epclust/R/computeWerDists.R

   1 #' computeWerDists
   2 #'
   3 #' Compute the WER distances between the series at specified indices, which are
   4 #' obtaind by \code{getSeries(indices)}
   5 #'
   6 #' @param indices Range of series indices to cluster
   7 #' @param getSeries Function to retrieve series (argument: 'indices', integer vector),
   8 #'   as columns of a matrix
   9 #' @param ncores Number of cores for parallel runs
  10 #' @inheritParams claws
  11 #'
  12 #' @return A distances matrix of size K x K where K == length(indices)
  13 #'
  14 #' @export
  15 computeWerDists <- function(indices, getSeries, nb_series_per_chunk, smooth_lvl, nvoice,
  16     nbytes, endian, ncores=3, verbose=FALSE, parll=TRUE)
  17 {
  18     n <- length(indices)
  19     L <- length(getSeries(1)) #TODO: not very neat way to get L
  20     noctave <- ceiling(log2(L)) #min power of 2 to cover serie range
  21     # Since a CWT contains noctave*nvoice complex series, we deduce the number of CWT to
  22     # retrieve/put in one chunk.
  23     nb_cwt_per_chunk <- max(1, floor(nb_series_per_chunk / (nvoice*noctave*2)))
  24
  25     # Initialize result as a square big.matrix of size 'number of medoids'
  26     Xwer_dist <- bigmemory::big.matrix(nrow=n, ncol=n, type="double")
  27
  28     cwt_file <- tempfile(pattern="epclust_cwt.bin_")
  29     # Compute the getSeries(indices) CWT, and store the results in the binary file
  30     computeSaveCWT <- function(indices)
  31     {
  32         # Obtain CWT as big vectors of real part + imaginary part (concatenate)
  33         ts_cwt <- sapply(indices, function(i) {
  34             ts <- scale(ts(getSeries(i)), center=TRUE, scale=FALSE)
  35             ts_cwt <- Rwave::cwt(ts, noctave, nvoice, w0=2*pi, twoD=TRUE, plot=FALSE)
  36             c( as.double(Re(ts_cwt)),as.double(Im(ts_cwt)) )
  37         })
  38
  39         # Serialization
  40         binarize(ts_cwt, cwt_file, nb_cwt_per_chunk, ",", nbytes, endian)
  41     }
  42
  43     # Function to retrieve a synchrone CWT from (binary) file
  44     getCWT <- function(index, L)
  45     {
  46         flat_cwt <- getDataInFile(index, cwt_file, nbytes, endian)
  47         cwt_length <- length(flat_cwt) / 2
  48         re_part <- as.matrix(flat_cwt[1:cwt_length], nrow=L)
  49         im_part <- as.matrix(flat_cwt[(cwt_length+1):(2*cwt_length)], nrow=L)
  50         re_part + 1i * im_part
  51     }
  52
  53     # Compute distances between columns i and j for j>i
  54     computeDistances <- function(i)
  55     {
  56         if (parll)
  57         {
  58             # parallel workers start with an empty environment
  59             require("epclust", quietly=TRUE)
  60             Xwer_dist <- bigmemory::attach.big.matrix(Xwer_dist_desc)
  61         }
  62
  63         if (verbose && !parll)
  64             cat(paste("   Distances from ",i," to ",i+1,"...",n,"\n", sep=""))
  65
  66         # Get CWT of column i, and run computations for columns j>i
  67         cwt_i <- getCWT(i, L)
  68         WX  <- filterMA(Mod(cwt_i * Conj(cwt_i)), smooth_lvl)
  69
  70         for (j in (i+1):n)
  71         {
  72             cwt_j <- getCWT(j, L)
  73
  74             # Compute the ratio of integrals formula 5.6 for WER^2
  75             # in https://arxiv.org/abs/1101.4744v2 paragraph 5.3
  76             num <- filterMA(Mod(cwt_i * Conj(cwt_j)), smooth_lvl)
  77             WY <- filterMA(Mod(cwt_j * Conj(cwt_j)), smooth_lvl)
  78             wer2 <- sum(colSums(num)^2) / sum(colSums(WX) * colSums(WY))
  79
  80             Xwer_dist[i,j] <- sqrt(L * ncol(cwt_i) * (1 - wer2))
  81             Xwer_dist[j,i] <- Xwer_dist[i,j]
  82         }
  83         Xwer_dist[i,i] <- 0.
  84     }
  85
  86     if (verbose)
  87         cat(paste("--- Precompute and serialize synchrones CWT\n", sep=""))
  88
  89     # Split indices by packets of length at most nb_cwt_per_chunk
  90     indices_cwt <- .splitIndices(seq_len(n), nb_cwt_per_chunk)
  91     # NOTE: next loop could potentially be run in //. Indices would be permuted (by
  92     # serialization order), and synchronicity would be required because of concurrent
  93     # writes. Probably not worth the effort - but possible to gain some bits of speed.
  94     for (inds in indices_cwt)
  95         computeSaveCWT(inds)
  96
  97     if (parll)
  98     {
  99         # outfile=="" to see stderr/stdout on terminal
 100         cl <-
 101             if (verbose)
 102                 parallel::makeCluster(ncores, outfile="")
 103             else
 104                 parallel::makeCluster(ncores)
 105         Xwer_dist_desc <- bigmemory::describe(Xwer_dist)
 106         parallel::clusterExport(cl, envir=environment(),
 107             varlist=c("parll","n","L","Xwer_dist_desc","getCWT","verbose"))
 108     }
 109
 110     if (verbose)
 111         cat(paste("--- Compute WER distances\n", sep=""))
 112
 113     ignored <-
 114         if (parll)
 115             parallel::parLapply(cl, seq_len(n-1), computeDistances)
 116         else
 117             lapply(seq_len(n-1), computeDistances)
 118     Xwer_dist[n,n] <- 0.
 119
 120     if (parll)
 121         parallel::stopCluster(cl)
 122
 123     unlink(cwt_file) #remove binary file
 124
 125     Xwer_dist[,] #~small matrix K1 x K1
 126 }