projects / valse.git / blame
| Commit | Line | Data |
|---|---|---|
| c3b2c1ab BG |
1 | source("/home/goehry/Documents/valse/valse/R/vec_bin.R") |
| 2 | require(MASS) | |
| 3 | EMGrank = function(Pi, Rho, mini, maxi, X, Y, tau, rank){ | |
| c2028869 BG |
4 | #matrix dimensions |
| 5 | n = dim(X)[1] | |
| 6 | p = dim(X)[2] | |
| 7 | m = dim(Rho)[2] | |
| 8 | k = dim(Rho)[3] | |
| 9 | ||
| 10 | #init outputs | |
| 11 | phi = array(0, dim=c(p,m,k)) | |
| 12 | Z = rep(1, n) | |
| 13 | Pi = piInit | |
| 14 | LLF = 0 | |
| 15 | ||
| 16 | #local variables | |
| 17 | Phi = array(0, dim=c(p,m,k)) | |
| 18 | deltaPhi = c(0) | |
| 19 | sumDeltaPhi = 0 | |
| 20 | deltaPhiBufferSize = 20 | |
| 21 | ||
| 22 | #main loop | |
| 23 | ite = 1 | |
| 24 | while(ite<=mini || (ite<=maxi && sumDeltaPhi>tau)){ | |
| 25 | #M step: Mise à jour de Beta (et donc phi) | |
| 26 | for(r in 1:k){ | |
| 27 | Z_bin = vec_bin(Z,r) | |
| 28 | Z_vec = Z_bin$vec #vecteur 0 et 1 aux endroits o? Z==r | |
| 29 | Z_indice = Z_bin$indice | |
| 30 | if(sum(Z_indice) == 0){ | |
| 31 | next | |
| 32 | } | |
| 33 | #U,S,V = SVD of (t(Xr)Xr)^{-1} * t(Xr) * Yr | |
| c3b2c1ab BG |
34 | sv = svd(ginv( crossprod(X[Z_indice,]) ) %*% crossprod(X[Z_indice,], Y[Z_indice,]) ) |
| 35 | S = diag(sv$d) | |
| 36 | U = sv$u | |
| 37 | V = sv$v | |
| c2028869 BG |
38 | #Set m-rank(r) singular values to zero, and recompose |
| 39 | #best rank(r) approximation of the initial product | |
| c3b2c1ab BG |
40 | if(r==k){ |
| 41 | j_r_1 = length(S) | |
| 42 | } | |
| 43 | else{ | |
| 44 | j_r_1 = c(rank[r]+1:length(S)) | |
| 45 | } | |
| 46 | S[j_r_1] = 0 | |
| 47 | S = diag(S, nrow = ncol(U)) | |
| 48 | phi[,,r] = U %*% S %*% t(V) %*% Rho[,,r] | |
| c2028869 BG |
49 | } |
| 50 | ||
| 51 | #Etape E et calcul de LLF | |
| 52 | sumLogLLF2 = 0 | |
| 53 | for(i in 1:n){ | |
| 54 | sumLLF1 = 0 | |
| 55 | maxLogGamIR = -Inf | |
| 56 | for(r in 1:k){ | |
| 57 | dotProduct = tcrossprod(Y[i,]%*%Rho[,,r]-X[i,]%*%phi[,,r]) | |
| 58 | logGamIR = log(Pi[r]) + log(det(Rho[,,r])) - 0.5*dotProduct | |
| 59 | #Z[i] = index of max (gam[i,]) | |
| 60 | if(logGamIR > maxLogGamIR){ | |
| 61 | Z[i] = r | |
| 62 | maxLogGamIR = logGamIR | |
| 63 | } | |
| 64 | sumLLF1 = sumLLF1 + exp(logGamIR) / (2*pi)^(m/2) | |
| 65 | } | |
| 66 | sumLogLLF2 = sumLogLLF2 + log(sumLLF1) | |
| 67 | } | |
| 68 | ||
| 69 | LLF = -1/n * sumLogLLF2 | |
| 70 | ||
| 71 | #update distance parameter to check algorithm convergence (delta(phi, Phi)) | |
| 72 | deltaPhi = c(deltaPhi, max(max(max((abs(phi-Phi))/(1+abs(phi))))) ) | |
| 73 | if(length(deltaPhi) > deltaPhiBufferSize){ | |
| c3b2c1ab BG |
74 | l_1 = c(2:length(deltaPhi)) |
| 75 | deltaPhi = deltaPhi[l_1] | |
| c2028869 BG |
76 | } |
| 77 | sumDeltaPhi = sum(abs(deltaPhi)) | |
| 78 | ||
| 79 | #update other local variables | |
| 80 | Phi = phi | |
| 81 | ite = ite+1 | |
| 82 | ||
| 83 | } | |
| 84 | return(list(phi=phi, LLF=LLF)) | |
| 85 | } |