epclust/R/clustering.R

   1 #' @name clustering
   2 #' @rdname clustering
   3 #' @aliases clusteringTask1 computeClusters1 computeClusters2
   4 #'
   5 #' @title Two-stage clustering, withing one task (see \code{claws()})
   6 #'
   7 #' @description \code{clusteringTask1()} runs one full stage-1 task, which consists in
   8 #'   iterated stage 1 clustering (on nb_curves / ntasks energy contributions, computed
   9 #'   through discrete wavelets coefficients).
  10 #'   \code{clusteringTask2()} runs a full stage-2 task, which consists in synchrones
  11 #'   and then WER distances computations, before applying the clustering algorithm.
  12 #'   \code{computeClusters1()} and \code{computeClusters2()} correspond to the atomic
  13 #'   clustering procedures respectively for stage 1 and 2. The former applies the
  14 #'   clustering algorithm (PAM) on a contributions matrix, while the latter clusters
  15 #'   a chunk of series inside one task (~max nb_series_per_chunk)
  16 #'
  17 #' @param indices Range of series indices to cluster in parallel (initial data)
  18 #' @param getContribs Function to retrieve contributions from initial series indices:
  19 #'   \code{getContribs(indices)} outpus a contributions matrix
  20 #' @param contribs matrix of contributions (e.g. output of \code{curvesToContribs()})
  21 #' @inheritParams computeSynchrones
  22 #' @inheritParams claws
  23 #'
  24 #' @return For \code{clusteringTask1()} and \code{computeClusters1()}, the indices of the
  25 #'   computed (K1) medoids. Indices are irrelevant for stage 2 clustering, thus
  26 #'   \code{computeClusters2()} outputs a big.matrix of medoids
  27 #'   (of size limited by nb_series_per_chunk)
  28 NULL
  29
  30 #' @rdname clustering
  31 #' @export
  32 clusteringTask1 = function(
  33     indices, getContribs, K1, nb_series_per_chunk, ncores_clust=1, verbose=FALSE, parll=TRUE)
  34 {
  35     if (verbose)
  36         cat(paste("*** Clustering task on ",length(indices)," lines\n", sep=""))
  37
  38     wrapComputeClusters1 = function(inds) {
  39         if (parll)
  40             require("epclust", quietly=TRUE)
  41         if (verbose)
  42             cat(paste("   computeClusters1() on ",length(inds)," lines\n", sep=""))
  43         inds[ computeClusters1(getContribs(inds), K1) ]
  44     }
  45
  46     if (parll)
  47     {
  48         cl = parallel::makeCluster(ncores_clust)
  49         parallel::clusterExport(cl, varlist=c("getContribs","K1","verbose"), envir=environment())
  50     }
  51     while (length(indices) > K1)
  52     {
  53         indices_workers = .spreadIndices(indices, nb_series_per_chunk)
  54         if (parll)
  55             indices = unlist( parallel::parLapply(cl, indices_workers, wrapComputeClusters1) )
  56         else
  57             indices = unlist( lapply(indices_workers, wrapComputeClusters1) )
  58     }
  59     if (parll)
  60         parallel::stopCluster(cl)
  61
  62     indices #medoids
  63 }
  64
  65 #' @rdname clustering
  66 #' @export
  67 clusteringTask2 = function(medoids, K2,
  68     getRefSeries, nb_ref_curves, nb_series_per_chunk, ncores_clust=1,verbose=FALSE,parll=TRUE)
  69 {
  70     if (nrow(medoids) <= K2)
  71         return (medoids)
  72     synchrones = computeSynchrones(medoids,
  73         getRefSeries, nb_ref_curves, nb_series_per_chunk, ncores_clust, verbose, parll)
  74     distances = computeWerDists(synchrones, ncores_clust, verbose, parll)
  75     # PAM in package 'cluster' cannot take big.matrix in input: need to cast it
  76     mat_dists = matrix(nrow=K1, ncol=K1)
  77     for (i in seq_len(K1))
  78         mat_dists[i,] = distances[i,]
  79     medoids[ computeClusters2(mat_dists,K2), ]
  80 }
  81
  82 #' @rdname clustering
  83 #' @export
  84 computeClusters1 = function(contribs, K1)
  85     cluster::pam(contribs, K1, diss=FALSE)$id.med
  86
  87 #' @rdname clustering
  88 #' @export
  89 computeClusters2 = function(distances, K2)
  90     cluster::pam(distances, K2, diss=TRUE)$id.med
  91
  92 #' computeSynchrones
  93 #'
  94 #' Compute the synchrones curves (sum of clusters elements) from a matrix of medoids,
  95 #' using L2 distances.
  96 #'
  97 #' @param medoids big.matrix of medoids (curves of same length as initial series)
  98 #' @param getRefSeries Function to retrieve initial series (e.g. in stage 2 after series
  99 #'   have been replaced by stage-1 medoids)
 100 #' @param nb_ref_curves How many reference series? (This number is known at this stage)
 101 #' @inheritParams claws
 102 #'
 103 #' @return A big.matrix of size K1 x L where L = data_length
 104 #'
 105 #' @export
 106 computeSynchrones = function(medoids, getRefSeries,
 107     nb_ref_curves, nb_series_per_chunk, ncores_clust=1,verbose=FALSE,parll=TRUE)
 108 {
 109
 110
 111
 112 #TODO: si parll, getMedoids + serialization, pass only getMedoids to nodes
 113 # --> BOF... chaque node chargera tous les medoids (efficacité) :/ ==> faut que ça tienne en RAM
 114 #au pire :: C-ifier et charger medoids 1 by 1...
 115
 116     #MIEUX :: medoids DOIT etre une big.matrix partagée !
 117
 118     computeSynchronesChunk = function(indices)
 119     {
 120         if (verbose)
 121             cat(paste("--- Compute synchrones for ",length(indices)," lines\n", sep=""))
 122         ref_series = getRefSeries(indices)
 123         #get medoids indices for this chunk of series
 124         for (i in seq_len(nrow(ref_series)))
 125         {
 126             j = which.min( rowSums( sweep(medoids, 2, ref_series[i,], '-')^2 ) )
 127             if (parll)
 128                 synchronicity::lock(m)
 129             synchrones[j,] = synchrones[j,] + ref_series[i,]
 130             counts[j,1] = counts[j,1] + 1
 131             if (parll)
 132                 synchronicity::unlock(m)
 133         }
 134     }
 135
 136     K = nrow(medoids)
 137     # Use bigmemory (shared==TRUE by default) + synchronicity to fill synchrones in //
 138     # TODO: if size > RAM (not our case), use file-backed big.matrix
 139     synchrones = bigmemory::big.matrix(nrow=K,ncol=ncol(medoids),type="double",init=0.)
 140     counts = bigmemory::big.matrix(nrow=K,ncol=1,type="double",init=0)
 141     # synchronicity is only for Linux & MacOS; on Windows: run sequentially
 142     parll = (requireNamespace("synchronicity",quietly=TRUE)
 143         && parll && Sys.info()['sysname'] != "Windows")
 144     if (parll)
 145         m <- synchronicity::boost.mutex()
 146
 147     if (parll)
 148     {
 149         cl = parallel::makeCluster(ncores_clust)
 150         parallel::clusterExport(cl,
 151             varlist=c("synchrones","counts","verbose","medoids","getRefSeries"),
 152             envir=environment())
 153     }
 154
 155     indices_workers = .spreadIndices(seq_len(nb_ref_curves), nb_series_per_chunk)
 156     ignored <-
 157         if (parll)
 158             parallel::parLapply(indices_workers, computeSynchronesChunk)
 159         else
 160             lapply(indices_workers, computeSynchronesChunk)
 161
 162     if (parll)
 163         parallel::stopCluster(cl)
 164
 165     #TODO: can we avoid this loop? ( synchrones = sweep(synchrones, 1, counts, '/') )
 166     for (i in seq_len(K))
 167         synchrones[i,] = synchrones[i,] / counts[i,1]
 168     #NOTE: odds for some clusters to be empty? (when series already come from stage 2)
 169     #      ...maybe; but let's hope resulting K1' be still quite bigger than K2
 170     noNA_rows = sapply(seq_len(K), function(i) all(!is.nan(synchrones[i,])))
 171     if (all(noNA_rows))
 172         return (synchrones)
 173     # Else: some clusters are empty, need to slice synchrones
 174     synchrones[noNA_rows,]
 175 }
 176
 177 #' computeWerDists
 178 #'
 179 #' Compute the WER distances between the synchrones curves (in rows), which are
 180 #' returned (e.g.) by \code{computeSynchrones()}
 181 #'
 182 #' @param synchrones A big.matrix of synchrones, in rows. The series have same length
 183 #'   as the series in the initial dataset
 184 #' @inheritParams claws
 185 #'
 186 #' @return A big.matrix of size K1 x K1
 187 #'
 188 #' @export
 189 computeWerDists = function(synchrones, ncores_clust=1,verbose=FALSE,parll=TRUE)
 190 {
 191
 192
 193
 194 #TODO: re-organize to call computeWerDist(x,y) [C] (in //?) from two indices + big.matrix
 195
 196
 197     n <- nrow(synchrones)
 198     delta <- ncol(synchrones)
 199     #TODO: automatic tune of all these parameters ? (for other users)
 200     nvoice   <- 4
 201     # noctave = 2^13 = 8192 half hours ~ 180 days ; ~log2(ncol(synchrones))
 202     noctave = 13
 203     # 4 here represent 2^5 = 32 half-hours ~ 1 day
 204     #NOTE: default scalevector == 2^(0:(noctave * nvoice) / nvoice) * s0 (?)
 205     scalevector  <- 2^(4:(noctave * nvoice) / nvoice + 1)
 206     #condition: ( log2(s0*w0/(2*pi)) - 1 ) * nvoice + 1.5 >= 1
 207     s0=2
 208     w0=2*pi
 209     scaled=FALSE
 210     s0log = as.integer( (log2( s0*w0/(2*pi) ) - 1) * nvoice + 1.5 )
 211     totnoct = noctave + as.integer(s0log/nvoice) + 1
 212
 213     computeCWT = function(i)
 214     {
 215         if (verbose)
 216             cat(paste("+++ Compute Rwave::cwt() on serie ",i,"\n", sep=""))
 217         ts <- scale(ts(synchrones[i,]), center=TRUE, scale=scaled)
 218         totts.cwt = Rwave::cwt(ts, totnoct, nvoice, w0, plot=FALSE)
 219         ts.cwt = totts.cwt[,s0log:(s0log+noctave*nvoice)]
 220         #Normalization
 221         sqs <- sqrt(2^(0:(noctave*nvoice)/nvoice)*s0)
 222         sqres <- sweep(ts.cwt,2,sqs,'*')
 223         sqres / max(Mod(sqres))
 224     }
 225
 226     if (parll)
 227     {
 228         cl = parallel::makeCluster(ncores_clust)
 229         parallel::clusterExport(cl,
 230             varlist=c("synchrones","totnoct","nvoice","w0","s0log","noctave","s0","verbose"),
 231             envir=environment())
 232     }
 233
 234     # list of CWT from synchrones
 235     # TODO: fit in RAM, OK? If not, 2 options: serialize, compute individual distances
 236     Xcwt4 <-
 237         if (parll)
 238             parallel::parLapply(cl, seq_len(n), computeCWT)
 239         else
 240             lapply(seq_len(n), computeCWT)
 241
 242     if (parll)
 243         parallel::stopCluster(cl)
 244
 245     Xwer_dist <- bigmemory::big.matrix(nrow=n, ncol=n, type="double")
 246     fcoefs = rep(1/3, 3) #moving average on 3 values (TODO: very slow! correct?!)
 247     if (verbose)
 248         cat("*** Compute WER distances from CWT\n")
 249
 250     #TODO: computeDistances(i,j), et répartir les n(n-1)/2 couples d'indices
 251     #là c'est trop déséquilibré
 252
 253     computeDistancesLineI = function(i)
 254     {
 255         if (verbose)
 256             cat(paste("   Line ",i,"\n", sep=""))
 257         for (j in (i+1):n)
 258         {
 259             #TODO: 'circular=TRUE' is wrong, should just take values on the sides; to rewrite in C
 260             num <- filter(Mod(Xcwt4[[i]] * Conj(Xcwt4[[j]])), fcoefs, circular=TRUE)
 261             WX <- filter(Mod(Xcwt4[[i]] * Conj(Xcwt4[[i]])), fcoefs, circular=TRUE)
 262             WY <- filter(Mod(Xcwt4[[j]] * Conj(Xcwt4[[j]])), fcoefs, circular=TRUE)
 263             wer2 <- sum(colSums(num)^2) / sum( sum(colSums(WX) * colSums(WY)) )
 264             if (parll)
 265                 synchronicity::lock(m)
 266             Xwer_dist[i,j] <- sqrt(delta * ncol(Xcwt4[[1]]) * (1 - wer2))
 267             Xwer_dist[j,i] <- Xwer_dist[i,j]
 268             if (parll)
 269                 synchronicity::unlock(m)
 270         }
 271         Xwer_dist[i,i] = 0.
 272     }
 273
 274     parll = (requireNamespace("synchronicity",quietly=TRUE)
 275         && parll && Sys.info()['sysname'] != "Windows")
 276     if (parll)
 277         m <- synchronicity::boost.mutex()
 278
 279     ignored <-
 280         if (parll)
 281         {
 282             parallel::mclapply(seq_len(n-1), computeDistancesLineI,
 283                 mc.cores=ncores_clust, mc.allow.recursive=FALSE)
 284         }
 285         else
 286             lapply(seq_len(n-1), computeDistancesLineI)
 287     Xwer_dist[n,n] = 0.
 288     Xwer_dist
 289 }
 290
 291 # Helper function to divide indices into balanced sets
 292 .spreadIndices = function(indices, nb_per_chunk)
 293 {
 294     L = length(indices)
 295     nb_workers = floor( L / nb_per_chunk )
 296     if (nb_workers == 0)
 297     {
 298         # L < nb_series_per_chunk, simple case
 299         indices_workers = list(indices)
 300     }
 301     else
 302     {
 303         indices_workers = lapply( seq_len(nb_workers), function(i)
 304             indices[(nb_per_chunk*(i-1)+1):(nb_per_chunk*i)] )
 305         # Spread the remaining load among the workers
 306         rem = L %% nb_per_chunk
 307         while (rem > 0)
 308         {
 309             index = rem%%nb_workers + 1
 310             indices_workers[[index]] = c(indices_workers[[index]], indices[L-rem+1])
 311             rem = rem - 1
 312         }
 313     }
 314     indices_workers
 315 }