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_clust)
  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 #' @param distances matrix of K1 x K1 (WER) distances between synchrones
  22 #' @inheritParams computeSynchrones
  23 #' @inheritParams claws
  24 #'
  25 #' @return For \code{clusteringTask1()} and \code{computeClusters1()}, the indices of the
  26 #'   computed (K1) medoids. Indices are irrelevant for stage 2 clustering, thus
  27 #'   \code{computeClusters2()} outputs a big.matrix of medoids
  28 #'   (of size limited by nb_series_per_chunk)
  29 NULL
  30
  31 #' @rdname clustering
  32 #' @export
  33 clusteringTask1 = function(indices, getContribs, K1, nb_items_clust1,
  34     ncores_clust=1, verbose=FALSE, parll=TRUE)
  35 {
  36     if (verbose)
  37         cat(paste("*** Clustering task 1 on ",length(indices)," lines\n", sep=""))
  38
  39     if (parll)
  40     {
  41         cl = parallel::makeCluster(ncores_clust, outfile = "")
  42         parallel::clusterExport(cl, varlist=c("getContribs","K1","verbose"), envir=environment())
  43     }
  44     while (length(indices) > K1)
  45     {
  46         indices_workers = .spreadIndices(indices, nb_items_clust1, K1+1)
  47         indices <-
  48             if (parll)
  49             {
  50                 unlist( parallel::parLapply(cl, indices_workers, function(inds) {
  51                     require("epclust", quietly=TRUE)
  52                     inds[ computeClusters1(getContribs(inds), K1, verbose) ]
  53                 }) )
  54             }
  55             else
  56             {
  57                 unlist( lapply(indices_workers, function(inds)
  58                     inds[ computeClusters1(getContribs(inds), K1, verbose) ]
  59                 ) )
  60             }
  61     }
  62     if (parll)
  63         parallel::stopCluster(cl)
  64
  65     indices #medoids
  66 }
  67
  68 #' @rdname clustering
  69 #' @export
  70 clusteringTask2 = function(medoids, K2, getRefSeries, nb_ref_curves,
  71     nb_series_per_chunk, nbytes,endian,ncores_clust=1,verbose=FALSE,parll=TRUE)
  72 {
  73     if (verbose)
  74         cat(paste("*** Clustering task 2 on ",nrow(medoids)," lines\n", sep=""))
  75
  76     if (nrow(medoids) <= K2)
  77         return (medoids)
  78     synchrones = computeSynchrones(medoids,
  79         getRefSeries, nb_ref_curves, nb_series_per_chunk, ncores_clust, verbose, parll)
  80     distances = computeWerDists(synchrones, nbytes, endian, ncores_clust, verbose, parll)
  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(        t(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 L x K1 where L = length of a serie
 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         if (parll)
 125         {
 126             require("bigmemory", quietly=TRUE)
 127             requireNamespace("synchronicity", quietly=TRUE)
 128             require("epclust", quietly=TRUE)
 129             synchrones <- bigmemory::attach.big.matrix(synchrones_desc)
 130             counts <- bigmemory::attach.big.matrix(counts_desc)
 131             medoids <- bigmemory::attach.big.matrix(medoids_desc)
 132             m <- synchronicity::attach.mutex(m_desc)
 133         }
 134
 135         ref_series = getRefSeries(indices)
 136         nb_series = nrow(ref_series)
 137
 138         #get medoids indices for this chunk of series
 139         mi = computeMedoidsIndices(medoids@address, ref_series)
 140
 141         for (i in seq_len(nb_series))
 142         {
 143             if (parll)
 144                 synchronicity::lock(m)
 145             synchrones[, mi[i] ] = synchrones[, mi[i] ] + ref_series[,i]
 146             counts[ mi[i] ] = counts[ mi[i] ] + 1 #TODO: remove counts? ...or as arg?!
 147             if (parll)
 148                 synchronicity::unlock(m)
 149         }
 150     }
 151
 152     K = nrow(medoids) ; L = ncol(medoids)
 153     # Use bigmemory (shared==TRUE by default) + synchronicity to fill synchrones in //
 154     # TODO: if size > RAM (not our case), use file-backed big.matrix
 155     synchrones = bigmemory::big.matrix(nrow=L, ncol=K, type="double", init=0.)
 156     counts = bigmemory::big.matrix(nrow=K, ncol=1, type="double", init=0)
 157     # synchronicity is only for Linux & MacOS; on Windows: run sequentially
 158     parll = (requireNamespace("synchronicity",quietly=TRUE)
 159         && parll && Sys.info()['sysname'] != "Windows")
 160     if (parll)
 161     {
 162         m <- synchronicity::boost.mutex()
 163         m_desc <- synchronicity::describe(m)
 164         synchrones_desc = bigmemory::describe(synchrones)
 165         counts_desc = bigmemory::describe(counts)
 166         medoids_desc = bigmemory::describe(medoids)
 167         cl = parallel::makeCluster(ncores_clust)
 168         parallel::clusterExport(cl, varlist=c("synchrones_desc","counts_desc","counts",
 169             "verbose","m_desc","medoids_desc","getRefSeries"), envir=environment())
 170     }
 171
 172     indices_workers = .spreadIndices(seq_len(nb_ref_curves), nb_series_per_chunk)
 173     ignored <-
 174         if (parll)
 175             parallel::parLapply(cl, indices_workers, computeSynchronesChunk)
 176         else
 177             lapply(indices_workers, computeSynchronesChunk)
 178
 179     if (parll)
 180         parallel::stopCluster(cl)
 181
 182     #TODO: can we avoid this loop? ( synchrones = sweep(synchrones, 1, counts, '/') )
 183     for (i in seq_len(K))
 184         synchrones[,i] = synchrones[,i] / counts[i]
 185     #NOTE: odds for some clusters to be empty? (when series already come from stage 2)
 186     #      ...maybe; but let's hope resulting K1' be still quite bigger than K2
 187     noNA_rows = sapply(seq_len(K), function(i) all(!is.nan(synchrones[,i])))
 188     if (all(noNA_rows))
 189         return (synchrones)
 190     # Else: some clusters are empty, need to slice synchrones
 191     bigmemory::as.big.matrix(synchrones[,noNA_rows])
 192 }
 193
 194 #' computeWerDists
 195 #'
 196 #' Compute the WER distances between the synchrones curves (in rows), which are
 197 #' returned (e.g.) by \code{computeSynchrones()}
 198 #'
 199 #' @param synchrones A big.matrix of synchrones, in rows. The series have same length
 200 #'   as the series in the initial dataset
 201 #' @inheritParams claws
 202 #'
 203 #' @return A matrix of size K1 x K1
 204 #'
 205 #' @export
 206 computeWerDists = function(synchrones, nbytes,endian,ncores_clust=1,verbose=FALSE,parll=TRUE)
 207 {
 208     if (verbose)
 209         cat(paste("--- Compute WER dists\n", sep=""))
 210
 211     n <- nrow(synchrones)
 212     delta <- ncol(synchrones)
 213     #TODO: automatic tune of all these parameters ? (for other users)
 214     nvoice   <- 4
 215     # noctave = 2^13 = 8192 half hours ~ 180 days ; ~log2(ncol(synchrones))
 216     noctave = 13
 217     # 4 here represent 2^5 = 32 half-hours ~ 1 day
 218     #NOTE: default scalevector == 2^(0:(noctave * nvoice) / nvoice) * s0 (?)
 219     scalevector  <- 2^(4:(noctave * nvoice) / nvoice + 1)
 220     #condition: ( log2(s0*w0/(2*pi)) - 1 ) * nvoice + 1.5 >= 1
 221     s0 = 2
 222     w0 = 2*pi
 223     scaled=FALSE
 224     s0log = as.integer( (log2( s0*w0/(2*pi) ) - 1) * nvoice + 1.5 )
 225     totnoct = noctave + as.integer(s0log/nvoice) + 1
 226
 227     Xwer_dist <- bigmemory::big.matrix(nrow=n, ncol=n, type="double")
 228
 229     cwt_file = ".epclust_bin/cwt"
 230     #TODO: args, nb_per_chunk, nbytes, endian
 231
 232     # Generate n(n-1)/2 pairs for WER distances computations
 233     pairs = list()
 234     V = seq_len(n)
 235     for (i in 1:n)
 236     {
 237         V = V[-1]
 238         pairs = c(pairs, lapply(V, function(v) c(i,v)))
 239     }
 240
 241     computeSaveCWT = function(index)
 242     {
 243         ts <- scale(ts(synchrones[index,]), center=TRUE, scale=scaled)
 244         totts.cwt = Rwave::cwt(ts, totnoct, nvoice, w0, plot=FALSE)
 245         ts.cwt = totts.cwt[,s0log:(s0log+noctave*nvoice)]
 246         #Normalization
 247         sqs <- sqrt(2^(0:(noctave*nvoice)/nvoice)*s0)
 248         sqres <- sweep(ts.cwt,2,sqs,'*')
 249         res <- sqres / max(Mod(sqres))
 250         #TODO: serializer les CWT, les récupérer via getDataInFile ;
 251         #--> OK, faut juste stocker comme séries simples de taille delta*ncol (53*17519)
 252         binarize(c(as.double(Re(res)),as.double(Im(res))), cwt_file, ncol(res), ",", nbytes, endian)
 253     }
 254
 255     if (parll)
 256     {
 257         cl = parallel::makeCluster(ncores_clust)
 258         synchrones_desc <- bigmemory::describe(synchrones)
 259         Xwer_dist_desc <- bigmemory::describe(Xwer_dist)
 260         parallel::clusterExport(cl, varlist=c("synchrones_desc","Xwer_dist_desc","totnoct",
 261             "nvoice","w0","s0log","noctave","s0","verbose","getCWT"), envir=environment())
 262     }
 263
 264     #precompute and serialize all CWT
 265     ignored <-
 266         if (parll)
 267             parallel::parLapply(cl, 1:n, computeSaveCWT)
 268         else
 269             lapply(1:n, computeSaveCWT)
 270
 271     getCWT = function(index)
 272     {
 273         #from cwt_file ...
 274         res <- getDataInFile(c(2*index-1,2*index), cwt_file, nbytes, endian)
 275     ###############TODO:
 276     }
 277
 278     # Distance between rows i and j
 279     computeDistancesIJ = function(pair)
 280     {
 281         if (parll)
 282         {
 283             require("bigmemory", quietly=TRUE)
 284             require("epclust", quietly=TRUE)
 285             synchrones <- bigmemory::attach.big.matrix(synchrones_desc)
 286             Xwer_dist <- bigmemory::attach.big.matrix(Xwer_dist_desc)
 287         }
 288
 289         i = pair[1] ; j = pair[2]
 290         if (verbose && j==i+1)
 291             cat(paste("   Distances (",i,",",j,"), (",i,",",j+1,") ...\n", sep=""))
 292         cwt_i <- getCWT(i)
 293         cwt_j <- getCWT(j)
 294
 295         num <- epclustFilter(Mod(cwt_i * Conj(cwt_j)))
 296         WX  <- epclustFilter(Mod(cwt_i * Conj(cwt_i)))
 297         WY <- epclustFilter(Mod(cwt_j * Conj(cwt_j)))
 298         wer2 <- sum(colSums(num)^2) / sum(colSums(WX) * colSums(WY))
 299         Xwer_dist[i,j] <- sqrt(delta * ncol(cwt_i) * max(1 - wer2, 0.)) #FIXME: wer2 should be < 1
 300         Xwer_dist[j,i] <- Xwer_dist[i,j]
 301         Xwer_dist[i,i] = 0.
 302     }
 303
 304     ignored <-
 305         if (parll)
 306             parallel::parLapply(cl, pairs, computeDistancesIJ)
 307         else
 308             lapply(pairs, computeDistancesIJ)
 309
 310     if (parll)
 311         parallel::stopCluster(cl)
 312
 313     Xwer_dist[n,n] = 0.
 314     distances <- Xwer_dist[,]
 315     rm(Xwer_dist) ; gc()
 316     distances #~small matrix K1 x K1
 317 }
 318
 319 # Helper function to divide indices into balanced sets
 320 .spreadIndices = function(indices, max_per_set, min_nb_per_set = 1)
 321 {
 322     L = length(indices)
 323     min_nb_workers = floor( L / max_per_set )
 324     rem = L %% max_per_set
 325     if (nb_workers == 0 || (nb_workers==1 && rem==0))
 326     {
 327         # L <= max_nb_per_set, simple case
 328         indices_workers = list(indices)
 329     }
 330     else
 331     {
 332         indices_workers = lapply( seq_len(nb_workers), function(i)
 333             indices[(max_nb_per_set*(i-1)+1):(max_per_set*i)] )
 334         # Two cases: remainder is >= min_per_set (easy)...
 335         if (rem >= min_nb_per_set)
 336             indices_workers = c( indices_workers, list(tail(indices,rem)) )
 337         #...or < min_per_set: harder, need to remove indices from current sets to feed
 338         # the too-small remainder. It may fail: then fallback to "slightly bigger sets"
 339         else
 340         {
 341             save_indices_workers = indices_workers
 342             small_set = tail(indices,rem)
 343             # Try feeding small_set until it reaches min_per_set, whle keeping the others big enough
 344             # Spread the remaining load among the workers
 345         rem = L %% nb_per_chunk
 346         while (rem > 0)
 347         {
 348             index = rem%%nb_workers + 1
 349             indices_workers[[index]] = c(indices_workers[[index]], indices[L-rem+1])
 350             rem = rem - 1
 351         }
 352     }
 353     indices_workers
 354 }