-# Cluster one full task (nb_curves / ntasks series); only step 1
-clusteringTask = function(indices, getCoefs, K1, nb_series_per_chunk, ncores)
+#' @name clustering
+#' @rdname clustering
+#' @aliases clusteringTask1 computeClusters1 computeClusters2
+#'
+#' @title Two-stage clustering, withing one task (see \code{claws()})
+#'
+#' @description \code{clusteringTask1()} runs one full stage-1 task, which consists in
+#' iterated stage 1 clustering (on nb_curves / ntasks energy contributions, computed
+#' through discrete wavelets coefficients).
+#' \code{clusteringTask2()} runs a full stage-2 task, which consists in synchrones
+#' and then WER distances computations, before applying the clustering algorithm.
+#' \code{computeClusters1()} and \code{computeClusters2()} correspond to the atomic
+#' clustering procedures respectively for stage 1 and 2. The former applies the
+#' clustering algorithm (PAM) on a contributions matrix, while the latter clusters
+#' a chunk of series inside one task (~max nb_series_per_chunk)
+#'
+#' @param indices Range of series indices to cluster in parallel (initial data)
+#' @param getContribs Function to retrieve contributions from initial series indices:
+#' \code{getContribs(indices)} outpus a contributions matrix
+#' @param contribs matrix of contributions (e.g. output of \code{curvesToContribs()})
+#' @inheritParams computeSynchrones
+#' @inheritParams claws
+#'
+#' @return For \code{clusteringTask1()} and \code{computeClusters1()}, the indices of the
+#' computed (K1) medoids. Indices are irrelevant for stage 2 clustering, thus
+#' \code{computeClusters2()} outputs a big.matrix of medoids
+#' (of size limited by nb_series_per_chunk)
+NULL
+
+#' @rdname clustering
+#' @export
+clusteringTask1 = function(
+ indices, getContribs, K1, nb_series_per_chunk, ncores_clust=1, verbose=FALSE, parll=TRUE)
{
- cl = parallel::makeCluster(ncores)
- repeat
+ if (verbose)
+ cat(paste("*** Clustering task 1 on ",length(indices)," lines\n", sep=""))
+
+ if (parll)
+ {
+ cl = parallel::makeCluster(ncores_clust)
+ parallel::clusterExport(cl, varlist=c("getContribs","K1","verbose"), envir=environment())
+ }
+ while (length(indices) > K1)
{
- nb_workers = max( 1, round( length(indices) / nb_series_per_chunk ) )
- indices_workers = lapply(seq_len(nb_workers), function(i) {
- upper_bound = ifelse( i<nb_workers,
- min(nb_series_per_chunk*i,length(indices)), length(indices) )
- indices[(nb_series_per_chunk*(i-1)+1):upper_bound]
- })
- indices = unlist( parallel::parLapply(cl, indices_workers, function(inds)
- computeClusters1(getCoefs(inds), K1)) )
- if (length(indices) == K1)
- break
+ indices_workers = .spreadIndices(indices, nb_series_per_chunk)
+ indices <-
+ if (parll)
+ {
+ unlist( parallel::parLapply(cl, indices_workers, function(inds) {
+ require("epclust", quietly=TRUE)
+ inds[ computeClusters1(getContribs(inds), K1, verbose) ]
+ }) )
+ }
+ else
+ {
+ unlist( lapply(indices_workers, function(inds)
+ inds[ computeClusters1(getContribs(inds), K1, verbose) ]
+ ) )
+ }
}
- parallel::stopCluster(cl)
+ if (parll)
+ parallel::stopCluster(cl)
+
indices #medoids
}
-# Apply the clustering algorithm (PAM) on a coeffs or distances matrix
-computeClusters1 = function(coefs, K1)
- indices[ cluster::pam(coefs, K1, diss=FALSE)$id.med ]
+#' @rdname clustering
+#' @export
+clusteringTask2 = function(medoids, K2,
+ getRefSeries, nb_ref_curves, nb_series_per_chunk, ncores_clust=1,verbose=FALSE,parll=TRUE)
+{
+ if (verbose)
+ cat(paste("*** Clustering task 2 on ",nrow(medoids)," lines\n", sep=""))
+
+ if (nrow(medoids) <= K2)
+ return (medoids)
+ synchrones = computeSynchrones(medoids,
+ getRefSeries, nb_ref_curves, nb_series_per_chunk, ncores_clust, verbose, parll)
+ distances = computeWerDists(synchrones, ncores_clust, verbose, parll)
+ # PAM in package 'cluster' cannot take big.matrix in input: need to cast it
+ medoids[ computeClusters2(distances[,],K2,verbose), ]
+}
+
+#' @rdname clustering
+#' @export
+computeClusters1 = function(contribs, K1, verbose=FALSE)
+{
+ if (verbose)
+ cat(paste(" computeClusters1() on ",nrow(contribs)," lines\n", sep=""))
+ cluster::pam(contribs, K1, diss=FALSE)$id.med
+}
-# Cluster a chunk of series inside one task (~max nb_series_per_chunk)
-computeClusters2 = function(medoids, K2, getRefSeries, nb_series_per_chunk)
+#' @rdname clustering
+#' @export
+computeClusters2 = function(distances, K2, verbose=FALSE)
{
- synchrones = computeSynchrones(medoids, getRefSeries, nb_series_per_chunk)
- cluster::pam(computeWerDists(synchrones), K2, diss=TRUE)$medoids
+ if (verbose)
+ cat(paste(" computeClusters2() on ",nrow(distances)," lines\n", sep=""))
+ cluster::pam(distances, K2, diss=TRUE)$id.med
}
-# Compute the synchrones curves (sum of clusters elements) from a clustering result
-computeSynchrones = function(medoids, getRefSeries, nb_series_per_chunk)
+#' computeSynchrones
+#'
+#' Compute the synchrones curves (sum of clusters elements) from a matrix of medoids,
+#' using L2 distances.
+#'
+#' @param medoids big.matrix of medoids (curves of same length as initial series)
+#' @param getRefSeries Function to retrieve initial series (e.g. in stage 2 after series
+#' have been replaced by stage-1 medoids)
+#' @param nb_ref_curves How many reference series? (This number is known at this stage)
+#' @inheritParams claws
+#'
+#' @return A big.matrix of size K1 x L where L = data_length
+#'
+#' @export
+computeSynchrones = function(medoids, getRefSeries,
+ nb_ref_curves, nb_series_per_chunk, ncores_clust=1,verbose=FALSE,parll=TRUE)
{
- #les getSeries(indices) sont les medoides --> init vect nul pour chacun, puis incr avec les
- #courbes (getSeriesForSynchrones) les plus proches... --> au sens de la norme L2 ?
- K = nrow(medoids)
- synchrones = matrix(0, nrow=K, ncol=ncol(medoids))
- counts = rep(0,K)
- index = 1
- repeat
+ if (verbose)
+ cat(paste("--- Compute synchrones\n", sep=""))
+
+ computeSynchronesChunk = function(indices)
{
- range = (index-1) + seq_len(nb_series_per_chunk)
- ref_series = getRefSeries(range)
- if (is.null(ref_series))
- break
+ ref_series = getRefSeries(indices)
+ nb_series = nrow(ref_series)
#get medoids indices for this chunk of series
- for (i in seq_len(nrow(ref_series)))
+
+ #TODO: debug this (address is OK but values are garbage: why?)
+# mi = .Call("computeMedoidsIndices", medoids@address, ref_series, PACKAGE="epclust")
+
+ #R-equivalent, requiring a matrix (thus potentially breaking "fit-in-memory" hope)
+ mat_meds = medoids[,]
+ mi = rep(NA,nb_series)
+ for (i in 1:nb_series)
+ mi[i] <- which.min( rowSums( sweep(mat_meds, 2, ref_series[i,], '-')^2 ) )
+ rm(mat_meds); gc()
+
+ for (i in seq_len(nb_series))
{
- j = which.min( rowSums( sweep(medoids, 2, series[i,], '-')^2 ) )
- synchrones[j,] = synchrones[j,] + series[i,]
- counts[j] = counts[j] + 1
+ if (parll)
+ synchronicity::lock(m)
+ synchrones[mi[i],] = synchrones[mi[i],] + ref_series[i,]
+ counts[mi[i],1] = counts[mi[i],1] + 1
+ if (parll)
+ synchronicity::unlock(m)
}
- index = index + nb_series_per_chunk
}
+
+ K = nrow(medoids) ; L = ncol(medoids)
+ # Use bigmemory (shared==TRUE by default) + synchronicity to fill synchrones in //
+ # TODO: if size > RAM (not our case), use file-backed big.matrix
+ synchrones = bigmemory::big.matrix(nrow=K, ncol=L, type="double", init=0.)
+ counts = bigmemory::big.matrix(nrow=K, ncol=1, type="double", init=0)
+ # synchronicity is only for Linux & MacOS; on Windows: run sequentially
+ parll = (requireNamespace("synchronicity",quietly=TRUE)
+ && parll && Sys.info()['sysname'] != "Windows")
+ if (parll)
+ m <- synchronicity::boost.mutex()
+
+ if (parll)
+ {
+ cl = parallel::makeCluster(ncores_clust)
+ parallel::clusterExport(cl,
+ varlist=c("synchrones","counts","verbose","medoids","getRefSeries"),
+ envir=environment())
+ }
+
+ indices_workers = .spreadIndices(seq_len(nb_ref_curves), nb_series_per_chunk)
+ browser()
+ ignored <-
+ if (parll)
+ parallel::parLapply(cl, indices_workers, computeSynchronesChunk)
+ else
+ lapply(indices_workers, computeSynchronesChunk)
+
+ if (parll)
+ parallel::stopCluster(cl)
+
+ #TODO: can we avoid this loop? ( synchrones = sweep(synchrones, 1, counts, '/') )
+ for (i in seq_len(K))
+ synchrones[i,] = synchrones[i,] / counts[i,1]
#NOTE: odds for some clusters to be empty? (when series already come from stage 2)
- sweep(synchrones, 1, counts, '/')
+ # ...maybe; but let's hope resulting K1' be still quite bigger than K2
+ noNA_rows = sapply(seq_len(K), function(i) all(!is.nan(synchrones[i,])))
+ if (all(noNA_rows))
+ return (synchrones)
+ # Else: some clusters are empty, need to slice synchrones
+ synchrones[noNA_rows,]
}
-# Compute the WER distance between the synchrones curves (in rows)
-computeWerDist = function(curves)
+#' computeWerDists
+#'
+#' Compute the WER distances between the synchrones curves (in rows), which are
+#' returned (e.g.) by \code{computeSynchrones()}
+#'
+#' @param synchrones A big.matrix of synchrones, in rows. The series have same length
+#' as the series in the initial dataset
+#' @inheritParams claws
+#'
+#' @return A big.matrix of size K1 x K1
+#'
+#' @export
+computeWerDists = function(synchrones, ncores_clust=1,verbose=FALSE,parll=TRUE)
{
- if (!require("Rwave", quietly=TRUE))
- stop("Unable to load Rwave library")
- n <- nrow(curves)
- delta <- ncol(curves)
+ if (verbose)
+ cat(paste("--- Compute WER dists\n", sep=""))
+
+ n <- nrow(synchrones)
+ delta <- ncol(synchrones)
#TODO: automatic tune of all these parameters ? (for other users)
nvoice <- 4
- # noctave = 2^13 = 8192 half hours ~ 180 days ; ~log2(ncol(curves))
+ # noctave = 2^13 = 8192 half hours ~ 180 days ; ~log2(ncol(synchrones))
noctave = 13
# 4 here represent 2^5 = 32 half-hours ~ 1 day
#NOTE: default scalevector == 2^(0:(noctave * nvoice) / nvoice) * s0 (?)
- scalevector <- 2^(4:(noctave * nvoice) / nvoice) * 2
+ scalevector <- 2^(4:(noctave * nvoice) / nvoice + 1)
#condition: ( log2(s0*w0/(2*pi)) - 1 ) * nvoice + 1.5 >= 1
s0=2
w0=2*pi
s0log = as.integer( (log2( s0*w0/(2*pi) ) - 1) * nvoice + 1.5 )
totnoct = noctave + as.integer(s0log/nvoice) + 1
- # (normalized) observations node with CWT
- Xcwt4 <- lapply(seq_len(n), function(i) {
- ts <- scale(ts(curves[i,]), center=TRUE, scale=scaled)
- totts.cwt = Rwave::cwt(ts,totnoct,nvoice,w0,plot=0)
+ Xwer_dist <- bigmemory::big.matrix(nrow=n, ncol=n, type="double")
+ fcoefs = rep(1/3, 3) #moving average on 3 values
+
+ # Generate n(n-1)/2 pairs for WER distances computations
+ pairs = list()
+ V = seq_len(n)
+ for (i in 1:n)
+ {
+ V = V[-1]
+ pairs = c(pairs, lapply(V, function(v) c(i,v)))
+ }
+
+ computeCWT = function(i)
+ {
+ ts <- scale(ts(synchrones[i,]), center=TRUE, scale=scaled)
+ totts.cwt = Rwave::cwt(ts, totnoct, nvoice, w0, plot=FALSE)
ts.cwt = totts.cwt[,s0log:(s0log+noctave*nvoice)]
#Normalization
sqs <- sqrt(2^(0:(noctave*nvoice)/nvoice)*s0)
- sqres <- sweep(ts.cwt,MARGIN=2,sqs,'*')
+ sqres <- sweep(ts.cwt,2,sqs,'*')
sqres / max(Mod(sqres))
- })
+ }
+
+ computeDistancesIJ = function(pair)
+ {
+ i = pair[1] ; j = pair[2]
+ if (verbose && j==i+1)
+ cat(paste(" Distances (",i,",",j,"), (",i,",",j+1,") ...\n", sep=""))
+ cwt_i = computeCWT(i)
+ cwt_j = computeCWT(j)
+ num <- .Call("filter", Mod(cwt_i * Conj(cwt_j)), PACKAGE="epclust")
+ WX <- .Call("filter", Mod(cwt_i * Conj(cwt_i)), PACKAGE="epclust")
+ WY <- .Call("filter", Mod(cwt_j * Conj(cwt_j)), PACKAGE="epclust")
+ wer2 <- sum(colSums(num)^2) / sum(colSums(WX) * colSums(WY))
+ Xwer_dist[i,j] <- sqrt(delta * ncol(cwt_i) * (1 - wer2))
+ Xwer_dist[j,i] <- Xwer_dist[i,j]
+ Xwer_dist[i,i] = 0.
+ }
+
+ if (parll)
+ {
+ cl = parallel::makeCluster(ncores_clust)
+ parallel::clusterExport(cl,
+ varlist=c("synchrones","totnoct","nvoice","w0","s0log","noctave","s0","verbose"),
+ envir=environment())
+ }
+
+ ignored <-
+ if (parll)
+ parallel::parLapply(cl, pairs, computeDistancesIJ)
+ else
+ lapply(pairs, computeDistancesIJ)
- Xwer_dist <- matrix(0., n, n)
- fcoefs = rep(1/3, 3) #moving average on 3 values (TODO: very slow! correct?!)
- for (i in 1:(n-1))
+ if (parll)
+ parallel::stopCluster(cl)
+
+ Xwer_dist[n,n] = 0.
+ Xwer_dist
+}
+
+# Helper function to divide indices into balanced sets
+.spreadIndices = function(indices, nb_per_chunk)
+{
+ L = length(indices)
+ nb_workers = floor( L / nb_per_chunk )
+ if (nb_workers == 0)
{
- for (j in (i+1):n)
+ # L < nb_series_per_chunk, simple case
+ indices_workers = list(indices)
+ }
+ else
+ {
+ indices_workers = lapply( seq_len(nb_workers), function(i)
+ indices[(nb_per_chunk*(i-1)+1):(nb_per_chunk*i)] )
+ # Spread the remaining load among the workers
+ rem = L %% nb_per_chunk
+ while (rem > 0)
{
- #TODO: later, compute CWT here (because not enough storage space for 200k series)
- # 'circular=TRUE' is wrong, should just take values on the sides; to rewrite in C
- num <- filter(Mod(Xcwt4[[i]] * Conj(Xcwt4[[j]])), fcoefs, circular=TRUE)
- WX <- filter(Mod(Xcwt4[[i]] * Conj(Xcwt4[[i]])), fcoefs, circular=TRUE)
- WY <- filter(Mod(Xcwt4[[j]] * Conj(Xcwt4[[j]])), fcoefs, circular=TRUE)
- wer2 <- sum(colSums(num)^2) / sum( sum(colSums(WX) * colSums(WY)) )
- Xwer_dist[i,j] <- sqrt(delta * ncol(Xcwt4[[1]]) * (1 - wer2))
- Xwer_dist[j,i] <- Xwer_dist[i,j]
+ index = rem%%nb_workers + 1
+ indices_workers[[index]] = c(indices_workers[[index]], indices[L-rem+1])
+ rem = rem - 1
}
}
- diag(Xwer_dist) <- numeric(n)
- Xwer_dist
+ indices_workers
}