| 1 | #include "utils.h" |
| 2 | #include <stdlib.h> |
| 3 | #include <math.h> |
| 4 | #include <gsl/gsl_linalg.h> |
| 5 | |
| 6 | // TODO: don't recompute indexes ai(...) and mi(...) when possible |
| 7 | void EMGLLF_core( |
| 8 | // IN parameters |
| 9 | const Real* phiInit, // parametre initial de moyenne renormalisé |
| 10 | const Real* rhoInit, // parametre initial de variance renormalisé |
| 11 | const Real* piInit, // parametre initial des proportions |
| 12 | const Real* gamInit, // paramètre initial des probabilités a posteriori de chaque échantillon |
| 13 | int mini, // nombre minimal d'itérations dans l'algorithme EM |
| 14 | int maxi, // nombre maximal d'itérations dans l'algorithme EM |
| 15 | Real gamma, // puissance des proportions dans la pénalisation pour un Lasso adaptatif |
| 16 | Real lambda, // valeur du paramètre de régularisation du Lasso |
| 17 | const Real* X, // régresseurs |
| 18 | const Real* Y, // réponse |
| 19 | Real eps, // seuil pour accepter la convergence |
| 20 | // OUT parameters (all pointers, to be modified) |
| 21 | Real* phi, // parametre de moyenne renormalisé, calculé par l'EM |
| 22 | Real* rho, // parametre de variance renormalisé, calculé par l'EM |
| 23 | Real* pi, // parametre des proportions renormalisé, calculé par l'EM |
| 24 | Real* llh, // (derniere) log vraisemblance associée à cet échantillon, |
| 25 | // pour les valeurs estimées des paramètres |
| 26 | Real* S, |
| 27 | int* affec, |
| 28 | // additional size parameters |
| 29 | int n, // nombre d'echantillons |
| 30 | int p, // nombre de covariables |
| 31 | int m, // taille de Y (multivarié) |
| 32 | int k) // nombre de composantes dans le mélange |
| 33 | { |
| 34 | //Initialize outputs |
| 35 | copyArray(phiInit, phi, p*m*k); |
| 36 | copyArray(rhoInit, rho, m*m*k); |
| 37 | copyArray(piInit, pi, k); |
| 38 | //S is already allocated, and doesn't need to be 'zeroed' |
| 39 | |
| 40 | //Other local variables: same as in R |
| 41 | Real* gam = (Real*)malloc(n*k*sizeof(Real)); |
| 42 | Real* logGam = (Real*)malloc(k*sizeof(Real)); |
| 43 | copyArray(gamInit, gam, n*k); |
| 44 | Real* Gram2 = (Real*)malloc(p*p*k*sizeof(Real)); |
| 45 | Real* ps2 = (Real*)malloc(p*m*k*sizeof(Real)); |
| 46 | Real* b = (Real*)malloc(k*sizeof(Real)); |
| 47 | Real* X2 = (Real*)malloc(n*p*k*sizeof(Real)); |
| 48 | Real* Y2 = (Real*)malloc(n*m*k*sizeof(Real)); |
| 49 | *llh = -INFINITY; |
| 50 | Real* pi2 = (Real*)malloc(k*sizeof(Real)); |
| 51 | // Additional (not at this place, in R file) |
| 52 | Real* gam2 = (Real*)malloc(k*sizeof(Real)); |
| 53 | Real* sqNorm2 = (Real*)malloc(k*sizeof(Real)); |
| 54 | Real* detRho = (Real*)malloc(k*sizeof(Real)); |
| 55 | gsl_matrix* matrix = gsl_matrix_alloc(m, m); |
| 56 | gsl_permutation* permutation = gsl_permutation_alloc(m); |
| 57 | Real* YiRhoR = (Real*)malloc(m*sizeof(Real)); |
| 58 | Real* XiPhiR = (Real*)malloc(m*sizeof(Real)); |
| 59 | const Real gaussConstM = pow(2.*M_PI,m/2.); |
| 60 | Real* Phi = (Real*)malloc(p*m*k*sizeof(Real)); |
| 61 | Real* Rho = (Real*)malloc(m*m*k*sizeof(Real)); |
| 62 | Real* Pi = (Real*)malloc(k*sizeof(Real)); |
| 63 | |
| 64 | for (int ite=1; ite<=maxi; ite++) |
| 65 | { |
| 66 | copyArray(phi, Phi, p*m*k); |
| 67 | copyArray(rho, Rho, m*m*k); |
| 68 | copyArray(pi, Pi, k); |
| 69 | |
| 70 | // Calculs associés a Y et X |
| 71 | for (int r=0; r<k; r++) |
| 72 | { |
| 73 | for (int mm=0; mm<m; mm++) |
| 74 | { |
| 75 | //Y2[,mm,r] = sqrt(gam[,r]) * Y[,mm] |
| 76 | for (int u=0; u<n; u++) |
| 77 | Y2[ai(u,mm,r,n,m,k)] = sqrt(gam[mi(u,r,n,k)]) * Y[mi(u,mm,n,m)]; |
| 78 | } |
| 79 | for (int i=0; i<n; i++) |
| 80 | { |
| 81 | //X2[i,,r] = sqrt(gam[i,r]) * X[i,] |
| 82 | for (int u=0; u<p; u++) |
| 83 | X2[ai(i,u,r,n,p,k)] = sqrt(gam[mi(i,r,n,k)]) * X[mi(i,u,n,p)]; |
| 84 | } |
| 85 | for (int mm=0; mm<m; mm++) |
| 86 | { |
| 87 | //ps2[,mm,r] = crossprod(X2[,,r],Y2[,mm,r]) |
| 88 | for (int u=0; u<p; u++) |
| 89 | { |
| 90 | Real dotProduct = 0.; |
| 91 | for (int v=0; v<n; v++) |
| 92 | dotProduct += X2[ai(v,u,r,n,p,k)] * Y2[ai(v,mm,r,n,m,k)]; |
| 93 | ps2[ai(u,mm,r,p,m,k)] = dotProduct; |
| 94 | } |
| 95 | } |
| 96 | for (int j=0; j<p; j++) |
| 97 | { |
| 98 | for (int s=0; s<p; s++) |
| 99 | { |
| 100 | //Gram2[j,s,r] = crossprod(X2[,j,r], X2[,s,r]) |
| 101 | Real dotProduct = 0.; |
| 102 | for (int u=0; u<n; u++) |
| 103 | dotProduct += X2[ai(u,j,r,n,p,k)] * X2[ai(u,s,r,n,p,k)]; |
| 104 | Gram2[ai(j,s,r,p,p,k)] = dotProduct; |
| 105 | } |
| 106 | } |
| 107 | } |
| 108 | |
| 109 | ///////////// |
| 110 | // Etape M // |
| 111 | ///////////// |
| 112 | |
| 113 | // Pour pi |
| 114 | for (int r=0; r<k; r++) |
| 115 | { |
| 116 | //b[r] = sum(abs(phi[,,r])) |
| 117 | Real sumAbsPhi = 0.; |
| 118 | for (int u=0; u<p; u++) |
| 119 | for (int v=0; v<m; v++) |
| 120 | sumAbsPhi += fabs(phi[ai(u,v,r,p,m,k)]); |
| 121 | b[r] = sumAbsPhi; |
| 122 | } |
| 123 | //gam2 = colSums(gam) |
| 124 | for (int u=0; u<k; u++) |
| 125 | { |
| 126 | Real sumOnColumn = 0.; |
| 127 | for (int v=0; v<n; v++) |
| 128 | sumOnColumn += gam[mi(v,u,n,k)]; |
| 129 | gam2[u] = sumOnColumn; |
| 130 | } |
| 131 | //a = sum(gam %*% log(pi)) |
| 132 | Real a = 0.; |
| 133 | for (int u=0; u<n; u++) |
| 134 | { |
| 135 | Real dotProduct = 0.; |
| 136 | for (int v=0; v<k; v++) |
| 137 | dotProduct += gam[mi(u,v,n,k)] * log(pi[v]); |
| 138 | a += dotProduct; |
| 139 | } |
| 140 | |
| 141 | //tant que les proportions sont negatives |
| 142 | int kk = 0, |
| 143 | pi2AllPositive = 0; |
| 144 | Real invN = 1./n; |
| 145 | while (!pi2AllPositive) |
| 146 | { |
| 147 | //pi2 = pi + 0.1^kk * ((1/n)*gam2 - pi) |
| 148 | Real pow_01_kk = pow(0.1,kk); |
| 149 | for (int r=0; r<k; r++) |
| 150 | pi2[r] = pi[r] + pow_01_kk * (invN*gam2[r] - pi[r]); |
| 151 | //pi2AllPositive = all(pi2 >= 0) |
| 152 | pi2AllPositive = 1; |
| 153 | for (int r=0; r<k; r++) |
| 154 | { |
| 155 | if (pi2[r] < 0) |
| 156 | { |
| 157 | pi2AllPositive = 0; |
| 158 | break; |
| 159 | } |
| 160 | } |
| 161 | kk++; |
| 162 | } |
| 163 | |
| 164 | //sum(pi^gamma * b) |
| 165 | Real piPowGammaDotB = 0.; |
| 166 | for (int v=0; v<k; v++) |
| 167 | piPowGammaDotB += pow(pi[v],gamma) * b[v]; |
| 168 | //sum(pi2^gamma * b) |
| 169 | Real pi2PowGammaDotB = 0.; |
| 170 | for (int v=0; v<k; v++) |
| 171 | pi2PowGammaDotB += pow(pi2[v],gamma) * b[v]; |
| 172 | //sum(gam2 * log(pi2)) |
| 173 | Real gam2DotLogPi2 = 0.; |
| 174 | for (int v=0; v<k; v++) |
| 175 | gam2DotLogPi2 += gam2[v] * log(pi2[v]); |
| 176 | |
| 177 | //t(m) la plus grande valeur dans la grille O.1^k tel que ce soit décroissante ou constante |
| 178 | while (-invN*a + lambda*piPowGammaDotB < -invN*gam2DotLogPi2 + lambda*pi2PowGammaDotB |
| 179 | && kk<1000) |
| 180 | { |
| 181 | Real pow_01_kk = pow(0.1,kk); |
| 182 | //pi2 = pi + 0.1^kk * (1/n*gam2 - pi) |
| 183 | for (int v=0; v<k; v++) |
| 184 | pi2[v] = pi[v] + pow_01_kk * (invN*gam2[v] - pi[v]); |
| 185 | //pi2 was updated, so we recompute pi2PowGammaDotB and gam2DotLogPi2 |
| 186 | pi2PowGammaDotB = 0.; |
| 187 | for (int v=0; v<k; v++) |
| 188 | pi2PowGammaDotB += pow(pi2[v],gamma) * b[v]; |
| 189 | gam2DotLogPi2 = 0.; |
| 190 | for (int v=0; v<k; v++) |
| 191 | gam2DotLogPi2 += gam2[v] * log(pi2[v]); |
| 192 | kk++; |
| 193 | } |
| 194 | Real t = pow(0.1,kk); |
| 195 | //sum(pi + t*(pi2-pi)) |
| 196 | Real sumPiPlusTbyDiff = 0.; |
| 197 | for (int v=0; v<k; v++) |
| 198 | sumPiPlusTbyDiff += (pi[v] + t*(pi2[v] - pi[v])); |
| 199 | //pi = (pi + t*(pi2-pi)) / sum(pi + t*(pi2-pi)) |
| 200 | for (int v=0; v<k; v++) |
| 201 | pi[v] = (pi[v] + t*(pi2[v] - pi[v])) / sumPiPlusTbyDiff; |
| 202 | |
| 203 | //Pour phi et rho |
| 204 | for (int r=0; r<k; r++) |
| 205 | { |
| 206 | for (int mm=0; mm<m; mm++) |
| 207 | { |
| 208 | Real ps = 0., |
| 209 | nY2 = 0.; |
| 210 | // Compute ps, and nY2 = sum(Y2[,mm,r]^2) |
| 211 | for (int i=0; i<n; i++) |
| 212 | { |
| 213 | //< X2[i,,r] , phi[,mm,r] > |
| 214 | Real dotProduct = 0.; |
| 215 | for (int u=0; u<p; u++) |
| 216 | dotProduct += X2[ai(i,u,r,n,p,k)] * phi[ai(u,mm,r,p,m,k)]; |
| 217 | //ps = ps + Y2[i,mm,r] * sum(X2[i,,r] * phi[,mm,r]) |
| 218 | ps += Y2[ai(i,mm,r,n,m,k)] * dotProduct; |
| 219 | nY2 += Y2[ai(i,mm,r,n,m,k)] * Y2[ai(i,mm,r,n,m,k)]; |
| 220 | } |
| 221 | //rho[mm,mm,r] = (ps+sqrt(ps^2+4*nY2*gam2[r])) / (2*nY2) |
| 222 | rho[ai(mm,mm,r,m,m,k)] = (ps + sqrt(ps*ps + 4*nY2 * gam2[r])) / (2*nY2); |
| 223 | } |
| 224 | } |
| 225 | |
| 226 | for (int r=0; r<k; r++) |
| 227 | { |
| 228 | for (int j=0; j<p; j++) |
| 229 | { |
| 230 | for (int mm=0; mm<m; mm++) |
| 231 | { |
| 232 | //sum(phi[-j,mm,r] * Gram2[j,-j,r]) |
| 233 | Real phiDotGram2 = 0.; |
| 234 | for (int u=0; u<p; u++) |
| 235 | { |
| 236 | if (u != j) |
| 237 | phiDotGram2 += phi[ai(u,mm,r,p,m,k)] * Gram2[ai(j,u,r,p,p,k)]; |
| 238 | } |
| 239 | //S[j,mm,r] = -rho[mm,mm,r]*ps2[j,mm,r] + sum(phi[-j,mm,r] * Gram2[j,-j,r]) |
| 240 | S[ai(j,mm,r,p,m,k)] = -rho[ai(mm,mm,r,m,m,k)] * ps2[ai(j,mm,r,p,m,k)] |
| 241 | + phiDotGram2; |
| 242 | Real pirPowGamma = pow(pi[r],gamma); |
| 243 | if (fabs(S[ai(j,mm,r,p,m,k)]) <= n*lambda*pirPowGamma) |
| 244 | phi[ai(j,mm,r,p,m,k)] = 0.; |
| 245 | else if (S[ai(j,mm,r,p,m,k)] > n*lambda*pirPowGamma) |
| 246 | { |
| 247 | phi[ai(j,mm,r,p,m,k)] = (n*lambda*pirPowGamma - S[ai(j,mm,r,p,m,k)]) |
| 248 | / Gram2[ai(j,j,r,p,p,k)]; |
| 249 | } |
| 250 | else |
| 251 | { |
| 252 | phi[ai(j,mm,r,p,m,k)] = -(n*lambda*pirPowGamma + S[ai(j,mm,r,p,m,k)]) |
| 253 | / Gram2[ai(j,j,r,p,p,k)]; |
| 254 | } |
| 255 | } |
| 256 | } |
| 257 | } |
| 258 | |
| 259 | ///////////// |
| 260 | // Etape E // |
| 261 | ///////////// |
| 262 | |
| 263 | // Precompute det(rho[,,r]) for r in 1...k |
| 264 | int signum; |
| 265 | for (int r=0; r<k; r++) |
| 266 | { |
| 267 | for (int u=0; u<m; u++) |
| 268 | { |
| 269 | for (int v=0; v<m; v++) |
| 270 | matrix->data[u*m+v] = rho[ai(u,v,r,m,m,k)]; |
| 271 | } |
| 272 | gsl_linalg_LU_decomp(matrix, permutation, &signum); |
| 273 | detRho[r] = gsl_linalg_LU_det(matrix, signum); |
| 274 | } |
| 275 | |
| 276 | Real sumLogLLH = 0.; |
| 277 | for (int i=0; i<n; i++) |
| 278 | { |
| 279 | for (int r=0; r<k; r++) |
| 280 | { |
| 281 | //compute Y[i,]%*%rho[,,r] |
| 282 | for (int u=0; u<m; u++) |
| 283 | { |
| 284 | YiRhoR[u] = 0.; |
| 285 | for (int v=0; v<m; v++) |
| 286 | YiRhoR[u] += Y[mi(i,v,n,m)] * rho[ai(v,u,r,m,m,k)]; |
| 287 | } |
| 288 | |
| 289 | //compute X[i,]%*%phi[,,r] |
| 290 | for (int u=0; u<m; u++) |
| 291 | { |
| 292 | XiPhiR[u] = 0.; |
| 293 | for (int v=0; v<p; v++) |
| 294 | XiPhiR[u] += X[mi(i,v,n,p)] * phi[ai(v,u,r,p,m,k)]; |
| 295 | } |
| 296 | |
| 297 | //compute sq norm || Y(:,i)*rho(:,:,r)-X(i,:)*phi(:,:,r) ||_2^2 |
| 298 | sqNorm2[r] = 0.; |
| 299 | for (int u=0; u<m; u++) |
| 300 | sqNorm2[r] += (YiRhoR[u]-XiPhiR[u]) * (YiRhoR[u]-XiPhiR[u]); |
| 301 | } |
| 302 | |
| 303 | // Update gam[,]; use log to avoid numerical problems |
| 304 | Real maxLogGam = -INFINITY; |
| 305 | for (int r=0; r<k; r++) |
| 306 | { |
| 307 | logGam[r] = log(pi[r]) - .5 * sqNorm2[r] + log(detRho[r]); |
| 308 | if (maxLogGam < logGam[r]) |
| 309 | maxLogGam = logGam[r]; |
| 310 | } |
| 311 | Real norm_fact = 0.; |
| 312 | for (int r=0; r<k; r++) |
| 313 | { |
| 314 | logGam[r] = logGam[r] - maxLogGam; //adjust without changing proportions |
| 315 | gam[mi(i,r,n,k)] = exp(logGam[r]); //gam[i, ] <- exp(logGam) |
| 316 | norm_fact += gam[mi(i,r,n,k)]; //norm_fact <- sum(gam[i, ]) |
| 317 | } |
| 318 | // gam[i, ] <- gam[i, ] / norm_fact |
| 319 | for (int r=0; r<k; r++) |
| 320 | gam[mi(i,r,n,k)] /= norm_fact; |
| 321 | |
| 322 | sumLogLLH += log(norm_fact) - log(gaussConstM); |
| 323 | } |
| 324 | |
| 325 | //sumPen = sum(pi^gamma * b) |
| 326 | Real sumPen = 0.; |
| 327 | for (int r=0; r<k; r++) |
| 328 | sumPen += pow(pi[r],gamma) * b[r]; |
| 329 | Real last_llh = *llh; |
| 330 | //llh = -sumLogLLH/n #+ lambda*sumPen |
| 331 | *llh = -invN * sumLogLLH; //+ lambda * sumPen; |
| 332 | Real dist = ( ite==1 ? *llh : (*llh - last_llh) / (1. + fabs(*llh)) ); |
| 333 | |
| 334 | //Dist1 = max( abs(phi-Phi) / (1+abs(phi)) ) |
| 335 | Real Dist1 = 0.; |
| 336 | for (int u=0; u<p; u++) |
| 337 | { |
| 338 | for (int v=0; v<m; v++) |
| 339 | { |
| 340 | for (int w=0; w<k; w++) |
| 341 | { |
| 342 | Real tmpDist = fabs(phi[ai(u,v,w,p,m,k)]-Phi[ai(u,v,w,p,m,k)]) |
| 343 | / (1.+fabs(phi[ai(u,v,w,p,m,k)])); |
| 344 | if (tmpDist > Dist1) |
| 345 | Dist1 = tmpDist; |
| 346 | } |
| 347 | } |
| 348 | } |
| 349 | //Dist2 = max( (abs(rho-Rho)) / (1+abs(rho)) ) |
| 350 | Real Dist2 = 0.; |
| 351 | for (int u=0; u<m; u++) |
| 352 | { |
| 353 | for (int v=0; v<m; v++) |
| 354 | { |
| 355 | for (int w=0; w<k; w++) |
| 356 | { |
| 357 | Real tmpDist = fabs(rho[ai(u,v,w,m,m,k)]-Rho[ai(u,v,w,m,m,k)]) |
| 358 | / (1.+fabs(rho[ai(u,v,w,m,m,k)])); |
| 359 | if (tmpDist > Dist2) |
| 360 | Dist2 = tmpDist; |
| 361 | } |
| 362 | } |
| 363 | } |
| 364 | //Dist3 = max( (abs(pi-Pi)) / (1+abs(Pi))) |
| 365 | Real Dist3 = 0.; |
| 366 | for (int u=0; u<n; u++) |
| 367 | { |
| 368 | for (int v=0; v<k; v++) |
| 369 | { |
| 370 | Real tmpDist = fabs(pi[v]-Pi[v]) / (1.+fabs(pi[v])); |
| 371 | if (tmpDist > Dist3) |
| 372 | Dist3 = tmpDist; |
| 373 | } |
| 374 | } |
| 375 | //dist2=max([max(Dist1),max(Dist2),max(Dist3)]); |
| 376 | Real dist2 = Dist1; |
| 377 | if (Dist2 > dist2) |
| 378 | dist2 = Dist2; |
| 379 | if (Dist3 > dist2) |
| 380 | dist2 = Dist3; |
| 381 | |
| 382 | if (ite >= mini && (dist >= eps || dist2 >= sqrt(eps))) |
| 383 | break; |
| 384 | } |
| 385 | |
| 386 | //affec = apply(gam, 1, which.max) |
| 387 | for (int i=0; i<n; i++) |
| 388 | { |
| 389 | Real rowMax = 0.; |
| 390 | affec[i] = 0; |
| 391 | for (int j=0; j<k; j++) |
| 392 | { |
| 393 | if (gam[mi(i,j,n,k)] > rowMax) |
| 394 | { |
| 395 | affec[i] = j+1; //R indices start at 1 |
| 396 | rowMax = gam[mi(i,j,n,k)]; |
| 397 | } |
| 398 | } |
| 399 | } |
| 400 | |
| 401 | //free memory |
| 402 | free(b); |
| 403 | free(gam); |
| 404 | free(logGam); |
| 405 | free(Phi); |
| 406 | free(Rho); |
| 407 | free(Pi); |
| 408 | free(Gram2); |
| 409 | free(ps2); |
| 410 | free(detRho); |
| 411 | gsl_matrix_free(matrix); |
| 412 | gsl_permutation_free(permutation); |
| 413 | free(XiPhiR); |
| 414 | free(YiRhoR); |
| 415 | free(gam2); |
| 416 | free(pi2); |
| 417 | free(X2); |
| 418 | free(Y2); |
| 419 | free(sqNorm2); |
| 420 | }\f |