epclust/src/WER.c

   1 #include <stdlib.h>
   2 #include <math.h>
   3 #include <stdbool.h>
   4
   5 #ifndef M_PI
   6 #define M_PI 3.14159265358979323846
   7 #endif
   8
   9 // n: number of synchrones, m: length of a synchrone
  10 float computeWerDist(float* s1, float* s2, int n, int m)
  11 {
  12     //TODO: automatic tune of all these parameters ? (for other users)
  13     int nvoice = 4;
  14     //noctave 2^13 = 8192 half hours ~ 180 days ; ~log2(ncol(synchrones))
  15     int noctave = 13
  16     // 4 here represent 2^5 = 32 half-hours ~ 1 day
  17     //NOTE: default scalevector == 2^(0:(noctave * nvoice) / nvoice) * s0 (?)
  18     //R: scalevector <- 2^(4:(noctave * nvoice) / nvoice + 1)
  19     int* scalevector = (int*)malloc( (noctave*nvoice-4 + 1) * sizeof(int))
  20     for (int i=4; i<=noctave*nvoice; i++)
  21         scalevector[i-4] = pow(2., (float)i/nvoice + 1.);
  22     //condition: ( log2(s0*w0/(2*pi)) - 1 ) * nvoice + 1.5 >= 1
  23     int s0 = 2;
  24     double w0 = 2*M_PI;
  25     bool scaled = false;
  26     int s0log = as.integer( (log2( s0*w0/(2*pi) ) - 1) * nvoice + 1.5 )
  27     int totnoct = noctave + as.integer(s0log/nvoice) + 1
  28
  29
  30
  31
  32
  33 ///TODO: continue
  34
  35
  36
  37     computeCWT = function(i)
  38     {
  39         if (verbose)
  40             cat(paste("+++ Compute Rwave::cwt() on serie ",i,"\n", sep=""))
  41         ts <- scale(ts(synchrones[i,]), center=TRUE, scale=scaled)
  42         totts.cwt = Rwave::cwt(ts,totnoct,nvoice,w0,plot=0)
  43         ts.cwt = totts.cwt[,s0log:(s0log+noctave*nvoice)]
  44         #Normalization
  45         sqs <- sqrt(2^(0:(noctave*nvoice)/nvoice)*s0)
  46         sqres <- sweep(ts.cwt,2,sqs,'*')
  47         sqres / max(Mod(sqres))
  48     }
  49
  50     if (parll)
  51     {
  52         cl = parallel::makeCluster(ncores_clust)
  53         parallel::clusterExport(cl,
  54             varlist=c("synchrones","totnoct","nvoice","w0","s0log","noctave","s0","verbose"),
  55             envir=environment())
  56     }
  57
  58     # (normalized) observations node with CWT
  59     Xcwt4 <-
  60         if (parll)
  61             parallel::parLapply(cl, seq_len(n), computeCWT)
  62         else
  63             lapply(seq_len(n), computeCWT)
  64
  65     if (parll)
  66         parallel::stopCluster(cl)
  67
  68     Xwer_dist <- bigmemory::big.matrix(nrow=n, ncol=n, type="double")
  69     fcoefs = rep(1/3, 3) #moving average on 3 values (TODO: very slow! correct?!)
  70     if (verbose)
  71         cat("*** Compute WER distances from CWT\n")
  72
  73     #TODO: computeDistances(i,j), et répartir les n(n-1)/2 couples d'indices
  74     #là c'est trop déséquilibré
  75
  76     computeDistancesLineI = function(i)
  77     {
  78         if (verbose)
  79             cat(paste("   Line ",i,"\n", sep=""))
  80         for (j in (i+1):n)
  81         {
  82             #TODO: 'circular=TRUE' is wrong, should just take values on the sides; to rewrite in C
  83             num <- filter(Mod(Xcwt4[[i]] * Conj(Xcwt4[[j]])), fcoefs, circular=TRUE)
  84             WX <- filter(Mod(Xcwt4[[i]] * Conj(Xcwt4[[i]])), fcoefs, circular=TRUE)
  85             WY <- filter(Mod(Xcwt4[[j]] * Conj(Xcwt4[[j]])), fcoefs, circular=TRUE)
  86             wer2    <- sum(colSums(num)^2) / sum( sum(colSums(WX) * colSums(WY)) )
  87             if (parll)
  88                 synchronicity::lock(m)
  89             Xwer_dist[i,j] <- sqrt(delta * ncol(Xcwt4[[1]]) * (1 - wer2))
  90             Xwer_dist[j,i] <- Xwer_dist[i,j]
  91             if (parll)
  92                 synchronicity::unlock(m)
  93         }
  94         Xwer_dist[i,i] = 0.
  95     }
  96
  97     parll = (requireNamespace("synchronicity",quietly=TRUE)
  98         && parll && Sys.info()['sysname'] != "Windows")
  99     if (parll)
 100         m <- synchronicity::boost.mutex()
 101
 102     ignored <-
 103         if (parll)
 104         {
 105             parallel::mclapply(seq_len(n-1), computeDistancesLineI,
 106                 mc.cores=ncores_clust, mc.allow.recursive=FALSE)
 107         }
 108         else
 109             lapply(seq_len(n-1), computeDistancesLineI)
 110     Xwer_dist[n,n] = 0.
 111
 112     mat_dists = matrix(nrow=n, ncol=n)
 113     #TODO: avoid this loop?
 114     for (i in 1:n)
 115         mat_dists[i,] = Xwer_dist[i,]
 116     mat_dists
 117