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 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)$id.med
  67
  68 #' @rdname clustering
  69 #' @export
  70 computeClusters2 = function(medoids, K2,
  71     getRefSeries, nb_ref_curves, nb_series_per_chunk, ncores_clust=1,verbose=FALSE,parll=TRUE)
  72 {
  73     synchrones = computeSynchrones(medoids,
  74         getRefSeries, nb_ref_curves, nb_series_per_chunk, ncores_clust, verbose, parll)
  75     distances = computeWerDists(synchrones, ncores_clust, verbose, parll)
  76     medoids[ cluster::pam(distances, K2, diss=TRUE)$medoids , ]
  77 }
  78
  79 #' computeSynchrones
  80 #'
  81 #' Compute the synchrones curves (sum of clusters elements) from a matrix of medoids,
  82 #' using L2 distances.
  83 #'
  84 #' @param medoids Matrix of medoids (curves of same legnth as initial series)
  85 #' @param getRefSeries Function to retrieve initial series (e.g. in stage 2 after series
  86 #'   have been replaced by stage-1 medoids)
  87 #' @param nb_ref_curves How many reference series? (This number is known at this stage)
  88 #' @inheritParams claws
  89 #'
  90 #' @export
  91 computeSynchrones = function(medoids, getRefSeries,
  92     nb_ref_curves, nb_series_per_chunk, ncores_clust=1,verbose=FALSE,parll=TRUE)
  93 {
  94     computeSynchronesChunk = function(indices)
  95     {
  96         if (verbose)
  97             cat(paste("--- Compute synchrones for ",length(indices)," lines\n", sep=""))
  98         ref_series = getRefSeries(indices)
  99         #get medoids indices for this chunk of series
 100         for (i in seq_len(nrow(ref_series)))
 101         {
 102             j = which.min( rowSums( sweep(medoids, 2, ref_series[i,], '-')^2 ) )
 103             if (parll)
 104                 synchronicity::lock(m)
 105             synchrones[j,] = synchrones[j,] + ref_series[i,]
 106             counts[j,1] = counts[j,1] + 1
 107             if (parll)
 108                 synchronicity::unlock(m)
 109         }
 110     }
 111
 112     K = nrow(medoids)
 113     # Use bigmemory (shared==TRUE by default) + synchronicity to fill synchrones in //
 114     synchrones = bigmemory::big.matrix(nrow=K,ncol=ncol(medoids),type="double",init=0.)
 115     counts = bigmemory::big.matrix(nrow=K,ncol=1,type="double",init=0)
 116     # Fork (// run) only on Linux & MacOS; on Windows: run sequentially
 117     parll = (requireNamespace("synchronicity",quietly=TRUE)
 118         && parll && Sys.info()['sysname'] != "Windows")
 119     if (parll)
 120         m <- synchronicity::boost.mutex()
 121
 122     indices_workers = .spreadIndices(seq_len(nb_ref_curves), nb_series_per_chunk)
 123     ignored <-
 124         if (parll)
 125         {
 126             parallel::mclapply(indices_workers, computeSynchronesChunk,
 127                 mc.cores=ncores_clust, mc.allow.recursive=FALSE)
 128         }
 129         else
 130             lapply(indices_workers, computeSynchronesChunk)
 131
 132     mat_syncs = matrix(nrow=K, ncol=ncol(medoids))
 133     vec_count = rep(NA, K)
 134     #TODO: can we avoid this loop?
 135     for (i in seq_len(K))
 136     {
 137         mat_syncs[i,] = synchrones[i,]
 138         vec_count[i] = counts[i,1]
 139     }
 140     #NOTE: odds for some clusters to be empty? (when series already come from stage 2)
 141     #      ...maybe; but let's hope resulting K1' be still quite bigger than K2
 142     mat_syncs = sweep(mat_syncs, 1, vec_count, '/')
 143     mat_syncs[ sapply(seq_len(K), function(i) all(!is.nan(mat_syncs[i,]))) , ]
 144 }
 145
 146 #' computeWerDists
 147 #'
 148 #' Compute the WER distances between the synchrones curves (in rows), which are
 149 #' returned (e.g.) by \code{computeSynchrones()}
 150 #'
 151 #' @param synchrones A matrix of synchrones, in rows. The series have same length as the
 152 #'   series in the initial dataset
 153 #' @inheritParams claws
 154 #'
 155 #' @export
 156 computeWerDists = function(synchrones, ncores_clust=1,verbose=FALSE,parll=TRUE)
 157 {
 158     n <- nrow(synchrones)
 159     delta <- ncol(synchrones)
 160     #TODO: automatic tune of all these parameters ? (for other users)
 161     nvoice   <- 4
 162     # noctave = 2^13 = 8192 half hours ~ 180 days ; ~log2(ncol(synchrones))
 163     noctave = 13
 164     # 4 here represent 2^5 = 32 half-hours ~ 1 day
 165     #NOTE: default scalevector == 2^(0:(noctave * nvoice) / nvoice) * s0 (?)
 166     scalevector  <- 2^(4:(noctave * nvoice) / nvoice) * 2
 167     #condition: ( log2(s0*w0/(2*pi)) - 1 ) * nvoice + 1.5 >= 1
 168     s0=2
 169     w0=2*pi
 170     scaled=FALSE
 171     s0log = as.integer( (log2( s0*w0/(2*pi) ) - 1) * nvoice + 1.5 )
 172     totnoct = noctave + as.integer(s0log/nvoice) + 1
 173
 174     computeCWT = function(i)
 175     {
 176         if (verbose)
 177             cat(paste("+++ Compute Rwave::cwt() on serie ",i,"\n", sep=""))
 178         ts <- scale(ts(synchrones[i,]), center=TRUE, scale=scaled)
 179         totts.cwt = Rwave::cwt(ts,totnoct,nvoice,w0,plot=0)
 180         ts.cwt = totts.cwt[,s0log:(s0log+noctave*nvoice)]
 181         #Normalization
 182         sqs <- sqrt(2^(0:(noctave*nvoice)/nvoice)*s0)
 183         sqres <- sweep(ts.cwt,2,sqs,'*')
 184         sqres / max(Mod(sqres))
 185     }
 186
 187     if (parll)
 188     {
 189         cl = parallel::makeCluster(ncores_clust)
 190         parallel::clusterExport(cl,
 191             varlist=c("synchrones","totnoct","nvoice","w0","s0log","noctave","s0","verbose"),
 192             envir=environment())
 193     }
 194
 195     # (normalized) observations node with CWT
 196     Xcwt4 <-
 197         if (parll)
 198             parallel::parLapply(cl, seq_len(n), computeCWT)
 199         else
 200             lapply(seq_len(n), computeCWT)
 201
 202     if (parll)
 203         parallel::stopCluster(cl)
 204
 205     Xwer_dist <- bigmemory::big.matrix(nrow=n, ncol=n, type="double")
 206     fcoefs = rep(1/3, 3) #moving average on 3 values (TODO: very slow! correct?!)
 207     if (verbose)
 208         cat("*** Compute WER distances from CWT\n")
 209
 210     computeDistancesLineI = function(i)
 211     {
 212         if (verbose)
 213             cat(paste("   Line ",i,"\n", sep=""))
 214         for (j in (i+1):n)
 215         {
 216             #TODO: 'circular=TRUE' is wrong, should just take values on the sides; to rewrite in C
 217             num <- filter(Mod(Xcwt4[[i]] * Conj(Xcwt4[[j]])), fcoefs, circular=TRUE)
 218             WX <- filter(Mod(Xcwt4[[i]] * Conj(Xcwt4[[i]])), fcoefs, circular=TRUE)
 219             WY <- filter(Mod(Xcwt4[[j]] * Conj(Xcwt4[[j]])), fcoefs, circular=TRUE)
 220             wer2    <- sum(colSums(num)^2) / sum( sum(colSums(WX) * colSums(WY)) )
 221             if (parll)
 222                 synchronicity::lock(m)
 223             Xwer_dist[i,j] <- sqrt(delta * ncol(Xcwt4[[1]]) * (1 - wer2))
 224             Xwer_dist[j,i] <- Xwer_dist[i,j]
 225             if (parll)
 226                 synchronicity::unlock(m)
 227         }
 228         Xwer_dist[i,i] = 0.
 229     }
 230
 231     parll = (requireNamespace("synchronicity",quietly=TRUE)
 232         && parll && Sys.info()['sysname'] != "Windows")
 233     if (parll)
 234         m <- synchronicity::boost.mutex()
 235
 236     ignored <-
 237         if (parll)
 238         {
 239             parallel::mclapply(seq_len(n-1), computeDistancesLineI,
 240                 mc.cores=ncores_clust, mc.allow.recursive=FALSE)
 241         }
 242         else
 243             lapply(seq_len(n-1), computeDistancesLineI)
 244     Xwer_dist[n,n] = 0.
 245
 246     mat_dists = matrix(nrow=n, ncol=n)
 247     #TODO: avoid this loop?
 248     for (i in 1:n)
 249         mat_dists[i,] = Xwer_dist[i,]
 250     mat_dists
 251 }
 252
 253 # Helper function to divide indices into balanced sets
 254 .spreadIndices = function(indices, nb_per_chunk)
 255 {
 256     L = length(indices)
 257     nb_workers = floor( L / nb_per_chunk )
 258     if (nb_workers == 0)
 259     {
 260         # L < nb_series_per_chunk, simple case
 261         indices_workers = list(indices)
 262     }
 263     else
 264     {
 265         indices_workers = lapply( seq_len(nb_workers), function(i)
 266             indices[(nb_per_chunk*(i-1)+1):(nb_per_chunk*i)] )
 267         # Spread the remaining load among the workers
 268         rem = L %% nb_per_chunk
 269         while (rem > 0)
 270         {
 271             index = rem%%nb_workers + 1
 272             indices_workers[[index]] = c(indices_workers[[index]], indices[L-rem+1])
 273             rem = rem - 1
 274         }
 275     }
 276     indices_workers
 277 }