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, c("getContribs","K1","verbose"), envir=environment())
  37     }
  38     # Iterate clustering algorithm 1 until K1 medoids are found
  39     while (length(indices) > K1)
  40     {
  41         # Balance tasks by splitting the indices set - as evenly as possible
  42         indices_workers = .spreadIndices(indices, nb_items_clust1)
  43         if (verbose)
  44             cat(paste("*** [iterated] Clustering task 1 on ",length(indices)," series\n", sep=""))
  45         indices <-
  46             if (parll)
  47             {
  48                 unlist( parallel::parLapply(cl, indices_workers, function(inds) {
  49                     require("epclust", quietly=TRUE)
  50                     inds[ algoClust1(getContribs(inds), K1) ]
  51                 }) )
  52             }
  53             else
  54             {
  55                 unlist( lapply(indices_workers, function(inds)
  56                     inds[ algoClust1(getContribs(inds), K1) ]
  57                 ) )
  58             }
  59     }
  60     if (parll)
  61         parallel::stopCluster(cl)
  62
  63     indices #medoids
  64 }
  65
  66 #' @rdname clustering
  67 #' @export
  68 clusteringTask2 = function(medoids, K2, algoClust2, getRefSeries, nb_ref_curves,
  69     nb_series_per_chunk, nbytes,endian,ncores_clust=1,verbose=FALSE,parll=TRUE)
  70 {
  71     if (verbose)
  72         cat(paste("*** Clustering task 2 on ",ncol(medoids)," synchrones\n", sep=""))
  73
  74     if (ncol(medoids) <= K2)
  75         return (medoids)
  76
  77     # A) Obtain synchrones, that is to say the cumulated power consumptions
  78     #    for each of the K1 initial groups
  79     synchrones = computeSynchrones(medoids, getRefSeries, nb_ref_curves,
  80         nb_series_per_chunk, ncores_clust, verbose, parll)
  81
  82     # B) Compute the WER distances (Wavelets Extended coefficient of deteRmination)
  83     distances = computeWerDists(synchrones, nbytes, endian, ncores_clust, verbose, parll)
  84
  85     # C) Apply clustering algorithm 2 on the WER distances matrix
  86     if (verbose)
  87         cat(paste("   algoClust2() on ",nrow(distances)," items\n", sep=""))
  88     medoids[ ,algoClust2(distances,K2) ]
  89 }
  90
  91 #' computeSynchrones
  92 #'
  93 #' Compute the synchrones curves (sum of clusters elements) from a matrix of medoids,
  94 #' using euclidian distance.
  95 #'
  96 #' @param medoids big.matrix of medoids (curves of same length as initial series)
  97 #' @param getRefSeries Function to retrieve initial series (e.g. in stage 2 after series
  98 #'   have been replaced by stage-1 medoids)
  99 #' @param nb_ref_curves How many reference series? (This number is known at this stage)
 100 #' @inheritParams claws
 101 #'
 102 #' @return A big.matrix of size L x K1 where L = length of a serie
 103 #'
 104 #' @export
 105 computeSynchrones = function(medoids, getRefSeries, nb_ref_curves,
 106     nb_series_per_chunk, ncores_clust=1,verbose=FALSE,parll=TRUE)
 107 {
 108     # Synchrones computation is embarassingly parallel: compute it by chunks of series
 109     computeSynchronesChunk = function(indices)
 110     {
 111         if (parll)
 112         {
 113             require("bigmemory", quietly=TRUE)
 114             requireNamespace("synchronicity", quietly=TRUE)
 115             require("epclust", quietly=TRUE)
 116             # The big.matrix objects need to be attached to be usable on the workers
 117             synchrones <- bigmemory::attach.big.matrix(synchrones_desc)
 118             medoids <- bigmemory::attach.big.matrix(medoids_desc)
 119             m <- synchronicity::attach.mutex(m_desc)
 120         }
 121
 122         # Obtain a chunk of reference series
 123         ref_series = getRefSeries(indices)
 124         nb_series = ncol(ref_series)
 125
 126         # Get medoids indices for this chunk of series
 127         mi = computeMedoidsIndices(medoids@address, ref_series)
 128
 129         # Update synchrones using mi above
 130         for (i in seq_len(nb_series))
 131         {
 132             if (parll)
 133                 synchronicity::lock(m) #locking required because several writes at the same time
 134             synchrones[, mi[i] ] = synchrones[, mi[i] ] + ref_series[,i]
 135             if (parll)
 136                 synchronicity::unlock(m)
 137         }
 138     }
 139
 140     K = ncol(medoids) ; L = nrow(medoids)
 141     # Use bigmemory (shared==TRUE by default) + synchronicity to fill synchrones in //
 142     synchrones = bigmemory::big.matrix(nrow=L, ncol=K, type="double", init=0.)
 143     # NOTE: synchronicity is only for Linux & MacOS; on Windows: run sequentially
 144     parll = (requireNamespace("synchronicity",quietly=TRUE)
 145         && parll && Sys.info()['sysname'] != "Windows")
 146     if (parll)
 147     {
 148         m <- synchronicity::boost.mutex() #for lock/unlock, see computeSynchronesChunk
 149         # mutex and big.matrix objects cannot be passed directly:
 150         # they will be accessed from their description
 151         m_desc <- synchronicity::describe(m)
 152         synchrones_desc = bigmemory::describe(synchrones)
 153         medoids_desc = bigmemory::describe(medoids)
 154         cl = parallel::makeCluster(ncores_clust)
 155         parallel::clusterExport(cl, envir=environment(),
 156             varlist=c("synchrones_desc","m_desc","medoids_desc","getRefSeries"))
 157     }
 158
 159     if (verbose)
 160         cat(paste("--- Compute ",K," synchrones with ",nb_ref_curves," series\n", sep=""))
 161
 162     # Balance tasks by splitting the indices set - maybe not so evenly, but
 163     # max==TRUE in next call ensures that no set has more than nb_series_per_chunk items.
 164     indices_workers = .spreadIndices(seq_len(nb_ref_curves), nb_series_per_chunk, max=TRUE)
 165     ignored <-
 166         if (parll)
 167             parallel::parLapply(cl, indices_workers, computeSynchronesChunk)
 168         else
 169             lapply(indices_workers, computeSynchronesChunk)
 170
 171     if (parll)
 172         parallel::stopCluster(cl)
 173
 174     return (synchrones)
 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 matrix of size K1 x K1
 187 #'
 188 #' @export
 189 computeWerDists = function(synchrones, nbytes,endian,ncores_clust=1,verbose=FALSE,parll=TRUE)
 190 {
 191     n <- ncol(synchrones)
 192     L <- nrow(synchrones)
 193     #TODO: automatic tune of all these parameters ? (for other users)
 194     # 4 here represent 2^5 = 32 half-hours ~ 1 day
 195     nvoice   <- 4
 196     # noctave = 2^13 = 8192 half hours ~ 180 days ; ~log2(ncol(synchrones))
 197     noctave = 13
 198
 199     Xwer_dist <- bigmemory::big.matrix(nrow=n, ncol=n, type="double")
 200
 201     cwt_file = ".epclust_bin/cwt"
 202     #TODO: args, nb_per_chunk, nbytes, endian
 203
 204     # Generate n(n-1)/2 pairs for WER distances computations
 205     pairs = list()
 206     V = seq_len(n)
 207     for (i in 1:n)
 208     {
 209         V = V[-1]
 210         pairs = c(pairs, lapply(V, function(v) c(i,v)))
 211     }
 212
 213     computeSaveCWT = function(index)
 214     {
 215         ts <- scale(ts(synchrones[,index]), center=TRUE, scale=FALSE)
 216         totts.cwt = Rwave::cwt(ts, totnoct, nvoice, w0=2*pi, twoD=TRUE, plot=FALSE)
 217         ts.cwt = totts.cwt[,s0log:(s0log+noctave*nvoice)]
 218         #Normalization
 219         sqs <- sqrt(2^(0:(noctave*nvoice)/nvoice)*s0)
 220         sqres <- sweep(ts.cwt,2,sqs,'*')
 221         res <- sqres / max(Mod(sqres))
 222         #TODO: serializer les CWT, les récupérer via getDataInFile ;
 223         #--> OK, faut juste stocker comme séries simples de taille L*n' (53*17519)
 224         binarize(c(as.double(Re(res)),as.double(Im(res))), cwt_file, ncol(res), ",", nbytes, endian)
 225     }
 226
 227     if (parll)
 228     {
 229         cl = parallel::makeCluster(ncores_clust)
 230         synchrones_desc <- bigmemory::describe(synchrones)
 231         Xwer_dist_desc <- bigmemory::describe(Xwer_dist)
 232         parallel::clusterExport(cl, envir=environment(),
 233             varlist=c("synchrones_desc","Xwer_dist_desc","totnoct","nvoice","w0","s0log",
 234                 "noctave","s0","verbose","getCWT"))
 235     }
 236
 237     if (verbose)
 238     {
 239         cat(paste("--- Compute WER dists\n", sep=""))
 240     #   precompute save all CWT........
 241     }
 242     #precompute and serialize all CWT
 243     ignored <-
 244         if (parll)
 245             parallel::parLapply(cl, 1:n, computeSaveCWT)
 246         else
 247             lapply(1:n, computeSaveCWT)
 248
 249     getCWT = function(index)
 250     {
 251         #from cwt_file ...
 252         res <- getDataInFile(c(2*index-1,2*index), cwt_file, nbytes, endian)
 253     ###############TODO:
 254     }
 255
 256     # Distance between rows i and j
 257     computeDistancesIJ = function(pair)
 258     {
 259         if (parll)
 260         {
 261             require("bigmemory", quietly=TRUE)
 262             require("epclust", quietly=TRUE)
 263             synchrones <- bigmemory::attach.big.matrix(synchrones_desc)
 264             Xwer_dist <- bigmemory::attach.big.matrix(Xwer_dist_desc)
 265         }
 266
 267         i = pair[1] ; j = pair[2]
 268         if (verbose && j==i+1)
 269             cat(paste("   Distances (",i,",",j,"), (",i,",",j+1,") ...\n", sep=""))
 270         cwt_i <- getCWT(i)
 271         cwt_j <- getCWT(j)
 272
 273         num <- epclustFilter(Mod(cwt_i * Conj(cwt_j)))
 274         WX  <- epclustFilter(Mod(cwt_i * Conj(cwt_i)))
 275         WY <- epclustFilter(Mod(cwt_j * Conj(cwt_j)))
 276         wer2 <- sum(colSums(num)^2) / sum(colSums(WX) * colSums(WY))
 277         Xwer_dist[i,j] <- sqrt(L * ncol(cwt_i) * max(1 - wer2, 0.))
 278         Xwer_dist[j,i] <- Xwer_dist[i,j]
 279         Xwer_dist[i,i] = 0.
 280     }
 281
 282     if (verbose)
 283     {
 284         cat(paste("--- Compute WER dists\n", sep=""))
 285     }
 286     ignored <-
 287         if (parll)
 288             parallel::parLapply(cl, pairs, computeDistancesIJ)
 289         else
 290             lapply(pairs, computeDistancesIJ)
 291
 292     if (parll)
 293         parallel::stopCluster(cl)
 294
 295     Xwer_dist[n,n] = 0.
 296     distances <- Xwer_dist[,]
 297     rm(Xwer_dist) ; gc()
 298     distances #~small matrix K1 x K1
 299 }
 300
 301 # Helper function to divide indices into balanced sets
 302 # If max == TRUE, sets sizes cannot exceed nb_per_set
 303 .spreadIndices = function(indices, nb_per_set, max=FALSE)
 304 {
 305     L = length(indices)
 306     nb_workers = floor( L / nb_per_set )
 307     rem = L %% nb_per_set
 308     if (nb_workers == 0 || (nb_workers==1 && rem==0))
 309     {
 310         # L <= nb_per_set, simple case
 311         indices_workers = list(indices)
 312     }
 313     else
 314     {
 315         indices_workers = lapply( seq_len(nb_workers), function(i)
 316             indices[(nb_per_set*(i-1)+1):(nb_per_set*i)] )
 317
 318         if (max)
 319         {
 320             # Sets are not so well balanced, but size is supposed to be critical
 321             return ( c( indices_workers, (L-rem+1):L ) )
 322         }
 323
 324         # Spread the remaining load among the workers
 325         rem = L %% nb_per_set
 326         while (rem > 0)
 327         {
 328             index = rem%%nb_workers + 1
 329             indices_workers[[index]] = c(indices_workers[[index]], indices[L-rem+1])
 330             rem = rem - 1
 331         }
 332     }
 333     indices_workers
 334 }