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 1 on ",length(indices)," lines\n", sep=""))
  37
  38     if (parll)
  39     {
  40         cl = parallel::makeCluster(ncores_clust)
  41         parallel::clusterExport(cl, varlist=c("getContribs","K1","verbose"), envir=environment())
  42     }
  43     while (length(indices) > K1)
  44     {
  45         indices_workers = .spreadIndices(indices, nb_series_per_chunk)
  46         indices <-
  47             if (parll)
  48             {
  49                 unlist( parallel::parLapply(cl, indices_workers, function(inds) {
  50                     require("epclust", quietly=TRUE)
  51                     inds[ computeClusters1(getContribs(inds), K1, verbose) ]
  52                 }) )
  53             }
  54             else
  55             {
  56                 unlist( lapply(indices_workers, function(inds)
  57                     inds[ computeClusters1(getContribs(inds), K1, verbose) ]
  58                 ) )
  59             }
  60     }
  61     if (parll)
  62         parallel::stopCluster(cl)
  63
  64     indices #medoids
  65 }
  66
  67 #' @rdname clustering
  68 #' @export
  69 clusteringTask2 = function(medoids, K2,
  70     getRefSeries, nb_ref_curves, nb_series_per_chunk, ncores_clust=1,verbose=FALSE,parll=TRUE)
  71 {
  72     if (verbose)
  73         cat(paste("*** Clustering task 2 on ",nrow(medoids)," lines\n", sep=""))
  74
  75     if (nrow(medoids) <= K2)
  76         return (medoids)
  77     synchrones = computeSynchrones(medoids,
  78         getRefSeries, nb_ref_curves, nb_series_per_chunk, ncores_clust, verbose, parll)
  79     distances = computeWerDists(synchrones, ncores_clust, verbose, parll)
  80     # PAM in package 'cluster' cannot take big.matrix in input: need to cast it
  81     medoids[ computeClusters2(distances[,],K2,verbose), ]
  82 }
  83
  84 #' @rdname clustering
  85 #' @export
  86 computeClusters1 = function(contribs, K1, verbose=FALSE)
  87 {
  88     if (verbose)
  89         cat(paste("   computeClusters1() on ",nrow(contribs)," lines\n", sep=""))
  90     cluster::pam(contribs, K1, diss=FALSE)$id.med
  91 }
  92
  93 #' @rdname clustering
  94 #' @export
  95 computeClusters2 = function(distances, K2, verbose=FALSE)
  96 {
  97     if (verbose)
  98         cat(paste("   computeClusters2() on ",nrow(distances)," lines\n", sep=""))
  99     cluster::pam(distances, K2, diss=TRUE)$id.med
 100 }
 101
 102 #' computeSynchrones
 103 #'
 104 #' Compute the synchrones curves (sum of clusters elements) from a matrix of medoids,
 105 #' using L2 distances.
 106 #'
 107 #' @param medoids big.matrix of medoids (curves of same length as initial series)
 108 #' @param getRefSeries Function to retrieve initial series (e.g. in stage 2 after series
 109 #'   have been replaced by stage-1 medoids)
 110 #' @param nb_ref_curves How many reference series? (This number is known at this stage)
 111 #' @inheritParams claws
 112 #'
 113 #' @return A big.matrix of size K1 x L where L = data_length
 114 #'
 115 #' @export
 116 computeSynchrones = function(medoids, getRefSeries,
 117     nb_ref_curves, nb_series_per_chunk, ncores_clust=1,verbose=FALSE,parll=TRUE)
 118 {
 119     if (verbose)
 120         cat(paste("--- Compute synchrones\n", sep=""))
 121
 122     computeSynchronesChunk = function(indices)
 123     {
 124         ref_series = getRefSeries(indices)
 125         nb_series = nrow(ref_series)
 126         #get medoids indices for this chunk of series
 127
 128         #TODO: debug this (address is OK but values are garbage: why?)
 129 #        mi = .Call("computeMedoidsIndices", medoids@address, ref_series, PACKAGE="epclust")
 130
 131         #R-equivalent, requiring a matrix (thus potentially breaking "fit-in-memory" hope)
 132         mat_meds = medoids[,]
 133         mi = rep(NA,nb_series)
 134         for (i in 1:nb_series)
 135             mi[i] <- which.min( rowSums( sweep(mat_meds, 2, ref_series[i,], '-')^2 ) )
 136         rm(mat_meds); gc()
 137
 138         for (i in seq_len(nb_series))
 139         {
 140             if (parll)
 141                 synchronicity::lock(m)
 142             synchrones[mi[i],] = synchrones[mi[i],] + ref_series[i,]
 143             counts[mi[i],1] = counts[mi[i],1] + 1
 144             if (parll)
 145                 synchronicity::unlock(m)
 146         }
 147     }
 148
 149     K = nrow(medoids) ; L = ncol(medoids)
 150     # Use bigmemory (shared==TRUE by default) + synchronicity to fill synchrones in //
 151     # TODO: if size > RAM (not our case), use file-backed big.matrix
 152     synchrones = bigmemory::big.matrix(nrow=K, ncol=L, type="double", init=0.)
 153     counts = bigmemory::big.matrix(nrow=K, ncol=1, type="double", init=0)
 154     # synchronicity is only for Linux & MacOS; on Windows: run sequentially
 155     parll = (requireNamespace("synchronicity",quietly=TRUE)
 156         && parll && Sys.info()['sysname'] != "Windows")
 157     if (parll)
 158         m <- synchronicity::boost.mutex()
 159
 160     if (parll)
 161     {
 162         cl = parallel::makeCluster(ncores_clust)
 163         parallel::clusterExport(cl,
 164             varlist=c("synchrones","counts","verbose","medoids","getRefSeries"),
 165             envir=environment())
 166     }
 167
 168     indices_workers = .spreadIndices(seq_len(nb_ref_curves), nb_series_per_chunk)
 169         browser()
 170     ignored <-
 171         if (parll)
 172             parallel::parLapply(cl, indices_workers, computeSynchronesChunk)
 173         else
 174             lapply(indices_workers, computeSynchronesChunk)
 175
 176     if (parll)
 177         parallel::stopCluster(cl)
 178
 179     #TODO: can we avoid this loop? ( synchrones = sweep(synchrones, 1, counts, '/') )
 180     for (i in seq_len(K))
 181         synchrones[i,] = synchrones[i,] / counts[i,1]
 182     #NOTE: odds for some clusters to be empty? (when series already come from stage 2)
 183     #      ...maybe; but let's hope resulting K1' be still quite bigger than K2
 184     noNA_rows = sapply(seq_len(K), function(i) all(!is.nan(synchrones[i,])))
 185     if (all(noNA_rows))
 186         return (synchrones)
 187     # Else: some clusters are empty, need to slice synchrones
 188     synchrones[noNA_rows,]
 189 }
 190
 191 #' computeWerDists
 192 #'
 193 #' Compute the WER distances between the synchrones curves (in rows), which are
 194 #' returned (e.g.) by \code{computeSynchrones()}
 195 #'
 196 #' @param synchrones A big.matrix of synchrones, in rows. The series have same length
 197 #'   as the series in the initial dataset
 198 #' @inheritParams claws
 199 #'
 200 #' @return A big.matrix of size K1 x K1
 201 #'
 202 #' @export
 203 computeWerDists = function(synchrones, ncores_clust=1,verbose=FALSE,parll=TRUE)
 204 {
 205     if (verbose)
 206         cat(paste("--- Compute WER dists\n", sep=""))
 207
 208     n <- nrow(synchrones)
 209     delta <- ncol(synchrones)
 210     #TODO: automatic tune of all these parameters ? (for other users)
 211     nvoice   <- 4
 212     # noctave = 2^13 = 8192 half hours ~ 180 days ; ~log2(ncol(synchrones))
 213     noctave = 13
 214     # 4 here represent 2^5 = 32 half-hours ~ 1 day
 215     #NOTE: default scalevector == 2^(0:(noctave * nvoice) / nvoice) * s0 (?)
 216     scalevector  <- 2^(4:(noctave * nvoice) / nvoice + 1)
 217     #condition: ( log2(s0*w0/(2*pi)) - 1 ) * nvoice + 1.5 >= 1
 218     s0=2
 219     w0=2*pi
 220     scaled=FALSE
 221     s0log = as.integer( (log2( s0*w0/(2*pi) ) - 1) * nvoice + 1.5 )
 222     totnoct = noctave + as.integer(s0log/nvoice) + 1
 223
 224     Xwer_dist <- bigmemory::big.matrix(nrow=n, ncol=n, type="double")
 225     fcoefs = rep(1/3, 3) #moving average on 3 values
 226
 227     # Generate n(n-1)/2 pairs for WER distances computations
 228     pairs = list()
 229     V = seq_len(n)
 230     for (i in 1:n)
 231     {
 232         V = V[-1]
 233         pairs = c(pairs, lapply(V, function(v) c(i,v)))
 234     }
 235
 236     computeCWT = function(i)
 237     {
 238         ts <- scale(ts(synchrones[i,]), center=TRUE, scale=scaled)
 239         totts.cwt = Rwave::cwt(ts, totnoct, nvoice, w0, plot=FALSE)
 240         ts.cwt = totts.cwt[,s0log:(s0log+noctave*nvoice)]
 241         #Normalization
 242         sqs <- sqrt(2^(0:(noctave*nvoice)/nvoice)*s0)
 243         sqres <- sweep(ts.cwt,2,sqs,'*')
 244         sqres / max(Mod(sqres))
 245     }
 246
 247     computeDistancesIJ = function(pair)
 248     {
 249         i = pair[1] ; j = pair[2]
 250         if (verbose && j==i+1)
 251             cat(paste("   Distances (",i,",",j,"), (",i,",",j+1,") ...\n", sep=""))
 252         cwt_i = computeCWT(i)
 253         cwt_j = computeCWT(j)
 254         num <- .Call("filter", Mod(cwt_i * Conj(cwt_j)), PACKAGE="epclust")
 255         WX  <- .Call("filter", Mod(cwt_i * Conj(cwt_i)), PACKAGE="epclust")
 256         WY  <- .Call("filter", Mod(cwt_j * Conj(cwt_j)), PACKAGE="epclust")
 257         wer2 <- sum(colSums(num)^2) / sum(colSums(WX) * colSums(WY))
 258         Xwer_dist[i,j] <- sqrt(delta * ncol(cwt_i) * (1 - wer2))
 259         Xwer_dist[j,i] <- Xwer_dist[i,j]
 260         Xwer_dist[i,i] = 0.
 261     }
 262
 263     if (parll)
 264     {
 265         cl = parallel::makeCluster(ncores_clust)
 266         parallel::clusterExport(cl,
 267             varlist=c("synchrones","totnoct","nvoice","w0","s0log","noctave","s0","verbose"),
 268             envir=environment())
 269     }
 270
 271     ignored <-
 272         if (parll)
 273             parallel::parLapply(cl, pairs, computeDistancesIJ)
 274         else
 275             lapply(pairs, computeDistancesIJ)
 276
 277     if (parll)
 278         parallel::stopCluster(cl)
 279
 280     Xwer_dist[n,n] = 0.
 281     Xwer_dist
 282 }
 283
 284 # Helper function to divide indices into balanced sets
 285 .spreadIndices = function(indices, nb_per_chunk)
 286 {
 287     L = length(indices)
 288     nb_workers = floor( L / nb_per_chunk )
 289     if (nb_workers == 0)
 290     {
 291         # L < nb_series_per_chunk, simple case
 292         indices_workers = list(indices)
 293     }
 294     else
 295     {
 296         indices_workers = lapply( seq_len(nb_workers), function(i)
 297             indices[(nb_per_chunk*(i-1)+1):(nb_per_chunk*i)] )
 298         # Spread the remaining load among the workers
 299         rem = L %% nb_per_chunk
 300         while (rem > 0)
 301         {
 302             index = rem%%nb_workers + 1
 303             indices_workers[[index]] = c(indices_workers[[index]], indices[L-rem+1])
 304             rem = rem - 1
 305         }
 306     }
 307     indices_workers
 308 }