X-Git-Url: https://git.auder.net/?p=valse.git;a=blobdiff_plain;f=pkg%2FR%2FconstructionModelesLassoMLE.R;h=692fbe190121eb99399b839376ed2c8e006d535c;hp=50879c935aea7e04042024e2f51a071cb554fa7f;hb=6af1d4897dbab92a7be05068e0e15823378965d9;hpb=f33f35efc9a01f93bb61959522d90ee6a76b892e

diff --git a/pkg/R/constructionModelesLassoMLE.R b/pkg/R/constructionModelesLassoMLE.R
index 50879c9..692fbe1 100644
--- a/pkg/R/constructionModelesLassoMLE.R
+++ b/pkg/R/constructionModelesLassoMLE.R
@@ -1,37 +1,117 @@
-constructionModelesLassoMLE = function(phiInit,rhoInit,piInit,gamInit,mini,maxi,gamma,glambda,
-	X,Y,seuil,tau,selected)
+#' constructionModelesLassoMLE
+#'
+#' Construct a collection of models with the Lasso-MLE procedure.
+#'
+#' @param phiInit an initialization for phi, get by initSmallEM.R
+#' @param rhoInit an initialization for rho, get by initSmallEM.R
+#' @param piInit an initialization for pi, get by initSmallEM.R
+#' @param gamInit an initialization for gam, get by initSmallEM.R
+#' @param mini integer, minimum number of iterations in the EM algorithm, by default = 10
+#' @param maxi integer, maximum number of iterations in the EM algorithm, by default = 100
+#' @param gamma integer for the power in the penaly, by default = 1
+#' @param X matrix of covariates (of size n*p)
+#' @param Y matrix of responses (of size n*m)
+#' @param eps real, threshold to say the EM algorithm converges, by default = 1e-4
+#' @param S output of selectVariables.R
+#' @param ncores Number of cores, by default = 3
+#' @param fast TRUE to use compiled C code, FALSE for R code only
+#' @param verbose TRUE to show some execution traces
+#'
+#' @return a list with several models, defined by phi (the regression parameter reparametrized),
+#' rho (the covariance parameter reparametrized), pi (the proportion parameter is the mixture model), llh
+#' (the value of the loglikelihood function for this estimator on the training dataset). The list is given
+#' for several levels of sparsity, given by several regularization parameters computed automatically.
+#'
+#' @export
+constructionModelesLassoMLE <- function(phiInit, rhoInit, piInit, gamInit, mini,
+  maxi, gamma, X, Y, eps, S, ncores, fast, verbose)
 {
-	#TODO: parameter ncores (chaque tâche peut aussi demander du parallélisme...)
-	cl = parallel::makeCluster( parallel::detectCores() / 4 )
-	parallel::clusterExport(cl=cl,
-		varlist=c("phiInit","rhoInit","gamInit","mini","maxi","glambda","X","Y","seuil","tau"),
-		envir=environment())
-	#Pour chaque lambda de la grille, on calcule les coefficients
-	out = parLapply( seq_along(glambda), function(lambdaindex)
-	{
-		n = dim(X)[1]
-		p = dim(phiInit)[1]
-		m = dim(phiInit)[2]
-		k = dim(phiInit)[3]
-
-		#TODO: phiInit[selected] et X[selected] sont bien sûr faux; par quoi remplacer ?
-		#lambda == 0 c'est normal ? -> ED : oui, ici on calcule le maximum de vraisembance, donc on ne pénalise plus
-    res = EMGLLF(phiInit[selected],rhoInit,piInit,gamInit,mini,maxi,gamma,0.,X[selected],Y,tau)
-
-		#comment évaluer la dimension à partir du résultat et de [not]selected ?
-    #dimension = ...
-
-    #on veut calculer la vraisemblance avec toutes nos estimations
-		densite = vector("double",n)
-		for (r in 1:k)
-		{
-			delta = Y%*%rho[,,r] - (X[selected]%*%res$phi[selected,,r])
-			densite = densite + pi[r] *
-				det(rho[,,r])/(sqrt(2*base::pi))^m * exp(-tcrossprod(delta)/2.0)
-		}
-		llh = c( sum(log(densite[,lambdaIndex])), (dimension+m+1)*k-1 )
-		list("phi"=res$phi, "rho"=res$rho, "pi"=res$pi, "llh" = llh)
-	})
-	parallel::stopCluster(cl)
-	out
+  if (ncores > 1)
+  {
+    cl <- parallel::makeCluster(ncores, outfile = "")
+    parallel::clusterExport(cl, envir = environment(), varlist = c("phiInit",
+      "rhoInit", "gamInit", "mini", "maxi", "gamma", "X", "Y", "eps", "S",
+      "ncores", "fast", "verbose"))
+  }
+
+  # Individual model computation
+  computeAtLambda <- function(lambda)
+  {
+    if (ncores > 1)
+      require("valse")  #nodes start with an empty environment
+
+    if (verbose)
+      print(paste("Computations for lambda=", lambda))
+
+    n <- nrow(X)
+    p <- ncol(X)
+    m <- ncol(Y)
+    k <- length(piInit)
+    sel.lambda <- S[[lambda]]$selected
+    # col.sel = which(colSums(sel.lambda)!=0) #if boolean matrix
+    col.sel <- which(sapply(sel.lambda, length) > 0)  #if list of selected vars
+    if (length(col.sel) == 0)
+      return(NULL)
+
+    # lambda == 0 because we compute the EMV: no penalization here
+    res <- EMGLLF(array(phiInit[col.sel, , ], dim=c(length(col.sel),m,k)),
+      rhoInit, piInit, gamInit, mini, maxi, gamma, 0,
+      as.matrix(X[, col.sel]), Y, eps, fast)
+
+    # Eval dimension from the result + selected
+    phiLambda2 <- res$phi
+    rhoLambda <- res$rho
+    piLambda <- res$pi
+    phiLambda <- array(0, dim = c(p, m, k))
+    for (j in seq_along(col.sel))
+      phiLambda[col.sel[j], sel.lambda[[j]], ] <- phiLambda2[j, sel.lambda[[j]], ]
+    dimension <- length(unlist(sel.lambda))
+
+    ## Affectations
+    Gam <- matrix(0, ncol = length(piLambda), nrow = n)
+    for (i in 1:n)
+    {
+      for (r in 1:length(piLambda))
+      {
+        sqNorm2 <- sum((Y[i, ] %*% rhoLambda[, , r] - X[i, ] %*% phiLambda[, , r])^2)
+        Gam[i, r] <- piLambda[r] * exp(-0.5 * sqNorm2) * det(rhoLambda[, , r])
+      }
+    }
+    Gam2 <- Gam/rowSums(Gam)
+    affec <- apply(Gam2, 1, which.max)
+    proba <- Gam2
+    LLH <- c(sum(log(apply(Gam,1,sum))), (dimension + m + 1) * k - 1)
+    # ## Computation of the loglikelihood
+    # # Precompute det(rhoLambda[,,r]) for r in 1...k
+    # detRho <- sapply(1:k, function(r) gdet(rhoLambda[, , r]))
+    # sumLogLLH <- 0
+    # for (i in 1:n)
+    # {
+    #   # Update gam[,]; use log to avoid numerical problems
+    #   logGam <- sapply(1:k, function(r) {
+    #     log(piLambda[r]) + log(detRho[r]) - 0.5 *
+    #       sum((Y[i, ] %*% rhoLambda[, , r] - X[i, ] %*% phiLambda[, , r])^2)
+    #   })
+    #
+    #   #logGam <- logGam - max(logGam) #adjust without changing proportions -> change the LLH
+    #   gam <- exp(logGam)
+    #   norm_fact <- sum(gam)
+    #   sumLogLLH <- sumLogLLH + log(norm_fact) - m/2* log(2 * base::pi)
+    # }
+    #llhLambda <- c(-sumLogLLH/n, (dimension + m + 1) * k - 1)
+    list(phi = phiLambda, rho = rhoLambda, pi = piLambda, llh = LLH, affec = affec, proba = proba)
+  }
+
+  # For each lambda, computation of the parameters
+  out <-
+    if (ncores > 1) {
+      parallel::parLapply(cl, 1:length(S), computeAtLambda)
+    } else {
+      lapply(1:length(S), computeAtLambda)
+    }
+
+  if (ncores > 1)
+    parallel::stopCluster(cl)
+
+  out
 }