[mi(i,v,n,m)] * rho[ai(v,u,r,m,m,k)];
}
- //compute X(i,:)*phi(:,:,r)
+ //compute X[i,]%*%phi[,,r]
for (int u=0; u<m; u++)
{
XiPhiR[u] = 0.;
sqNorm2[r] += (YiRhoR[u]-XiPhiR[u]) * (YiRhoR[u]-XiPhiR[u]);
}
- Real sumLLF1 = 0.;
Real sumGamI = 0.;
for (int r=0; r<k; r++)
{
- //compute det(rho[,,r]) [TODO: avoid re-computations]
- for (int u=0; u<m; u++)
- {
- for (int v=0; v<m; v++)
- matrix->data[u*m+v] = rho[ai(u,v,r,m,m,k)];
- }
- gsl_linalg_LU_decomp(matrix, permutation, &signum);
- Real detRhoR = gsl_linalg_LU_det(matrix, signum);
- Gam[mi(i,r,n,k)] = pi[r] * exp(-.5*sqNorm2[r]) * detRhoR;
- sumLLF1 += Gam[mi(i,r,n,k)] / gaussConstM;
- sumGamI += Gam[mi(i,r,n,k)];
+ gam[mi(i,r,n,k)] = pi[r] * exp(-.5*sqNorm2[r]) * detRho[r];
+ sumGamI += gam[mi(i,r,n,k)];
}
- sumLogLLF2 += log(sumLLF1);
- for (int r=0; r<k; r++)
+
+ sumLogLLH += log(sumGamI) - log(gaussConstM);
+ if (sumGamI > EPS) //else: gam[i,] is already ~=0
{
- //gam[i,] = Gam[i,] / sumGamI
- gam[mi(i,r,n,k)] = sumGamI > EPS ? Gam[mi(i,r,n,k)] / sumGamI : 0.;
+ for (int r=0; r<k; r++)
+ gam[mi(i,r,n,k)] /= sumGamI;
}
}
Real sumPen = 0.;
for (int r=0; r<k; r++)
sumPen += pow(pi[r],gamma) * b[r];
- //LLF[ite] = -sumLogLLF2/n + lambda*sumPen
- LLF[ite] = -invN * sumLogLLF2 + lambda * sumPen;
- dist = ite==0 ? LLF[ite] : (LLF[ite] - LLF[ite-1]) / (1. + fabs(LLF[ite]));
+ Real last_llh = *llh;
+ //llh = -sumLogLLH/n + lambda*sumPen
+ *llh = -invN * sumLogLLH + lambda * sumPen;
+ Real dist = ite==1 ? *llh : (*llh - last_llh) / (1. + fabs(*llh));
//Dist1 = max( abs(phi-Phi) / (1+abs(phi)) )
Real Dist1 = 0.;
}
}
//dist2=max([max(Dist1),max(Dist2),max(Dist3)]);
- dist2 = Dist1;
+ Real dist2 = Dist1;
if (Dist2 > dist2)
dist2 = Dist2;
if (Dist3 > dist2)
dist2 = Dist3;
- ite++;
+ if (ite >= mini && (dist >= tau || dist2 >= sqrt(tau)))
+ break;
}
- //TODO: affec = apply(gam, 1,which.max) à traduire, et retourner affec aussi.
+ //affec = apply(gam, 1, which.max)
+ for (int i=0; i<n; i++)
+ {
+ Real rowMax = 0.;
+ affec[i] = 0;
+ for (int j=0; j<k; j++)
+ {
+ if (gam[mi(i,j,n,k)] > rowMax)
+ {
+ affec[i] = j+1; //R indices start at 1
+ rowMax = gam[mi(i,j,n,k)];
+ }
+ }
+ }
//free memory
free(b);
free(gam);
- free(Gam);
free(Phi);
free(Rho);
free(Pi);
- free(ps);
- free(nY2);
- free(ps1);
free(Gram2);
free(ps2);
+ free(detRho);
gsl_matrix_free(matrix);
gsl_permutation_free(permutation);
free(XiPhiR);