pkg/R/multiRun.R

   1 #' multiRun
   2 #'
   3 #' Estimate N times some parameters, outputs of some list of functions. This method is
   4 #' thus very generic, allowing typically bootstrap or Monte-Carlo estimations of matrices
   5 #' μ or β. Passing a list of functions opens the possibility to compare them on a fair
   6 #' basis (exact same inputs). It's even possible to compare methods on some deterministic
   7 #' design of experiments.
   8 #'
   9 #' @param fargs List of arguments for the estimation functions
  10 #' @param estimParams List of nf function(s) to apply on fargs - shared signature
  11 #' @param prepareArgs Prepare arguments for the functions inside estimParams
  12 #' @param N Number of runs
  13 #' @param ncores Number of cores for parallel runs (<=1: sequential)
  14 #' @param agg Aggregation method (default: lapply)
  15 #' @param verbose TRUE to indicate runs + methods numbers
  16 #'
  17 #' @return A list of nf aggregates of N results (matrices).
  18 #'
  19 #' @examples
  20 #' \donttest{
  21 #' β <- matrix(c(1,-2,3,1),ncol=2)
  22 #'
  23 #' # Bootstrap + computeMu, morpheus VS flexmix ; assumes fargs first 3 elts X,Y,K
  24 #' io <- generateSampleIO(n=1000, p=1/2, β=β, b=c(0,0), "logit")
  25 #' μ <- normalize(β)
  26 #' res <- multiRun(list(X=io$X,Y=io$Y,optargs=list(K=2)), list(
  27 #'   # morpheus
  28 #'   function(fargs) {
  29 #'     library(morpheus)
  30 #'     ind <- fargs$ind
  31 #'     computeMu(fargs$X[ind,],fargs$Y[ind],fargs$optargs)
  32 #'   },
  33 #'   # flexmix
  34 #'   function(fargs) {
  35 #'     library(flexmix)
  36 #'     ind <- fargs$ind
  37 #'     K <- fargs$optargs$K
  38 #'     dat = as.data.frame( cbind(fargs$Y[ind],fargs$X[ind,]) )
  39 #'     out = refit( flexmix( cbind(V1, 1 - V1) ~ 0+., data=dat, k=K,
  40 #'       model=FLXMRglm(family="binomial") ) )
  41 #'     normalize( matrix(out@@coef[1:(ncol(fargs$X)*K)], ncol=K) )
  42 #'   } ),
  43 #'   prepareArgs = function(fargs,index) {
  44 #'     if (index == 1)
  45 #'       fargs$ind <- 1:nrow(fargs$X)
  46 #'     else
  47 #'       fargs$ind <- sample(1:nrow(fargs$X),replace=TRUE)
  48 #'     fargs
  49 #'   }, N=10, ncores=3)
  50 #' for (i in 1:2)
  51 #'   res[[i]] <- alignMatrices(res[[i]], ref=μ, ls_mode="exact")
  52 #'
  53 #' # Monte-Carlo + optimParams from X,Y, morpheus VS flexmix ; first args n,p,β,b
  54 #' res <- multiRun(list(n=1000,p=1/2,β=β,b=c(0,0),optargs=list(link="logit")),list(
  55 #'   # morpheus
  56 #'   function(fargs) {
  57 #'     library(morpheus)
  58 #'     K <- fargs$optargs$K
  59 #'     μ <- computeMu(fargs$X, fargs$Y, fargs$optargs)
  60 #'     optimParams(fargs$K,fargs$link,fargs$optargs)$run(list(β=μ))$β
  61 #'   },
  62 #'   # flexmix
  63 #'   function(fargs) {
  64 #'     library(flexmix)
  65 #'     K <- fargs$optargs$K
  66 #'     dat <- as.data.frame( cbind(fargs$Y,fargs$X) )
  67 #'     out <- refit( flexmix( cbind(V1, 1 - V1) ~ 0+., data=dat, k=K,
  68 #'       model=FLXMRglm(family="binomial") ) )
  69 #'     sapply( seq_len(K), function(i) as.double( out@@components[[1]][[i]][,1] ) )
  70 #'   } ),
  71 #'   prepareArgs = function(fargs,index) {
  72 #'     library(morpheus)
  73 #'     io = generateSampleIO(fargs$n, fargs$p, fargs$β, fargs$b, fargs$optargs$link)
  74 #'     fargs$X = io$X
  75 #'     fargs$Y = io$Y
  76 #'     fargs$K = ncol(fargs$β)
  77 #'     fargs$link = fargs$optargs$link
  78 #'     fargs$optargs$M = computeMoments(io$X,io$Y)
  79 #'     fargs
  80 #'   }, N=10, ncores=3)
  81 #' for (i in 1:2)
  82 #'   res[[i]] <- alignMatrices(res[[i]], ref=β, ls_mode="exact")}
  83 #' @export
  84 multiRun <- function(fargs, estimParams,
  85   prepareArgs = function(x,i) x, N=10, ncores=3, agg=lapply, verbose=FALSE)
  86 {
  87   if (!is.list(fargs))
  88     stop("fargs: list")
  89   # No checks on fargs: supposedly done in estimParams[[i]]()
  90   if (!is.list(estimParams))
  91     estimParams = list(estimParams)
  92   # Verify that the provided parameters estimations are indeed functions
  93   lapply(seq_along(estimParams), function(i) {
  94     if (!is.function(estimParams[[i]]))
  95       stop("estimParams: list of function(fargs)")
  96   })
  97   if (!is.numeric(N) || N < 1)
  98     stop("N: positive integer")
  99
 100   estimParamAtIndex <- function(index)
 101   {
 102     fargs <- prepareArgs(fargs, index)
 103     if (verbose)
 104       cat("Run ",index,"\n")
 105     lapply(seq_along(estimParams), function(i) {
 106       if (verbose)
 107         cat("   Method ",i,"\n")
 108       out <- estimParams[[i]](fargs)
 109       if (is.list(out))
 110         do.call(rbind, out)
 111       else
 112         out
 113     })
 114   }
 115
 116   if (ncores > 1)
 117   {
 118     cl = parallel::makeCluster(ncores, outfile="")
 119     parallel::clusterExport(cl, c("fargs","verbose"), environment())
 120     list_res = parallel::clusterApplyLB(cl, 1:N, estimParamAtIndex)
 121     parallel::stopCluster(cl)
 122   }
 123   else
 124     list_res = lapply(1:N, estimParamAtIndex)
 125
 126   # De-interlace results: output one list per function
 127   nf <- length(estimParams)
 128   lapply( seq_len(nf), function(i) lapply(seq_len(N), function(j) list_res[[j]][[i]]) )
 129 }