epclust/R/clustering.R

   1 oneIteration = function(..........)
   2 {
   3         cl_clust = parallel::makeCluster(ncores_clust)
   4         parallel::clusterExport(cl_clust, .............., envir=........)
   5         indices_clust = indices_task[[i]]
   6         repeat
   7         {
   8             nb_workers = max( 1, round( length(indices_clust) / nb_series_per_chunk ) )
   9             indices_workers = list()
  10             #indices[[i]] == (start_index,number_of_elements)
  11             for (i in 1:nb_workers)
  12             {
  13                 upper_bound = ifelse( i<nb_workers,
  14                     min(nb_series_per_chunk*i,length(indices_clust)), length(indices_clust) )
  15                 indices_workers[[i]] = indices_clust[(nb_series_per_chunk*(i-1)+1):upper_bound]
  16             }
  17             indices_clust = parallel::parSapply(cl, indices_workers, processChunk, K1, K2*(WER=="mix"))
  18             if ( (WER=="end" && length(indices_clust) == K1) ||
  19                 (WER=="mix" && length(indices_clust) == K2) )
  20             {
  21                 break
  22             }
  23         }
  24         parallel::stopCluster(cl_clust)
  25         res_clust
  26 }
  27
  28 processChunk = function(indices, K1, K2)
  29 {
  30     #1) retrieve data (coeffs)
  31     coeffs = getCoeffs(indices)
  32     #2) cluster
  33     cl = computeClusters(as.matrix(coeffs[,2:ncol(coeffs)]), K1)
  34     #3) WER (optional)
  35     if (K2 > 0)
  36     {
  37         curves = computeSynchrones(cl)
  38         dists = computeWerDists(curves)
  39         cl = computeClusters(dists, K2)
  40     }
  41     cl
  42 }
  43
  44 computeClusters = function(data, K)
  45 {
  46     library(cluster)
  47     pam_output = cluster::pam(data, K)
  48     return ( list( clusts=pam_output$clustering, medoids=pam_output$medoids,
  49         ranks=pam_output$id.med ) )
  50 }
  51
  52 #TODO: appendCoeffs() en C --> serialize et append to file
  53
  54 computeSynchrones = function(...)
  55 {
  56
  57 }
  58
  59 #Entrée : courbes synchrones, soit après étape 1 itérée, soit après chaqure étape 1
  60 computeWerDist = function(conso)
  61 {
  62     if (!require("Rwave", quietly=TRUE))
  63         stop("Unable to load Rwave library")
  64     n <- nrow(conso)
  65     delta <- ncol(conso)
  66     #TODO: automatic tune of all these parameters ? (for other users)
  67     nvoice   <- 4
  68     # noctave = 2^13 = 8192 half hours ~ 180 days ; ~log2(ncol(conso))
  69     noctave = 13
  70     # 4 here represent 2^5 = 32 half-hours ~ 1 day
  71     #NOTE: default scalevector == 2^(0:(noctave * nvoice) / nvoice) * s0 (?)
  72     scalevector  <- 2^(4:(noctave * nvoice) / nvoice) * 2
  73     #condition: ( log2(s0*w0/(2*pi)) - 1 ) * nvoice + 1.5 >= 1
  74     s0=2
  75     w0=2*pi
  76     scaled=FALSE
  77     s0log = as.integer( (log2( s0*w0/(2*pi) ) - 1) * nvoice + 1.5 )
  78     totnoct = noctave + as.integer(s0log/nvoice) + 1
  79
  80     # (normalized) observations node with CWT
  81     Xcwt4 <- lapply(seq_len(n), function(i) {
  82         ts <- scale(ts(conso[i,]), center=TRUE, scale=scaled)
  83         totts.cwt = Rwave::cwt(ts,totnoct,nvoice,w0,plot=0)
  84         ts.cwt = totts.cwt[,s0log:(s0log+noctave*nvoice)]
  85         #Normalization
  86         sqs <- sqrt(2^(0:(noctave*nvoice)/nvoice)*s0)
  87         sqres <- sweep(ts.cwt,MARGIN=2,sqs,'*')
  88         sqres / max(Mod(sqres))
  89     })
  90
  91     Xwer_dist <- matrix(0., n, n)
  92     fcoefs = rep(1/3, 3) #moving average on 3 values (TODO: very slow! correct?!)
  93     for (i in 1:(n-1))
  94     {
  95         for (j in (i+1):n)
  96         {
  97             #TODO: later, compute CWT here (because not enough storage space for 32M series)
  98             #      'circular=TRUE' is wrong, should just take values on the sides; to rewrite in C
  99             num <- filter(Mod(Xcwt4[[i]] * Conj(Xcwt4[[j]])), fcoefs, circular=TRUE)
 100             WX <- filter(Mod(Xcwt4[[i]] * Conj(Xcwt4[[i]])), fcoefs, circular=TRUE)
 101             WY <- filter(Mod(Xcwt4[[j]] * Conj(Xcwt4[[j]])), fcoefs, circular=TRUE)
 102             wer2    <- sum(colSums(num)^2) / sum( sum(colSums(WX) * colSums(WY)) )
 103             Xwer_dist[i,j] <- sqrt(delta * ncol(Xcwt4[[1]]) * (1 - wer2))
 104             Xwer_dist[j,i] <- Xwer_dist[i,j]
 105         }
 106     }
 107     diag(Xwer_dist) <- numeric(n)
 108     Xwer_dist
 109 }