epclust/R/clustering.R

   1 #' @name clustering
   2 #' @rdname clustering
   3 #' @aliases clusteringTask1 clusteringTask2 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 #'   first clustering algorithm on a contributions matrix, while the latter clusters
  15 #'   a set of series inside one task (~nb_items_clust1)
  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 #' @inheritParams computeSynchrones
  21 #' @inheritParams claws
  22 #'
  23 #' @return For \code{clusteringTask1()}, the indices of the computed (K1) medoids.
  24 #'   Indices are irrelevant for stage 2 clustering, thus \code{clusteringTask2()}
  25 #'   outputs a big.matrix of medoids (of size LxK2, K2 = final number of clusters)
  26 NULL
  27
  28 #' @rdname clustering
  29 #' @export
  30 clusteringTask1 = function(indices, getContribs, K1, algoClust1, nb_items_clust1,
  31     ncores_clust=1, verbose=FALSE, parll=TRUE)
  32 {
  33     if (parll)
  34     {
  35         cl = parallel::makeCluster(ncores_clust, outfile = "")
  36         parallel::clusterExport(cl, varlist=c("getContribs","K1","verbose"), envir=environment())
  37     }
  38     while (length(indices) > K1)
  39     {
  40         indices_workers = .spreadIndices(indices, nb_items_clust1)
  41         if (verbose)
  42             cat(paste("*** [iterated] Clustering task 1 on ",length(indices)," series\n", sep=""))
  43         indices <-
  44             if (parll)
  45             {
  46                 unlist( parallel::parLapply(cl, indices_workers, function(inds) {
  47                     require("epclust", quietly=TRUE)
  48                     inds[ algoClust1(getContribs(inds), K1) ]
  49                 }) )
  50             }
  51             else
  52             {
  53                 unlist( lapply(indices_workers, function(inds)
  54                     inds[ algoClust1(getContribs(inds), K1) ]
  55                 ) )
  56             }
  57     }
  58     if (parll)
  59         parallel::stopCluster(cl)
  60
  61     indices #medoids
  62 }
  63
  64 #' @rdname clustering
  65 #' @export
  66 clusteringTask2 = function(medoids, K2, algoClust2, getRefSeries, nb_ref_curves,
  67     nb_series_per_chunk, sync_mean, nbytes,endian,ncores_clust=1,verbose=FALSE,parll=TRUE)
  68 {
  69     if (verbose)
  70         cat(paste("*** Clustering task 2 on ",ncol(medoids)," synchrones\n", sep=""))
  71
  72     if (ncol(medoids) <= K2)
  73         return (medoids)
  74     synchrones = computeSynchrones(medoids, getRefSeries, nb_ref_curves,
  75         nb_series_per_chunk, sync_mean, ncores_clust, verbose, parll)
  76     distances = computeWerDists(synchrones, nbytes, endian, ncores_clust, verbose, parll)
  77     if (verbose)
  78         cat(paste("   algoClust2() on ",nrow(distances)," items\n", sep=""))
  79     medoids[ algoClust2(distances,K2), ]
  80 }
  81
  82 #' computeSynchrones
  83 #'
  84 #' Compute the synchrones curves (sum of clusters elements) from a matrix of medoids,
  85 #' using L2 distances.
  86 #'
  87 #' @param medoids big.matrix of medoids (curves of same length as initial series)
  88 #' @param getRefSeries Function to retrieve initial series (e.g. in stage 2 after series
  89 #'   have been replaced by stage-1 medoids)
  90 #' @param nb_ref_curves How many reference series? (This number is known at this stage)
  91 #' @inheritParams claws
  92 #'
  93 #' @return A big.matrix of size L x K1 where L = length of a serie
  94 #'
  95 #' @export
  96 computeSynchrones = function(medoids, getRefSeries, nb_ref_curves,
  97     nb_series_per_chunk, sync_mean, ncores_clust=1,verbose=FALSE,parll=TRUE)
  98 {
  99     computeSynchronesChunk = function(indices)
 100     {
 101         if (parll)
 102         {
 103             require("bigmemory", quietly=TRUE)
 104             requireNamespace("synchronicity", quietly=TRUE)
 105             require("epclust", quietly=TRUE)
 106             synchrones <- bigmemory::attach.big.matrix(synchrones_desc)
 107             if (sync_mean)
 108                 counts <- bigmemory::attach.big.matrix(counts_desc)
 109             medoids <- bigmemory::attach.big.matrix(medoids_desc)
 110             m <- synchronicity::attach.mutex(m_desc)
 111         }
 112
 113         ref_series = getRefSeries(indices)
 114         nb_series = nrow(ref_series)
 115
 116         # Get medoids indices for this chunk of series
 117         mi = computeMedoidsIndices(medoids@address, ref_series)
 118
 119         for (i in seq_len(nb_series))
 120         {
 121             if (parll)
 122                 synchronicity::lock(m)
 123             synchrones[, mi[i] ] = synchrones[, mi[i] ] + ref_series[,i]
 124             if (sync_mean)
 125                 counts[ mi[i] ] = counts[ mi[i] ] + 1
 126             if (parll)
 127                 synchronicity::unlock(m)
 128         }
 129     }
 130
 131     K = ncol(medoids) ; L = nrow(medoids)
 132     # Use bigmemory (shared==TRUE by default) + synchronicity to fill synchrones in //
 133     # TODO: if size > RAM (not our case), use file-backed big.matrix
 134     synchrones = bigmemory::big.matrix(nrow=L, ncol=K, type="double", init=0.)
 135     if (sync_mean)
 136         counts = bigmemory::big.matrix(nrow=K, ncol=1, type="double", init=0)
 137     # synchronicity is only for Linux & MacOS; on Windows: run sequentially
 138     parll = (requireNamespace("synchronicity",quietly=TRUE)
 139         && parll && Sys.info()['sysname'] != "Windows")
 140     if (parll)
 141     {
 142         m <- synchronicity::boost.mutex()
 143         m_desc <- synchronicity::describe(m)
 144         synchrones_desc = bigmemory::describe(synchrones)
 145         if (sync_mean)
 146             counts_desc = bigmemory::describe(counts)
 147         medoids_desc = bigmemory::describe(medoids)
 148         cl = parallel::makeCluster(ncores_clust)
 149         varlist=c("synchrones_desc","sync_mean","m_desc","medoids_desc","getRefSeries")
 150         if (sync_mean)
 151             varlist = c(varlist, "counts_desc")
 152         parallel::clusterExport(cl, varlist, envir=environment())
 153     }
 154
 155     if (verbose)
 156     {
 157         if (verbose)
 158             cat(paste("--- Compute ",K," synchrones with ",nb_ref_curves," series\n", sep=""))
 159     }
 160     indices_workers = .spreadIndices(seq_len(nb_ref_curves), nb_series_per_chunk)
 161     ignored <-
 162         if (parll)
 163             parallel::parLapply(cl, indices_workers, computeSynchronesChunk)
 164         else
 165             lapply(indices_workers, computeSynchronesChunk)
 166
 167     if (parll)
 168         parallel::stopCluster(cl)
 169
 170     if (!sync_mean)
 171         return (synchrones)
 172
 173     #TODO: can we avoid this loop? ( synchrones = sweep(synchrones, 2, counts, '/') )
 174     for (i in seq_len(K))
 175         synchrones[,i] = synchrones[,i] / counts[i]
 176     #NOTE: odds for some clusters to be empty? (when series already come from stage 2)
 177     #      ...maybe; but let's hope resulting K1' be still quite bigger than K2
 178     noNA_rows = sapply(seq_len(K), function(i) all(!is.nan(synchrones[,i])))
 179     if (all(noNA_rows))
 180         return (synchrones)
 181     # Else: some clusters are empty, need to slice synchrones
 182     bigmemory::as.big.matrix(synchrones[,noNA_rows])
 183 }
 184
 185 #' computeWerDists
 186 #'
 187 #' Compute the WER distances between the synchrones curves (in rows), which are
 188 #' returned (e.g.) by \code{computeSynchrones()}
 189 #'
 190 #' @param synchrones A big.matrix of synchrones, in rows. The series have same length
 191 #'   as the series in the initial dataset
 192 #' @inheritParams claws
 193 #'
 194 #' @return A matrix of size K1 x K1
 195 #'
 196 #' @export
 197 computeWerDists = function(synchrones, nbytes,endian,ncores_clust=1,verbose=FALSE,parll=TRUE)
 198 {
 199     n <- nrow(synchrones)
 200     delta <- ncol(synchrones)
 201     #TODO: automatic tune of all these parameters ? (for other users)
 202     nvoice   <- 4
 203     # noctave = 2^13 = 8192 half hours ~ 180 days ; ~log2(ncol(synchrones))
 204     noctave = 13
 205     # 4 here represent 2^5 = 32 half-hours ~ 1 day
 206     #NOTE: default scalevector == 2^(0:(noctave * nvoice) / nvoice) * s0 (?)
 207     scalevector  <- 2^(4:(noctave * nvoice) / nvoice + 1)
 208     #condition: ( log2(s0*w0/(2*pi)) - 1 ) * nvoice + 1.5 >= 1
 209     s0 = 2
 210     w0 = 2*pi
 211     scaled=FALSE
 212     s0log = as.integer( (log2( s0*w0/(2*pi) ) - 1) * nvoice + 1.5 )
 213     totnoct = noctave + as.integer(s0log/nvoice) + 1
 214
 215     Xwer_dist <- bigmemory::big.matrix(nrow=n, ncol=n, type="double")
 216
 217     cwt_file = ".epclust_bin/cwt"
 218     #TODO: args, nb_per_chunk, nbytes, endian
 219
 220     # Generate n(n-1)/2 pairs for WER distances computations
 221     pairs = list()
 222     V = seq_len(n)
 223     for (i in 1:n)
 224     {
 225         V = V[-1]
 226         pairs = c(pairs, lapply(V, function(v) c(i,v)))
 227     }
 228
 229     computeSaveCWT = function(index)
 230     {
 231         ts <- scale(ts(synchrones[index,]), center=TRUE, scale=scaled)
 232         totts.cwt = Rwave::cwt(ts, totnoct, nvoice, w0, plot=FALSE)
 233         ts.cwt = totts.cwt[,s0log:(s0log+noctave*nvoice)]
 234         #Normalization
 235         sqs <- sqrt(2^(0:(noctave*nvoice)/nvoice)*s0)
 236         sqres <- sweep(ts.cwt,2,sqs,'*')
 237         res <- sqres / max(Mod(sqres))
 238         #TODO: serializer les CWT, les récupérer via getDataInFile ;
 239         #--> OK, faut juste stocker comme séries simples de taille delta*ncol (53*17519)
 240         binarize(c(as.double(Re(res)),as.double(Im(res))), cwt_file, ncol(res), ",", nbytes, endian)
 241     }
 242
 243     if (parll)
 244     {
 245         cl = parallel::makeCluster(ncores_clust)
 246         synchrones_desc <- bigmemory::describe(synchrones)
 247         Xwer_dist_desc <- bigmemory::describe(Xwer_dist)
 248         parallel::clusterExport(cl, varlist=c("synchrones_desc","Xwer_dist_desc","totnoct",
 249             "nvoice","w0","s0log","noctave","s0","verbose","getCWT"), envir=environment())
 250     }
 251
 252     if (verbose)
 253     {
 254         cat(paste("--- Compute WER dists\n", sep=""))
 255     #   precompute save all CWT........
 256     }
 257     #precompute and serialize all CWT
 258     ignored <-
 259         if (parll)
 260             parallel::parLapply(cl, 1:n, computeSaveCWT)
 261         else
 262             lapply(1:n, computeSaveCWT)
 263
 264     getCWT = function(index)
 265     {
 266         #from cwt_file ...
 267         res <- getDataInFile(c(2*index-1,2*index), cwt_file, nbytes, endian)
 268     ###############TODO:
 269     }
 270
 271     # Distance between rows i and j
 272     computeDistancesIJ = function(pair)
 273     {
 274         if (parll)
 275         {
 276             require("bigmemory", quietly=TRUE)
 277             require("epclust", quietly=TRUE)
 278             synchrones <- bigmemory::attach.big.matrix(synchrones_desc)
 279             Xwer_dist <- bigmemory::attach.big.matrix(Xwer_dist_desc)
 280         }
 281
 282         i = pair[1] ; j = pair[2]
 283         if (verbose && j==i+1)
 284             cat(paste("   Distances (",i,",",j,"), (",i,",",j+1,") ...\n", sep=""))
 285         cwt_i <- getCWT(i)
 286         cwt_j <- getCWT(j)
 287
 288         num <- epclustFilter(Mod(cwt_i * Conj(cwt_j)))
 289         WX  <- epclustFilter(Mod(cwt_i * Conj(cwt_i)))
 290         WY <- epclustFilter(Mod(cwt_j * Conj(cwt_j)))
 291         wer2 <- sum(colSums(num)^2) / sum(colSums(WX) * colSums(WY))
 292         Xwer_dist[i,j] <- sqrt(delta * ncol(cwt_i) * max(1 - wer2, 0.)) #FIXME: wer2 should be < 1
 293         Xwer_dist[j,i] <- Xwer_dist[i,j]
 294         Xwer_dist[i,i] = 0.
 295     }
 296
 297     if (verbose)
 298     {
 299         cat(paste("--- Compute WER dists\n", sep=""))
 300     }
 301     ignored <-
 302         if (parll)
 303             parallel::parLapply(cl, pairs, computeDistancesIJ)
 304         else
 305             lapply(pairs, computeDistancesIJ)
 306
 307     if (parll)
 308         parallel::stopCluster(cl)
 309
 310     Xwer_dist[n,n] = 0.
 311     distances <- Xwer_dist[,]
 312     rm(Xwer_dist) ; gc()
 313     distances #~small matrix K1 x K1
 314 }
 315
 316 # Helper function to divide indices into balanced sets
 317 .spreadIndices = function(indices, nb_per_set)
 318 {
 319     L = length(indices)
 320     nb_workers = floor( L / nb_per_set )
 321     rem = L %% max_nb_per_set
 322     if (nb_workers == 0 || (nb_workers==1 && rem==0))
 323     {
 324         # L <= max_nb_per_set, simple case
 325         indices_workers = list(indices)
 326     }
 327     else
 328     {
 329         indices_workers = lapply( seq_len(nb_workers), function(i)
 330             indices[(nb_per_chunk*(i-1)+1):(nb_per_set*i)] )
 331         # Spread the remaining load among the workers
 332         rem = L %% nb_per_set
 333         while (rem > 0)
 334         {
 335             index = rem%%nb_workers + 1
 336             indices_workers[[index]] = c(indices_workers[[index]], indices[L-rem+1])
 337             rem = rem - 1
 338         }
 339     }
 340     indices_workers
 341 }