pkg/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 Indices of the series to consider
   7 #' @param getSeries Function to retrieve series (argument: 'inds', 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=3,
  16     nvoice=4, nbytes=4, endian=.Platform$endian, ncores=3, verbose=FALSE)
  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     shift <- 1 #roughly equivalent to s0 in biwavelets & cie. TODO: as arg?
  29
  30     cwt_file <- tempfile(pattern="epclust_cwt.bin_")
  31     # Compute the getSeries(indices) CWT, and store the results in the binary file
  32     computeSaveCWT <- function(inds)
  33     {
  34         if (verbose)
  35             cat("   Compute save CWT on ",length(inds)," indices\n", sep="")
  36
  37         # Obtain CWT as big vectors of real part + imaginary part (concatenate)
  38         ts_cwt <- sapply(inds, function(i) {
  39             ts <- scale(ts(getSeries(i)), center=TRUE, scale=FALSE)
  40             ts_cwt <- Rwave::cwt(ts, noctave+ceiling(shift/nvoice), nvoice,
  41                 w0=2*pi, twoD=TRUE, plot=FALSE)
  42             ts_cwt <- ts_cwt[,(1+shift):(noctave*nvoice+shift)]
  43             c( as.double(Re(ts_cwt)),as.double(Im(ts_cwt)) )
  44         })
  45
  46         # Serialization
  47         binarize(ts_cwt, cwt_file, nb_cwt_per_chunk, ",", nbytes, endian)
  48     }
  49
  50     # Function to retrieve a synchrone CWT from (binary) file
  51     getCWT <- function(index, L)
  52     {
  53         flat_cwt <- getDataInFile(index, cwt_file, nbytes, endian)
  54         cwt_length <- length(flat_cwt) / 2
  55         re_part <- as.matrix(flat_cwt[1:cwt_length], nrow=L)
  56         im_part <- as.matrix(flat_cwt[(cwt_length+1):(2*cwt_length)], nrow=L)
  57         re_part + 1i * im_part
  58     }
  59
  60     # Compute distances between columns i and j for j>i
  61     computeDistances <- function(i)
  62     {
  63         if (parll)
  64         {
  65             # parallel workers start with an empty environment
  66             require("epclust", quietly=TRUE)
  67             Xwer_dist <- bigmemory::attach.big.matrix(Xwer_dist_desc)
  68         }
  69
  70         if (verbose)
  71             cat(paste("   Distances from ",i," to ",i+1,"...",n,"\n", sep=""))
  72
  73         # Get CWT of column i, and run computations for columns j>i
  74         cwt_i <- getCWT(i, L)
  75         WX  <- filterMA(Mod(cwt_i * Conj(cwt_i)), smooth_lvl)
  76
  77         for (j in (i+1):n)
  78         {
  79             cwt_j <- getCWT(j, L)
  80
  81             # Compute the ratio of integrals formula 5.6 for WER^2
  82             # in https://arxiv.org/abs/1101.4744v2 paragraph 5.3
  83             num <- filterMA(Mod(cwt_i * Conj(cwt_j)), smooth_lvl)
  84             WY <- filterMA(Mod(cwt_j * Conj(cwt_j)), smooth_lvl)
  85             wer2 <- sum(colSums(num)^2) / sum(colSums(WX) * colSums(WY))
  86
  87             Xwer_dist[i,j] <- sqrt(L * ncol(cwt_i) * (1 - wer2))
  88             Xwer_dist[j,i] <- Xwer_dist[i,j]
  89         }
  90         Xwer_dist[i,i] <- 0.
  91     }
  92
  93     if (verbose)
  94         cat(paste("--- Precompute and serialize synchrones CWT\n", sep=""))
  95
  96     # Split indices by packets of length at most nb_cwt_per_chunk
  97     indices_cwt <- .splitIndices(indices, nb_cwt_per_chunk)
  98     # NOTE: next loop could potentially be run in //. Indices would be permuted (by
  99     # serialization order), and synchronicity would be required because of concurrent
 100     # writes. Probably not worth the effort - but possible.
 101     for (inds in indices_cwt)
 102         computeSaveCWT(inds)
 103
 104     parll <- (ncores > 1)
 105     if (parll)
 106     {
 107         # outfile=="" to see stderr/stdout on terminal
 108         cl <-
 109             if (verbose)
 110                 parallel::makeCluster(ncores, outfile="")
 111             else
 112                 parallel::makeCluster(ncores)
 113         Xwer_dist_desc <- bigmemory::describe(Xwer_dist)
 114         parallel::clusterExport(cl, envir=environment(),
 115             varlist=c("parll","n","L","Xwer_dist_desc","getCWT","verbose"))
 116     }
 117
 118     if (verbose)
 119         cat(paste("--- Compute WER distances\n", sep=""))
 120
 121     ignored <-
 122         if (parll)
 123             parallel::clusterApplyLB(cl, seq_len(n-1), computeDistances)
 124         else
 125             lapply(seq_len(n-1), computeDistances)
 126     Xwer_dist[n,n] <- 0.
 127
 128     if (parll)
 129         parallel::stopCluster(cl)
 130
 131     unlink(cwt_file) #remove binary file
 132
 133     Xwer_dist[,] #~small matrix K1 x K1
 134 }