epclust/R/clustering.R

   1 # Cluster one full task (nb_curves / ntasks series)
   2 clusteringTask = function(indices_clust)
   3 {
   4     cl_clust = parallel::makeCluster(ncores_clust)
   5     parallel::clusterExport(cl_clust,
   6         varlist=c("K1","K2","WER"),
   7         envir=environment())
   8     repeat
   9     {
  10         nb_workers = max( 1, round( length(indices_clust) / nb_series_per_chunk ) )
  11         indices_workers = lapply(seq_len(nb_workers), function(i) {
  12             upper_bound = ifelse( i<nb_workers,
  13                 min(nb_series_per_chunk*i,length(indices_clust)), length(indices_clust) )
  14             indices_clust[(nb_series_per_chunk*(i-1)+1):upper_bound]
  15         })
  16         indices_clust = parallel::parLapply(cl, indices_workers, clusterChunk)
  17         # TODO: soft condition between K2 and K1, before applying final WER step
  18         if ((WER=="end" && length(indices_clust)==K1) || (WER=="mix" && length(indices_clust)==K2))
  19             break
  20     }
  21     parallel::stopCluster(cl_clust)
  22     unlist(indices_clust)
  23 }
  24
  25 # Cluster a chunk of series inside one task (~max nb_series_per_chunk)
  26 clusterChunk = function(indices_chunk)
  27 {
  28     coeffs = readCoeffs(indices_chunk)
  29     cl = computeClusters(as.matrix(coeffs[,2:ncol(coeffs)]), K1, diss=FALSE)
  30     if (WER=="mix" > 0)
  31     {
  32         curves = computeSynchrones(cl)
  33         dists = computeWerDists(curves)
  34         cl = computeClusters(dists, K2, diss=TRUE)
  35     }
  36     indices_chunk[cl]
  37 }
  38
  39 # Apply the clustering algorithm (PAM) on a coeffs or distances matrix
  40 computeClusters = function(md, K, diss)
  41 {
  42     if (!require(cluster, quietly=TRUE))
  43         stop("Unable to load cluster library")
  44     cluster::pam(md, K, diss=diss)$id.med
  45 }
  46
  47 # Compute the synchrones curves (sum of clusters elements) from a clustering result
  48 computeSynchrones = function(indices)
  49 {
  50     colSums( getData(indices) )
  51 }
  52
  53 # Compute the WER distance between the synchrones curves
  54 computeWerDist = function(curves)
  55 {
  56     if (!require("Rwave", quietly=TRUE))
  57         stop("Unable to load Rwave library")
  58     n <- nrow(curves)
  59     delta <- ncol(curves)
  60     #TODO: automatic tune of all these parameters ? (for other users)
  61     nvoice   <- 4
  62     # noctave = 2^13 = 8192 half hours ~ 180 days ; ~log2(ncol(curves))
  63     noctave = 13
  64     # 4 here represent 2^5 = 32 half-hours ~ 1 day
  65     #NOTE: default scalevector == 2^(0:(noctave * nvoice) / nvoice) * s0 (?)
  66     scalevector  <- 2^(4:(noctave * nvoice) / nvoice) * 2
  67     #condition: ( log2(s0*w0/(2*pi)) - 1 ) * nvoice + 1.5 >= 1
  68     s0=2
  69     w0=2*pi
  70     scaled=FALSE
  71     s0log = as.integer( (log2( s0*w0/(2*pi) ) - 1) * nvoice + 1.5 )
  72     totnoct = noctave + as.integer(s0log/nvoice) + 1
  73
  74     # (normalized) observations node with CWT
  75     Xcwt4 <- lapply(seq_len(n), function(i) {
  76         ts <- scale(ts(curves[i,]), center=TRUE, scale=scaled)
  77         totts.cwt = Rwave::cwt(ts,totnoct,nvoice,w0,plot=0)
  78         ts.cwt = totts.cwt[,s0log:(s0log+noctave*nvoice)]
  79         #Normalization
  80         sqs <- sqrt(2^(0:(noctave*nvoice)/nvoice)*s0)
  81         sqres <- sweep(ts.cwt,MARGIN=2,sqs,'*')
  82         sqres / max(Mod(sqres))
  83     })
  84
  85     Xwer_dist <- matrix(0., n, n)
  86     fcoefs = rep(1/3, 3) #moving average on 3 values (TODO: very slow! correct?!)
  87     for (i in 1:(n-1))
  88     {
  89         for (j in (i+1):n)
  90         {
  91             #TODO: later, compute CWT here (because not enough storage space for 32M series)
  92             #      'circular=TRUE' is wrong, should just take values on the sides; to rewrite in C
  93             num <- filter(Mod(Xcwt4[[i]] * Conj(Xcwt4[[j]])), fcoefs, circular=TRUE)
  94             WX <- filter(Mod(Xcwt4[[i]] * Conj(Xcwt4[[i]])), fcoefs, circular=TRUE)
  95             WY <- filter(Mod(Xcwt4[[j]] * Conj(Xcwt4[[j]])), fcoefs, circular=TRUE)
  96             wer2    <- sum(colSums(num)^2) / sum( sum(colSums(WX) * colSums(WY)) )
  97             Xwer_dist[i,j] <- sqrt(delta * ncol(Xcwt4[[1]]) * (1 - wer2))
  98             Xwer_dist[j,i] <- Xwer_dist[i,j]
  99         }
 100     }
 101     diag(Xwer_dist) <- numeric(n)
 102     Xwer_dist
 103 }