epclust/R/clustering.R

   1 # Cluster one full task (nb_curves / ntasks series)
   2 clusteringTask = function(indices,getSeries,getSeriesForSynchrones,synchrones_file,
   3     getCoefs,K1,K2,nb_series_per_chunk,ncores,to_file,ftype)
   4 {
   5     cl = parallel::makeCluster(ncores)
   6     repeat
   7     {
   8         nb_workers = max( 1, round( length(indices) / nb_series_per_chunk ) )
   9         indices_workers = lapply(seq_len(nb_workers), function(i) {
  10             upper_bound = ifelse( i<nb_workers,
  11                 min(nb_series_per_chunk*i,length(indices)), length(indices) )
  12             indices[(nb_series_per_chunk*(i-1)+1):upper_bound]
  13         })
  14         indices = unlist( parallel::parLapply(cl, indices_workers, function(inds)
  15             computeClusters1(inds, getCoefs, K1)) )
  16         if (length(indices_clust) == K1)
  17             break
  18     }
  19     parallel::stopCluster(cl)
  20     if (K2 == 0)
  21         return (indices)
  22     computeClusters2(indices, K2, getSeries, getSeriesForSynchrones, to_file,
  23                                      nb_series_per_chunk,ftype)
  24     vector("integer",0)
  25 }
  26
  27 # Apply the clustering algorithm (PAM) on a coeffs or distances matrix
  28 computeClusters1 = function(indices, getCoefs, K1)
  29 {
  30     coefs = getCoefs(indices)
  31     indices[ cluster::pam(coefs, K1, diss=FALSE)$id.med ]
  32 }
  33
  34 # Cluster a chunk of series inside one task (~max nb_series_per_chunk)
  35 computeClusters2 = function(indices, K2, getSeries, getSeriesForSynchrones, to_file,
  36                                                         nb_series_per_chunk, ftype)
  37 {
  38     curves = computeSynchrones(indices, getSeries, getSeriesForSynchrones, nb_series_per_chunk)
  39     dists = computeWerDists(curves)
  40     medoids = cluster::pam(dists, K2, diss=TRUE)$medoids
  41     if (to_file)
  42     {
  43         serialize(medoids, synchrones_file, ftype, nb_series_per_chunk)
  44         return (NULL)
  45     }
  46     medoids
  47 }
  48
  49 # Compute the synchrones curves (sum of clusters elements) from a clustering result
  50 computeSynchrones = function(indices, getSeries, getSeriesForSynchrones, nb_series_per_chunk)
  51 {
  52     #les getSeries(indices) sont les medoides --> init vect nul pour chacun, puis incr avec les
  53     #courbes (getSeriesForSynchrones) les plus proches... --> au sens de la norme L2 ?
  54     medoids = getSeries(indices)
  55     K = nrow(medoids)
  56     synchrones = matrix(0, nrow=K, ncol=ncol(medoids))
  57     counts = rep(0,K)
  58     index = 1
  59     repeat
  60     {
  61         series = getSeriesForSynchrones((index-1)+seq_len(nb_series_per_chunk))
  62         if (is.null(series))
  63             break
  64         #get medoids indices for this chunk of series
  65         index = which.min( rowSums( sweep(medoids, 2, series[i,], '-')^2 ) )
  66         synchrones[index,] = synchrones[index,] + series[i,]
  67         counts[index] = counts[index] + 1
  68     }
  69     #NOTE: odds for some clusters to be empty? (when series already come from stage 2)
  70     synchrones = sweep(synchrones, 1, counts, '/')
  71 }
  72
  73 # Compute the WER distance between the synchrones curves (in rows)
  74 computeWerDist = function(curves)
  75 {
  76     if (!require("Rwave", quietly=TRUE))
  77         stop("Unable to load Rwave library")
  78     n <- nrow(curves)
  79     delta <- ncol(curves)
  80     #TODO: automatic tune of all these parameters ? (for other users)
  81     nvoice   <- 4
  82     # noctave = 2^13 = 8192 half hours ~ 180 days ; ~log2(ncol(curves))
  83     noctave = 13
  84     # 4 here represent 2^5 = 32 half-hours ~ 1 day
  85     #NOTE: default scalevector == 2^(0:(noctave * nvoice) / nvoice) * s0 (?)
  86     scalevector  <- 2^(4:(noctave * nvoice) / nvoice) * 2
  87     #condition: ( log2(s0*w0/(2*pi)) - 1 ) * nvoice + 1.5 >= 1
  88     s0=2
  89     w0=2*pi
  90     scaled=FALSE
  91     s0log = as.integer( (log2( s0*w0/(2*pi) ) - 1) * nvoice + 1.5 )
  92     totnoct = noctave + as.integer(s0log/nvoice) + 1
  93
  94     # (normalized) observations node with CWT
  95     Xcwt4 <- lapply(seq_len(n), function(i) {
  96         ts <- scale(ts(curves[i,]), center=TRUE, scale=scaled)
  97         totts.cwt = Rwave::cwt(ts,totnoct,nvoice,w0,plot=0)
  98         ts.cwt = totts.cwt[,s0log:(s0log+noctave*nvoice)]
  99         #Normalization
 100         sqs <- sqrt(2^(0:(noctave*nvoice)/nvoice)*s0)
 101         sqres <- sweep(ts.cwt,MARGIN=2,sqs,'*')
 102         sqres / max(Mod(sqres))
 103     })
 104
 105     Xwer_dist <- matrix(0., n, n)
 106     fcoefs = rep(1/3, 3) #moving average on 3 values (TODO: very slow! correct?!)
 107     for (i in 1:(n-1))
 108     {
 109         for (j in (i+1):n)
 110         {
 111             #TODO: later, compute CWT here (because not enough storage space for 200k series)
 112             #      'circular=TRUE' is wrong, should just take values on the sides; to rewrite in C
 113             num <- filter(Mod(Xcwt4[[i]] * Conj(Xcwt4[[j]])), fcoefs, circular=TRUE)
 114             WX <- filter(Mod(Xcwt4[[i]] * Conj(Xcwt4[[i]])), fcoefs, circular=TRUE)
 115             WY <- filter(Mod(Xcwt4[[j]] * Conj(Xcwt4[[j]])), fcoefs, circular=TRUE)
 116             wer2    <- sum(colSums(num)^2) / sum( sum(colSums(WX) * colSums(WY)) )
 117             Xwer_dist[i,j] <- sqrt(delta * ncol(Xcwt4[[1]]) * (1 - wer2))
 118             Xwer_dist[j,i] <- Xwer_dist[i,j]
 119         }
 120     }
 121     diag(Xwer_dist) <- numeric(n)
 122     Xwer_dist
 123 }