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