--- /dev/null
+#include <stdlib.h>
+#include <gsl/gsl_linalg.h>
+#include "utils.h"
+
+// Compute pseudo-inverse of a square matrix
+static Real* pinv(const Real* matrix, int dim)
+{
+ gsl_matrix* U = gsl_matrix_alloc(dim,dim);
+ gsl_matrix* V = gsl_matrix_alloc(dim,dim);
+ gsl_vector* S = gsl_vector_alloc(dim);
+ gsl_vector* work = gsl_vector_alloc(dim);
+ Real EPS = 1e-10; //threshold for singular value "== 0"
+
+ //copy matrix into U
+ copyArray(matrix, U->data, dim*dim);
+
+ //U,S,V = SVD of matrix
+ gsl_linalg_SV_decomp(U, V, S, work);
+ gsl_vector_free(work);
+
+ // Obtain pseudo-inverse by V*S^{-1}*t(U)
+ Real* inverse = (Real*)malloc(dim*dim*sizeof(Real));
+ for (int i=0; i<dim; i++)
+ {
+ for (int ii=0; ii<dim; ii++)
+ {
+ Real dotProduct = 0.0;
+ for (int j=0; j<dim; j++)
+ dotProduct += V->data[i*dim+j] * (S->data[j] > EPS ? 1.0/S->data[j] : 0.0) * U->data[ii*dim+j];
+ inverse[i*dim+ii] = dotProduct;
+ }
+ }
+
+ gsl_matrix_free(U);
+ gsl_matrix_free(V);
+ gsl_vector_free(S);
+ return inverse;
+}
+
+// TODO: comment EMGrank purpose
+void EMGrank_core(
+ // IN parameters
+ const Real* Pi, // parametre de proportion
+ const Real* Rho, // parametre initial de variance renormalisé
+ int mini, // nombre minimal d'itérations dans l'algorithme EM
+ int maxi, // nombre maximal d'itérations dans l'algorithme EM
+ const Real* X, // régresseurs
+ const Real* Y, // réponse
+ Real tau, // seuil pour accepter la convergence
+ const int* rank, // vecteur des rangs possibles
+ // OUT parameters
+ Real* phi, // parametre de moyenne renormalisé, calculé par l'EM
+ Real* LLF, // log vraisemblance associé à cet échantillon, pour les valeurs estimées des paramètres
+ // additional size parameters
+ int n, // taille de l'echantillon
+ int p, // nombre de covariables
+ int m, // taille de Y (multivarié)
+ int k) // nombre de composantes
+{
+ // Allocations, initializations
+ Real* Phi = (Real*)calloc(p*m*k,sizeof(Real));
+ Real* hatBetaR = (Real*)malloc(p*m*sizeof(Real));
+ int signum;
+ Real invN = 1.0/n;
+ int deltaPhiBufferSize = 20;
+ Real* deltaPhi = (Real*)malloc(deltaPhiBufferSize*sizeof(Real));
+ int ite = 0;
+ Real sumDeltaPhi = 0.0;
+ Real* YiRhoR = (Real*)malloc(m*sizeof(Real));
+ Real* XiPhiR = (Real*)malloc(m*sizeof(Real));
+ Real* Xr = (Real*)malloc(n*p*sizeof(Real));
+ Real* Yr = (Real*)malloc(n*m*sizeof(Real));
+ Real* tXrXr = (Real*)malloc(p*p*sizeof(Real));
+ Real* tXrYr = (Real*)malloc(p*m*sizeof(Real));
+ gsl_matrix* matrixM = gsl_matrix_alloc(p, m);
+ gsl_matrix* matrixE = gsl_matrix_alloc(m, m);
+ gsl_permutation* permutation = gsl_permutation_alloc(m);
+ gsl_matrix* V = gsl_matrix_alloc(m,m);
+ gsl_vector* S = gsl_vector_alloc(m);
+ gsl_vector* work = gsl_vector_alloc(m);
+
+ //Initialize class memberships (all elements in class 0; TODO: randomize ?)
+ int* Z = (int*)calloc(n, sizeof(int));
+
+ //Initialize phi to zero, because some M loops might exit before phi affectation
+ zeroArray(phi, p*m*k);
+
+ while (ite<mini || (ite<maxi && sumDeltaPhi>tau))
+ {
+ /////////////
+ // Etape M //
+ /////////////
+
+ //M step: Mise à jour de Beta (et donc phi)
+ for (int r=0; r<k; r++)
+ {
+ //Compute Xr = X(Z==r,:) and Yr = Y(Z==r,:)
+ int cardClustR=0;
+ for (int i=0; i<n; i++)
+ {
+ if (Z[i] == r)
+ {
+ for (int j=0; j<p; j++)
+ Xr[mi(cardClustR,j,n,p)] = X[mi(i,j,n,p)];
+ for (int j=0; j<m; j++)
+ Yr[mi(cardClustR,j,n,m)] = Y[mi(i,j,n,m)];
+ cardClustR++;
+ }
+ }
+ if (cardClustR == 0)
+ continue;
+
+ //Compute tXrXr = t(Xr) * Xr
+ for (int j=0; j<p; j++)
+ {
+ for (int jj=0; jj<p; jj++)
+ {
+ Real dotProduct = 0.0;
+ for (int u=0; u<cardClustR; u++)
+ dotProduct += Xr[mi(u,j,n,p)] * Xr[mi(u,jj,n,p)];
+ tXrXr[mi(j,jj,p,p)] = dotProduct;
+ }
+ }
+
+ //Get pseudo inverse = (t(Xr)*Xr)^{-1}
+ Real* invTXrXr = pinv(tXrXr, p);
+
+ // Compute tXrYr = t(Xr) * Yr
+ for (int j=0; j<p; j++)
+ {
+ for (int jj=0; jj<m; jj++)
+ {
+ Real dotProduct = 0.0;
+ for (int u=0; u<cardClustR; u++)
+ dotProduct += Xr[mi(u,j,n,p)] * Yr[mi(u,jj,n,m)];
+ tXrYr[mi(j,jj,p,m)] = dotProduct;
+ }
+ }
+
+ //Fill matrixM with inverse * tXrYr = (t(Xr)*Xr)^{-1} * t(Xr) * Yr
+ for (int j=0; j<p; j++)
+ {
+ for (int jj=0; jj<m; jj++)
+ {
+ Real dotProduct = 0.0;
+ for (int u=0; u<p; u++)
+ dotProduct += invTXrXr[mi(j,u,p,p)] * tXrYr[mi(u,jj,p,m)];
+ matrixM->data[j*m+jj] = dotProduct;
+ }
+ }
+ free(invTXrXr);
+
+ //U,S,V = SVD of (t(Xr)Xr)^{-1} * t(Xr) * Yr
+ gsl_linalg_SV_decomp(matrixM, V, S, work);
+
+ //Set m-rank(r) singular values to zero, and recompose
+ //best rank(r) approximation of the initial product
+ for (int j=rank[r]; j<m; j++)
+ S->data[j] = 0.0;
+
+ //[intermediate step] Compute hatBetaR = U * S * t(V)
+ double* U = matrixM->data; //GSL require double precision
+ for (int j=0; j<p; j++)
+ {
+ for (int jj=0; jj<m; jj++)
+ {
+ Real dotProduct = 0.0;
+ for (int u=0; u<m; u++)
+ dotProduct += U[j*m+u] * S->data[u] * V->data[jj*m+u];
+ hatBetaR[mi(j,jj,p,m)] = dotProduct;
+ }
+ }
+
+ //Compute phi(:,:,r) = hatBetaR * Rho(:,:,r)
+ for (int j=0; j<p; j++)
+ {
+ for (int jj=0; jj<m; jj++)
+ {
+ Real dotProduct=0.0;
+ for (int u=0; u<m; u++)
+ dotProduct += hatBetaR[mi(j,u,p,m)] * Rho[ai(u,jj,r,m,m,k)];
+ phi[ai(j,jj,r,p,m,k)] = dotProduct;
+ }
+ }
+ }
+
+ /////////////
+ // Etape E //
+ /////////////
+
+ Real sumLogLLF2 = 0.0;
+ for (int i=0; i<n; i++)
+ {
+ Real sumLLF1 = 0.0;
+ Real maxLogGamIR = -INFINITY;
+ for (int r=0; r<k; r++)
+ {
+ //Compute
+ //Gam(i,r) = Pi(r) * det(Rho(:,:,r)) * exp( -1/2 * (Y(i,:)*Rho(:,:,r) - X(i,:)...
+ //*phi(:,:,r)) * transpose( Y(i,:)*Rho(:,:,r) - X(i,:)*phi(:,:,r) ) );
+ //split in several sub-steps
+
+ //compute det(Rho(:,:,r)) [TODO: avoid re-computations]
+ for (int j=0; j<m; j++)
+ {
+ for (int jj=0; jj<m; jj++)
+ matrixE->data[j*m+jj] = Rho[ai(j,jj,r,m,m,k)];
+ }
+ gsl_linalg_LU_decomp(matrixE, permutation, &signum);
+ Real detRhoR = gsl_linalg_LU_det(matrixE, signum);
+
+ //compute Y(i,:)*Rho(:,:,r)
+ for (int j=0; j<m; j++)
+ {
+ YiRhoR[j] = 0.0;
+ for (int u=0; u<m; u++)
+ YiRhoR[j] += Y[mi(i,u,n,m)] * Rho[ai(u,j,r,m,m,k)];
+ }
+
+ //compute X(i,:)*phi(:,:,r)
+ for (int j=0; j<m; j++)
+ {
+ XiPhiR[j] = 0.0;
+ for (int u=0; u<p; u++)
+ XiPhiR[j] += X[mi(i,u,n,p)] * phi[ai(u,j,r,p,m,k)];
+ }
+
+ //compute dotProduct < Y(:,i)*rho(:,:,r)-X(i,:)*phi(:,:,r) . Y(:,i)*rho(:,:,r)-X(i,:)*phi(:,:,r) >
+ Real dotProduct = 0.0;
+ for (int u=0; u<m; u++)
+ dotProduct += (YiRhoR[u]-XiPhiR[u]) * (YiRhoR[u]-XiPhiR[u]);
+ Real logGamIR = log(Pi[r]) + log(detRhoR) - 0.5*dotProduct;
+
+ //Z(i) = index of max (gam(i,:))
+ if (logGamIR > maxLogGamIR)
+ {
+ Z[i] = r;
+ maxLogGamIR = logGamIR;
+ }
+ sumLLF1 += exp(logGamIR) / pow(2*M_PI,m/2.0);
+ }
+
+ sumLogLLF2 += log(sumLLF1);
+ }
+
+ // Assign output variable LLF
+ *LLF = -invN * sumLogLLF2;
+
+ //newDeltaPhi = max(max((abs(phi-Phi))./(1+abs(phi))));
+ Real newDeltaPhi = 0.0;
+ for (int j=0; j<p; j++)
+ {
+ for (int jj=0; jj<m; jj++)
+ {
+ for (int r=0; r<k; r++)
+ {
+ Real tmpDist = fabs(phi[ai(j,jj,r,p,m,k)]-Phi[ai(j,jj,r,p,m,k)])
+ / (1.0+fabs(phi[ai(j,jj,r,p,m,k)]));
+ if (tmpDist > newDeltaPhi)
+ newDeltaPhi = tmpDist;
+ }
+ }
+ }
+
+ //update distance parameter to check algorithm convergence (delta(phi, Phi))
+ //TODO: deltaPhi should be a linked list for perf.
+ if (ite < deltaPhiBufferSize)
+ deltaPhi[ite] = newDeltaPhi;
+ else
+ {
+ sumDeltaPhi -= deltaPhi[0];
+ for (int u=0; u<deltaPhiBufferSize-1; u++)
+ deltaPhi[u] = deltaPhi[u+1];
+ deltaPhi[deltaPhiBufferSize-1] = newDeltaPhi;
+ }
+ sumDeltaPhi += newDeltaPhi;
+
+ // update other local variables
+ for (int j=0; j<m; j++)
+ {
+ for (int jj=0; jj<p; jj++)
+ {
+ for (int r=0; r<k; r++)
+ Phi[ai(jj,j,r,p,m,k)] = phi[ai(jj,j,r,p,m,k)];
+ }
+ }
+ ite++;
+ }
+
+ //free memory
+ free(hatBetaR);
+ free(deltaPhi);
+ free(Phi);
+ gsl_matrix_free(matrixE);
+ gsl_matrix_free(matrixM);
+ gsl_permutation_free(permutation);
+ gsl_vector_free(work);
+ gsl_matrix_free(V);
+ gsl_vector_free(S);
+ free(XiPhiR);
+ free(YiRhoR);
+ free(Xr);
+ free(Yr);
+ free(tXrXr);
+ free(tXrYr);
+ free(Z);
+}
projects / valse.git / blobdiff