default values for computeWerDists
[epclust.git] / epclust / R / computeWerDists.R
CommitLineData
40f12a2f
BA
1#' computeWerDists
2#'
3fb6e823
BA
3#' Compute the WER distances between the series at specified indices, which are
4#' obtaind by \code{getSeries(indices)}
40f12a2f 5#'
3c5a4b08 6#' @param indices Range of series indices to cluster
dc86eb0c
BA
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
40f12a2f
BA
10#' @inheritParams claws
11#'
3c5a4b08 12#' @return A distances matrix of size K x K where K == length(indices)
40f12a2f
BA
13#'
14#' @export
c25d83a0
BA
15computeWerDists <- function(indices, getSeries, nb_series_per_chunk, smooth_lvl=3, nvoice=4,
16 nbytes=4, endian=.Platform$endian, ncores=3, verbose=FALSE, parll=TRUE)
40f12a2f 17{
3c5a4b08 18 n <- length(indices)
282342ba
BA
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
3c5a4b08
BA
21 # Since a CWT contains noctave*nvoice complex series, we deduce the number of CWT to
22 # retrieve/put in one chunk.
282342ba 23 nb_cwt_per_chunk <- max(1, floor(nb_series_per_chunk / (nvoice*noctave*2)))
40f12a2f 24
3c5a4b08 25 # Initialize result as a square big.matrix of size 'number of medoids'
40f12a2f
BA
26 Xwer_dist <- bigmemory::big.matrix(nrow=n, ncol=n, type="double")
27
282342ba 28 cwt_file <- tempfile(pattern="epclust_cwt.bin_")
3c5a4b08 29 # Compute the getSeries(indices) CWT, and store the results in the binary file
e0154a59 30 computeSaveCWT <- function(inds)
40f12a2f 31 {
e0154a59
BA
32 if (verbose)
33 cat(" Compute save CWT on ",length(inds)," indices\n", sep="")
34
40f12a2f 35 # Obtain CWT as big vectors of real part + imaginary part (concatenate)
e0154a59 36 ts_cwt <- sapply(inds, function(i) {
3c5a4b08 37 ts <- scale(ts(getSeries(i)), center=TRUE, scale=FALSE)
40f12a2f
BA
38 ts_cwt <- Rwave::cwt(ts, noctave, nvoice, w0=2*pi, twoD=TRUE, plot=FALSE)
39 c( as.double(Re(ts_cwt)),as.double(Im(ts_cwt)) )
40 })
41
42 # Serialization
3c5a4b08 43 binarize(ts_cwt, cwt_file, nb_cwt_per_chunk, ",", nbytes, endian)
40f12a2f
BA
44 }
45
40f12a2f 46 # Function to retrieve a synchrone CWT from (binary) file
282342ba 47 getCWT <- function(index, L)
40f12a2f
BA
48 {
49 flat_cwt <- getDataInFile(index, cwt_file, nbytes, endian)
282342ba
BA
50 cwt_length <- length(flat_cwt) / 2
51 re_part <- as.matrix(flat_cwt[1:cwt_length], nrow=L)
52 im_part <- as.matrix(flat_cwt[(cwt_length+1):(2*cwt_length)], nrow=L)
40f12a2f
BA
53 re_part + 1i * im_part
54 }
55
dc86eb0c 56 # Compute distances between columns i and j for j>i
282342ba 57 computeDistances <- function(i)
40f12a2f
BA
58 {
59 if (parll)
60 {
61 # parallel workers start with an empty environment
40f12a2f 62 require("epclust", quietly=TRUE)
40f12a2f
BA
63 Xwer_dist <- bigmemory::attach.big.matrix(Xwer_dist_desc)
64 }
65
e0154a59 66 if (verbose)
282342ba 67 cat(paste(" Distances from ",i," to ",i+1,"...",n,"\n", sep=""))
40f12a2f 68
282342ba 69 # Get CWT of column i, and run computations for columns j>i
3c5a4b08 70 cwt_i <- getCWT(i, L)
282342ba
BA
71 WX <- filterMA(Mod(cwt_i * Conj(cwt_i)), smooth_lvl)
72
73 for (j in (i+1):n)
74 {
75 cwt_j <- getCWT(j, L)
40f12a2f 76
282342ba 77 # Compute the ratio of integrals formula 5.6 for WER^2
3fb6e823 78 # in https://arxiv.org/abs/1101.4744v2 paragraph 5.3
282342ba
BA
79 num <- filterMA(Mod(cwt_i * Conj(cwt_j)), smooth_lvl)
80 WY <- filterMA(Mod(cwt_j * Conj(cwt_j)), smooth_lvl)
81 wer2 <- sum(colSums(num)^2) / sum(colSums(WX) * colSums(WY))
40f12a2f 82
282342ba
BA
83 Xwer_dist[i,j] <- sqrt(L * ncol(cwt_i) * (1 - wer2))
84 Xwer_dist[j,i] <- Xwer_dist[i,j]
85 }
40f12a2f
BA
86 Xwer_dist[i,i] <- 0.
87 }
88
dc86eb0c
BA
89 if (verbose)
90 cat(paste("--- Precompute and serialize synchrones CWT\n", sep=""))
91
92 # Split indices by packets of length at most nb_cwt_per_chunk
e0154a59 93 indices_cwt <- .splitIndices(indices, nb_cwt_per_chunk)
dc86eb0c
BA
94 # NOTE: next loop could potentially be run in //. Indices would be permuted (by
95 # serialization order), and synchronicity would be required because of concurrent
e0154a59 96 # writes. Probably not worth the effort - but possible.
dc86eb0c
BA
97 for (inds in indices_cwt)
98 computeSaveCWT(inds)
99
282342ba
BA
100 if (parll)
101 {
102 # outfile=="" to see stderr/stdout on terminal
3fb6e823
BA
103 cl <-
104 if (verbose)
dc86eb0c 105 parallel::makeCluster(ncores, outfile="")
3fb6e823 106 else
dc86eb0c 107 parallel::makeCluster(ncores)
282342ba 108 Xwer_dist_desc <- bigmemory::describe(Xwer_dist)
dc86eb0c
BA
109 parallel::clusterExport(cl, envir=environment(),
110 varlist=c("parll","n","L","Xwer_dist_desc","getCWT","verbose"))
282342ba
BA
111 }
112
40f12a2f
BA
113 if (verbose)
114 cat(paste("--- Compute WER distances\n", sep=""))
115
116 ignored <-
117 if (parll)
282342ba 118 parallel::parLapply(cl, seq_len(n-1), computeDistances)
40f12a2f 119 else
282342ba
BA
120 lapply(seq_len(n-1), computeDistances)
121 Xwer_dist[n,n] <- 0.
40f12a2f
BA
122
123 if (parll)
124 parallel::stopCluster(cl)
125
282342ba 126 unlink(cwt_file) #remove binary file
40f12a2f 127
40f12a2f
BA
128 Xwer_dist[,] #~small matrix K1 x K1
129}