pkg/R/selectVariables.R

   1 #' selectVariables
   2 #'
   3 #' For a given lambda, construct the sets of relevant variables for each cluster.
   4 #'
   5 #' @param phiInit an initial estimator for phi (size: p*m*k)
   6 #' @param rhoInit an initial estimator for rho (size: m*m*k)
   7 #' @param piInit an initial estimator for pi (size : k)
   8 #' @param gamInit an initial estimator for gamma
   9 #' @param mini  minimum number of iterations in EM algorithm
  10 #' @param maxi  maximum number of iterations in EM algorithm
  11 #' @param gamma  power in the penalty
  12 #' @param glambda grid of regularization parameters
  13 #' @param X    matrix of regressors
  14 #' @param Y    matrix of responses
  15 #' @param thresh real, threshold to say a variable is relevant, by default = 1e-8
  16 #' @param eps   threshold to say that EM algorithm has converged
  17 #' @param ncores Number or cores for parallel execution (1 to disable)
  18 #' @param fast boolean to enable or not the C function call
  19 #'
  20 #' @return a list of outputs, for each lambda in grid: selected,Rho,Pi
  21 #'
  22 #' @export
  23 selectVariables <- function(phiInit, rhoInit, piInit, gamInit, mini, maxi, gamma,
  24   glambda, X, Y, thresh = 1e-08, eps, ncores = 3, fast)
  25 {
  26   if (ncores > 1) {
  27     cl <- parallel::makeCluster(ncores, outfile = "")
  28     parallel::clusterExport(cl = cl, varlist = c("phiInit", "rhoInit", "gamInit",
  29       "mini", "maxi", "glambda", "X", "Y", "thresh", "eps"), envir = environment())
  30   }
  31
  32   # Computation for a fixed lambda
  33   computeCoefs <- function(lambda)
  34   {
  35     params <- EMGLLF(phiInit, rhoInit, piInit, gamInit, mini, maxi, gamma, lambda,
  36       X, Y, eps, fast)
  37
  38     p <- ncol(X)
  39     m <- ncol(Y)
  40
  41     # selectedVariables: list where element j contains vector of selected variables
  42     # in [1,m]
  43     selectedVariables <- lapply(1:p, function(j) {
  44       # from boolean matrix mxk of selected variables obtain the corresponding boolean
  45       # m-vector, and finally return the corresponding indices
  46       if (m>1) {
  47         seq_len(m)[apply(abs(params$phi[j, , ]) > thresh, 1, any)]
  48       } else {
  49         if (any(params$phi[j, , ] > thresh))
  50           1
  51         else
  52           numeric(0)
  53       }
  54     })
  55
  56     list(selected = selectedVariables, Rho = params$rho, Pi = params$pi)
  57   }
  58
  59   # For each lambda in the grid, we compute the coefficients
  60   out <-
  61     if (ncores > 1) {
  62       parLapply(cl, glambda, computeCoefs)
  63     } else {
  64       lapply(glambda, computeCoefs)
  65     }
  66   if (ncores > 1)
  67     parallel::stopCluster(cl)
  68
  69   # Suppress models which are computed twice
  70   # sha1_array <- lapply(out, digest::sha1) out[ duplicated(sha1_array) ]
  71   selec <- lapply(out, function(model) model$selected)
  72   ind_dup <- duplicated(selec)
  73   ind_uniq <- which(!ind_dup)
  74   out2 <- list()
  75   for (l in 1:length(ind_uniq))
  76     out2[[l]] <- out[[ind_uniq[l]]]
  77   out2
  78 }