From 435cb8419ebcdb624aa053f351c981133d58d6b6 Mon Sep 17 00:00:00 2001 From: Benjamin Auder <benjamin.auder@somewhere> Date: Sun, 2 Apr 2017 11:11:35 +0200 Subject: [PATCH] Fix errors in EMGLLF.c :) --- pkg/src/sources/EMGLLF.c | 55 +++++++++++++++++--------------- test/generate_test_data/EMGLLF.R | 34 ++++++++------------ 2 files changed, 43 insertions(+), 46 deletions(-) diff --git a/pkg/src/sources/EMGLLF.c b/pkg/src/sources/EMGLLF.c index 7ce6ded..520d57b 100644 --- a/pkg/src/sources/EMGLLF.c +++ b/pkg/src/sources/EMGLLF.c @@ -77,7 +77,7 @@ void EMGLLF_core( { //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++) { @@ -164,33 +164,34 @@ void EMGLLF_core( 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); @@ -209,7 +210,7 @@ void EMGLLF_core( { 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)]; @@ -221,16 +222,17 @@ void EMGLLF_core( 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.; + //nY2[mm,r] = sum(Y2[,mm,r]^2) + Real sumY2 = 0.; for (int u=0; u<n; u++) - sumNy2 += Y2[ai(u,mm,r,n,m,k)]; - nY2[mi(mm,r,m,k)] = sumNy2; + sumY2 += Y2[ai(u,mm,r,n,m,k)] * Y2[ai(u,mm,r,n,m,k)]; + nY2[mi(mm,r,m,k)] = sumY2; //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)]); } } + for (int r=0; r<k; r++) { for (int j=0; j<p; j++) @@ -238,25 +240,25 @@ void EMGLLF_core( for (int mm=0; mm<m; mm++) { //sum(phi[-j,mm,r] * Gram2[j, setdiff(1:p,j),r]) - Real dotPhiGram2 = 0.0; + 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)]; } } @@ -281,7 +283,7 @@ void EMGLLF_core( 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.; @@ -311,6 +313,7 @@ void EMGLLF_core( sumLLF1 += Gam[mi(i,r,n,k)] / gaussConstM; sumGamI += Gam[mi(i,r,n,k)]; } + sumLogLLF2 += log(sumLLF1); for (int r=0; r<k; r++) { diff --git a/test/generate_test_data/EMGLLF.R b/test/generate_test_data/EMGLLF.R index 37859d8..fce0a8f 100644 --- a/test/generate_test_data/EMGLLF.R +++ b/test/generate_test_data/EMGLLF.R @@ -50,17 +50,17 @@ EMGLLF_R = function(phiInit,rhoInit,piInit,gamInit,mini,maxi,gamma,lambda,X,Y,ta Gram2[j,s,r] = crossprod(X2[,j,r], X2[,s,r]) } } - + ########## #Etape M # ########## #pour pi - for (r in 1:k){ - b[r] = sum(abs(phi[,,r]))} + for (r in 1:k) + b[r] = sum(abs(phi[,,r])) gam2 = colSums(gam) a = sum(gam %*% log(pi)) - + #tant que les props sont negatives kk = 0 pi2AllPositive = FALSE @@ -71,8 +71,7 @@ EMGLLF_R = function(phiInit,rhoInit,piInit,gamInit,mini,maxi,gamma,lambda,X,Y,ta kk = kk+1 } -#if (ite==2) browser() - #t[m] la plus grande valeur dans la grille O.1^k tel que ce soit décroissante ou constante + #t(m) la plus grande valeur dans la grille O.1^k tel que ce soit décroissante ou constante while( kk < 1000 && -a/n + lambda * sum(pi^gamma * b) < -sum(gam2 * log(pi2))/n + lambda * sum(pi2^gamma * b) ) { @@ -81,7 +80,7 @@ EMGLLF_R = function(phiInit,rhoInit,piInit,gamInit,mini,maxi,gamma,lambda,X,Y,ta } t = 0.1^kk pi = (pi + t*(pi2-pi)) / sum(pi + t*(pi2-pi)) - + #Pour phi et rho for (r in 1:k) { @@ -94,18 +93,17 @@ EMGLLF_R = function(phiInit,rhoInit,piInit,gamInit,mini,maxi,gamma,lambda,X,Y,ta ps[mm,r] = sum(ps1[,mm,r]) nY2[mm,r] = sum(Y2[,mm,r]^2) rho[mm,mm,r] = (ps[mm,r]+sqrt(ps[mm,r]^2+4*nY2[mm,r]*gam2[r])) / (2*nY2[mm,r]) - } + } } + for (r in 1:k) { for (j in 1:p) { for (mm in 1:m) { - S[j,mm,r] = -rho[mm,mm,r]*ps2[j,mm,r] + sum(phi[-j,mm,r] * Gram2[j, setdiff(1:p, j),r]) -# (if(j>1) sum(phi[1:(j-1),mm,r] * Gram2[j,1:(j-1),r]) else 0) + -# (if(j<p) sum(phi[(j+1):p,mm,r] * Gram2[j,(j+1):p,r]) else 0) - if (abs(S[j,mm,r]) <= n*lambda*(pi[r]^gamma)) + S[j,mm,r] = -rho[mm,mm,r]*ps2[j,mm,r] + sum(phi[-j,mm,r] * Gram2[j,-j,r]) + if (abs(S[j,mm,r]) <= n*lambda*(pi[r]^gamma)) phi[j,mm,r]=0 else if(S[j,mm,r] > n*lambda*(pi[r]^gamma)) phi[j,mm,r] = (n*lambda*(pi[r]^gamma)-S[j,mm,r]) / Gram2[j,j,r] @@ -124,16 +122,14 @@ EMGLLF_R = function(phiInit,rhoInit,piInit,gamInit,mini,maxi,gamma,lambda,X,Y,ta { #precompute sq norms to numerically adjust their values sqNorm2 = rep(0,k) - for (r in 1:k){ - sqNorm2[r] = sum( (Y[i,]%*%rho[,,r]-X[i,]%*%phi[,,r])^2 )} + for (r in 1:k) + sqNorm2[r] = sum( (Y[i,]%*%rho[,,r]-X[i,]%*%phi[,,r])^2 ) - #compute Gam(:,:) + #compute Gam[,] sumLLF1 = 0.0; for (r in 1:k) { - #FIXME: numerical problems, because 0 < det(Rho[,,r] < EPS; what to do ?! - # consequence: error in while() at line 77 - Gam[i,r] = pi[r] * exp(-0.5*sqNorm2[r])* det(rho[,,r]) + Gam[i,r] = pi[r] * exp(-0.5*sqNorm2[r]) * det(rho[,,r]) sumLLF1 = sumLLF1 + Gam[i,r] / (2*base::pi)^(m/2) } sumLogLLF2 = sumLogLLF2 + log(sumLLF1) @@ -146,9 +142,7 @@ EMGLLF_R = function(phiInit,rhoInit,piInit,gamInit,mini,maxi,gamma,lambda,X,Y,ta sumPen = sum(pi^gamma * b) LLF[ite] = -sumLogLLF2/n + lambda*sumPen - dist = ifelse( ite == 1, LLF[ite], (LLF[ite]-LLF[ite-1]) / (1+abs(LLF[ite])) ) - Dist1 = max( (abs(phi-Phi)) / (1+abs(phi)) ) Dist2 = max( (abs(rho-Rho)) / (1+abs(rho)) ) Dist3 = max( (abs(pi-Pi)) / (1+abs(Pi)) ) -- 2.44.0