-# 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 clusteringTask2 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
+#' first clustering algorithm on a contributions matrix, while the latter clusters
+#' a set of series inside one task (~nb_items_clust1)
+#'
+#' @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
+#' @inheritParams computeSynchrones
+#' @inheritParams claws
+#'
+#' @return For \code{clusteringTask1()}, the indices of the computed (K1) medoids.
+#' Indices are irrelevant for stage 2 clustering, thus \code{clusteringTask2()}
+#' outputs a big.matrix of medoids (of size LxK2, K2 = final number of clusters)
+NULL
+
+#' @rdname clustering
+#' @export
+clusteringTask1 = function(indices, getContribs, K1, algoClust1, nb_items_clust1,
+ ncores_clust=1, verbose=FALSE, parll=TRUE)
{
- cl = parallel::makeCluster(ncores)
- repeat
+ if (parll)
+ {
+ cl = parallel::makeCluster(ncores_clust, outfile = "")
+ parallel::clusterExport(cl, c("getContribs","K1","verbose"), envir=environment())
+ }
+ # Iterate clustering algorithm 1 until K1 medoids are found
+ 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
+ # Balance tasks by splitting the indices set - as evenly as possible
+ indices_workers = .spreadIndices(indices, nb_items_clust1)
+ if (verbose)
+ cat(paste("*** [iterated] Clustering task 1 on ",length(indices)," series\n", sep=""))
+ indices <-
+ if (parll)
+ {
+ unlist( parallel::parLapply(cl, indices_workers, function(inds) {
+ require("epclust", quietly=TRUE)
+ inds[ algoClust1(getContribs(inds), K1) ]
+ }) )
+ }
+ else
+ {
+ unlist( lapply(indices_workers, function(inds)
+ inds[ algoClust1(getContribs(inds), K1) ]
+ ) )
+ }
}
- 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)$id.med ]
+#' @rdname clustering
+#' @export
+clusteringTask2 = function(medoids, K2, algoClust2, getRefSeries, nb_ref_curves,
+ nb_series_per_chunk, nvoice, nbytes,endian,ncores_clust=1,verbose=FALSE,parll=TRUE)
+{
+ if (verbose)
+ cat(paste("*** Clustering task 2 on ",ncol(medoids)," synchrones\n", sep=""))
+
+ if (ncol(medoids) <= K2)
+ return (medoids)
+
+ # A) Obtain synchrones, that is to say the cumulated power consumptions
+ # for each of the K1 initial groups
+ synchrones = computeSynchrones(medoids, getRefSeries, nb_ref_curves,
+ nb_series_per_chunk, ncores_clust, verbose, parll)
+
+ # B) Compute the WER distances (Wavelets Extended coefficient of deteRmination)
+ distances = computeWerDists(
+ synchrones, nvoice, nbytes, endian, ncores_clust, verbose, parll)
-# Cluster a chunk of series inside one task (~max nb_series_per_chunk)
-computeClusters2 = function(medoids, K2, getRefSeries, nb_series_per_chunk)
+ # C) Apply clustering algorithm 2 on the WER distances matrix
+ if (verbose)
+ cat(paste("*** algoClust2() on ",nrow(distances)," items\n", sep=""))
+ medoids[ ,algoClust2(distances,K2) ]
+}
+
+#' computeSynchrones
+#'
+#' Compute the synchrones curves (sum of clusters elements) from a matrix of medoids,
+#' using euclidian distance.
+#'
+#' @param medoids big.matrix of medoids (curves of same length as initial series)
+#' @param getRefSeries Function to retrieve initial series (e.g. in stage 2 after series
+#' have been replaced by stage-1 medoids)
+#' @param nb_ref_curves How many reference series? (This number is known at this stage)
+#' @inheritParams claws
+#'
+#' @return A big.matrix of size L x K1 where L = length of a serie
+#'
+#' @export
+computeSynchrones = function(medoids, getRefSeries, nb_ref_curves,
+ nb_series_per_chunk, ncores_clust=1,verbose=FALSE,parll=TRUE)
{
- synchrones = computeSynchrones(medoids, getRefSeries, nb_series_per_chunk)
- cluster::pam(computeWerDists(synchrones), K2, diss=TRUE)$medoids
+ # Synchrones computation is embarassingly parallel: compute it by chunks of series
+ computeSynchronesChunk = function(indices)
+ {
+ if (parll)
+ {
+ require("bigmemory", quietly=TRUE)
+ requireNamespace("synchronicity", quietly=TRUE)
+ require("epclust", quietly=TRUE)
+ # The big.matrix objects need to be attached to be usable on the workers
+ synchrones <- bigmemory::attach.big.matrix(synchrones_desc)
+ medoids <- bigmemory::attach.big.matrix(medoids_desc)
+ m <- synchronicity::attach.mutex(m_desc)
+ }
+
+ # Obtain a chunk of reference series
+ ref_series = getRefSeries(indices)
+ nb_series = ncol(ref_series)
+
+ # Get medoids indices for this chunk of series
+ mi = computeMedoidsIndices(medoids@address, ref_series)
+
+ # Update synchrones using mi above
+ for (i in seq_len(nb_series))
+ {
+ if (parll)
+ synchronicity::lock(m) #locking required because several writes at the same time
+ synchrones[, mi[i] ] = synchrones[, mi[i] ] + ref_series[,i]
+ if (parll)
+ synchronicity::unlock(m)
+ }
+ NULL
+ }
+
+ K = ncol(medoids) ; L = nrow(medoids)
+ # Use bigmemory (shared==TRUE by default) + synchronicity to fill synchrones in //
+ synchrones = bigmemory::big.matrix(nrow=L, ncol=K, type="double", init=0.)
+ # NOTE: synchronicity is only for Linux & MacOS; on Windows: run sequentially
+ parll = (parll && requireNamespace("synchronicity",quietly=TRUE)
+ && Sys.info()['sysname'] != "Windows")
+ if (parll)
+ {
+ m <- synchronicity::boost.mutex() #for lock/unlock, see computeSynchronesChunk
+ # mutex and big.matrix objects cannot be passed directly:
+ # they will be accessed from their description
+ m_desc <- synchronicity::describe(m)
+ synchrones_desc = bigmemory::describe(synchrones)
+ medoids_desc = bigmemory::describe(medoids)
+ cl = parallel::makeCluster(ncores_clust)
+ parallel::clusterExport(cl, envir=environment(),
+ varlist=c("synchrones_desc","m_desc","medoids_desc","getRefSeries"))
+ }
+
+ if (verbose)
+ cat(paste("--- Compute ",K," synchrones with ",nb_ref_curves," series\n", sep=""))
+
+ # Balance tasks by splitting the indices set - maybe not so evenly, but
+ # max==TRUE in next call ensures that no set has more than nb_series_per_chunk items.
+ indices_workers = .spreadIndices(seq_len(nb_ref_curves), nb_series_per_chunk, max=TRUE)
+ ignored <-
+ if (parll)
+ parallel::parLapply(cl, indices_workers, computeSynchronesChunk)
+ else
+ lapply(indices_workers, computeSynchronesChunk)
+
+ if (parll)
+ parallel::stopCluster(cl)
+
+ return (synchrones)
}
-# Compute the synchrones curves (sum of clusters elements) from a clustering result
-computeSynchrones = function(medoids, getRefSeries, nb_series_per_chunk)
+#' computeWerDists
+#'
+#' Compute the WER distances between the synchrones curves (in columns), which are
+#' returned (e.g.) by \code{computeSynchrones()}
+#'
+#' @param synchrones A big.matrix of synchrones, in columns. The series have same
+#' length as the series in the initial dataset
+#' @inheritParams claws
+#'
+#' @return A distances matrix of size K1 x K1
+#'
+#' @export
+computeWerDists = function(synchrones, nvoice, nbytes,endian,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
+ n <- ncol(synchrones)
+ L <- nrow(synchrones)
+ noctave = ceiling(log2(L)) #min power of 2 to cover serie range
+
+ # Initialize result as a square big.matrix of size 'number of synchrones'
+ Xwer_dist <- bigmemory::big.matrix(nrow=n, ncol=n, type="double")
+
+ # 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)))
+ }
+
+ cwt_file = ".cwt.bin"
+ # Compute the synchrones[,index] CWT, and store it in the binary file above
+ computeSaveCWT = function(index)
+ {
+ if (parll && !exists(synchrones)) #avoid going here after first call on a worker
+ {
+ require("bigmemory", quietly=TRUE)
+ require("Rwave", quietly=TRUE)
+ require("epclust", quietly=TRUE)
+ synchrones <- bigmemory::attach.big.matrix(synchrones_desc)
+ }
+ ts <- scale(ts(synchrones[,index]), center=TRUE, scale=FALSE)
+ ts_cwt = Rwave::cwt(ts, noctave, nvoice, w0=2*pi, twoD=TRUE, plot=FALSE)
+
+ # Serialization
+ binarize(as.matrix(c(as.double(Re(ts_cwt)),as.double(Im(ts_cwt)))), cwt_file, 1,
+ ",", nbytes, endian)
+ }
+
+ if (parll)
+ {
+ cl = parallel::makeCluster(ncores_clust)
+ synchrones_desc <- bigmemory::describe(synchrones)
+ Xwer_dist_desc <- bigmemory::describe(Xwer_dist)
+ parallel::clusterExport(cl, varlist=c("parll","synchrones_desc","Xwer_dist_desc",
+ "noctave","nvoice","verbose","getCWT"), envir=environment())
+ }
+
+ if (verbose)
+ cat(paste("--- Precompute and serialize synchrones CWT\n", sep=""))
+
+ ignored <-
+ if (parll)
+ parallel::parLapply(cl, 1:n, computeSaveCWT)
+ else
+ lapply(1:n, computeSaveCWT)
+
+ # Function to retrieve a synchrone CWT from (binary) file
+ getSynchroneCWT = function(index, L)
+ {
+ flat_cwt <- getDataInFile(index, cwt_file, nbytes, endian)
+ cwt_length = length(flat_cwt) / 2
+ re_part = as.matrix(flat_cwt[1:cwt_length], nrow=L)
+ im_part = as.matrix(flat_cwt[(cwt_length+1):(2*cwt_length)], nrow=L)
+ re_part + 1i * im_part
+ }
+
+ # Compute distance between columns i and j in synchrones
+ computeDistanceIJ = function(pair)
{
- range = (index-1) + seq_len(nb_series_per_chunk)
- ref_series = getRefSeries(range)
- if (is.null(ref_series))
- break
- #get medoids indices for this chunk of series
- for (i in seq_len(nrow(ref_series)))
+ if (parll)
{
- j = which.min( rowSums( sweep(medoids, 2, series[i,], '-')^2 ) )
- synchrones[j,] = synchrones[j,] + series[i,]
- counts[j] = counts[j] + 1
+ # parallel workers start with an empty environment
+ require("bigmemory", quietly=TRUE)
+ require("epclust", quietly=TRUE)
+ synchrones <- bigmemory::attach.big.matrix(synchrones_desc)
+ Xwer_dist <- bigmemory::attach.big.matrix(Xwer_dist_desc)
}
- index = index + nb_series_per_chunk
+
+ i = pair[1] ; j = pair[2]
+ if (verbose && j==i+1 && !parll)
+ cat(paste(" Distances (",i,",",j,"), (",i,",",j+1,") ...\n", sep=""))
+
+ # Compute CWT of columns i and j in synchrones
+ L = nrow(synchrones)
+ cwt_i <- getSynchroneCWT(i, L)
+ cwt_j <- getSynchroneCWT(j, L)
+
+ # Compute the ratio of integrals formula 5.6 for WER^2
+ # in https://arxiv.org/abs/1101.4744v2 ยง5.3
+ num <- filterMA(Mod(cwt_i * Conj(cwt_j)))
+ WX <- filterMA(Mod(cwt_i * Conj(cwt_i)))
+ WY <- filterMA(Mod(cwt_j * Conj(cwt_j)))
+ wer2 <- sum(colSums(num)^2) / sum(colSums(WX) * colSums(WY))
+
+ Xwer_dist[i,j] <- sqrt(L * ncol(cwt_i) * (1 - wer2))
+ Xwer_dist[j,i] <- Xwer_dist[i,j]
+ Xwer_dist[i,i] <- 0.
}
- #NOTE: odds for some clusters to be empty? (when series already come from stage 2)
- sweep(synchrones, 1, counts, '/')
+
+ if (verbose)
+ cat(paste("--- Compute WER distances\n", sep=""))
+
+ ignored <-
+ if (parll)
+ parallel::parLapply(cl, pairs, computeDistanceIJ)
+ else
+ lapply(pairs, computeDistanceIJ)
+
+ if (parll)
+ parallel::stopCluster(cl)
+
+ unlink(cwt_file)
+
+ Xwer_dist[n,n] = 0.
+ Xwer_dist[,] #~small matrix K1 x K1
}
-# Compute the WER distance between the synchrones curves (in rows)
-computeWerDist = function(curves)
+# Helper function to divide indices into balanced sets
+# If max == TRUE, sets sizes cannot exceed nb_per_set
+.spreadIndices = function(indices, nb_per_set, max=FALSE)
{
- if (!require("Rwave", quietly=TRUE))
- stop("Unable to load Rwave library")
- n <- nrow(curves)
- delta <- ncol(curves)
- #TODO: automatic tune of all these parameters ? (for other users)
- nvoice <- 4
- # noctave = 2^13 = 8192 half hours ~ 180 days ; ~log2(ncol(curves))
- 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
- #condition: ( log2(s0*w0/(2*pi)) - 1 ) * nvoice + 1.5 >= 1
- s0=2
- w0=2*pi
- scaled=FALSE
- 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)
- 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 / max(Mod(sqres))
- })
-
- 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))
+ L = length(indices)
+ nb_workers = floor( L / nb_per_set )
+ rem = L %% nb_per_set
+ if (nb_workers == 0 || (nb_workers==1 && rem==0))
+ {
+ # L <= nb_per_set, simple case
+ indices_workers = list(indices)
+ }
+ else
{
- for (j in (i+1):n)
+ indices_workers = lapply( seq_len(nb_workers), function(i)
+ indices[(nb_per_set*(i-1)+1):(nb_per_set*i)] )
+
+ if (max)
+ {
+ # Sets are not so well balanced, but size is supposed to be critical
+ return ( c( indices_workers, if (rem>0) list((L-rem+1):L) else NULL ) )
+ }
+
+ # Spread the remaining load among the workers
+ rem = L %% nb_per_set
+ 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
}
projects / epclust.git / blobdiff