[epclust.git] / epclust / R / clustering.R

#' @name clustering
#' @rdname clustering
#' @aliases clusteringTask computeClusters1 computeClusters2
#'
#' @title Two-stages clustering, withing one task (see \code{claws()})
#'
#' @description \code{clusteringTask()} runs one full task, which consists in iterated stage 1
#'   clustering (on nb_curves / ntasks energy contributions, computed through discrete
#'   wavelets coefficients). \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{clusteringTask()} and \code{computeClusters1()}, the indices of the
#'   computed (K1) medoids. Indices are irrelevant for stage 2 clustering, thus
#'   \code{computeClusters2()} outputs a matrix of medoids
#'   (of size limited by nb_series_per_chunk)
NULL

#' @rdname clustering
#' @export
clusteringTask = function(indices, getContribs, K1, nb_series_per_chunk, ncores_clust)
{

#NOTE: comment out parallel sections for debugging
#propagate verbose arg ?!

#	cl = parallel::makeCluster(ncores_clust)
#	parallel::clusterExport(cl, varlist=c("getContribs","K1"), envir=environment())
	repeat
	{

print(length(indices))

		nb_workers = max( 1, floor( length(indices) / nb_series_per_chunk ) )
		indices_workers = lapply( seq_len(nb_workers), function(i)
			indices[(nb_series_per_chunk*(i-1)+1):(nb_series_per_chunk*i)] )
		# Spread the remaining load among the workers
		rem = length(indices) %% nb_series_per_chunk
		while (rem > 0)
		{
			index = rem%%nb_workers + 1
			indices_workers[[index]] = c(indices_workers[[index]], tail(indices,rem))
			rem = rem - 1
		}
#		indices = unlist( parallel::parLapply( cl, indices_workers, function(inds) {
		indices = unlist( lapply( indices_workers, function(inds) {
#			require("epclust", quietly=TRUE)

print(paste("   ",length(inds))) ## PROBLEME ICI : 21104 ??!

			inds[ computeClusters1(getContribs(inds), K1) ]
		} ) )
		if (length(indices) == K1)
			break
	}
#	parallel::stopCluster(cl)
	indices #medoids
}

#' @rdname clustering
#' @export
computeClusters1 = function(contribs, K1)
	cluster::pam(contribs, K1, diss=FALSE)$id.med

#' @rdname clustering
#' @export
computeClusters2 = function(medoids, K2, getRefSeries, nb_series_per_chunk)
{
	synchrones = computeSynchrones(medoids, getRefSeries, nb_series_per_chunk)
	medoids[ cluster::pam(computeWerDists(synchrones), K2, diss=TRUE)$medoids , ]
}

#' computeSynchrones
#'
#' Compute the synchrones curves (sum of clusters elements) from a matrix of medoids,
#' using L2 distances.
#'
#' @param medoids Matrix of medoids (curves of same legnth as initial series)
#' @param getRefSeries Function to retrieve initial series (e.g. in stage 2 after series
#'   have been replaced by stage-1 medoids)
#' @inheritParams claws
#'
#' @export
computeSynchrones = function(medoids, getRefSeries, nb_series_per_chunk)
{
	K = nrow(medoids)
	synchrones = matrix(0, nrow=K, ncol=ncol(medoids))
	counts = rep(0,K)
	index = 1
	repeat
	{
		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)))
		{
			j = which.min( rowSums( sweep(medoids, 2, ref_series[i,], '-')^2 ) )
			synchrones[j,] = synchrones[j,] + ref_series[i,]
			counts[j] = counts[j] + 1
		}
		index = index + nb_series_per_chunk
	}
	#NOTE: odds for some clusters to be empty? (when series already come from stage 2)
	#      ...maybe; but let's hope resulting K1' be still quite bigger than K2
	synchrones = sweep(synchrones, 1, counts, '/')
	synchrones[ sapply(seq_len(K), function(i) all(!is.nan(synchrones[i,]))) , ]
}

#' computeWerDists
#'
#' Compute the WER distances between the synchrones curves (in rows), which are
#' returned (e.g.) by \code{computeSynchrones()}
#'
#' @param synchrones A matrix of synchrones, in rows. The series have same length as the
#' series in the initial dataset
#'
#' @export
computeWerDists = function(synchrones)
{
	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(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
	#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(synchrones[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))
	{
		for (j in (i+1):n)
		{
			#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]
		}
	}
	diag(Xwer_dist) <- numeric(n)
	Xwer_dist
}
Commit	Line	Data
	1	#' @name clustering
	2	#' @rdname clustering
	3	#' @aliases clusteringTask computeClusters1 computeClusters2
	4	#'
	5	#' @title Two-stages clustering, withing one task (see \code{claws()})
	6	#'
	7	#' @description \code{clusteringTask()} runs one full task, which consists in iterated stage 1
	8	#' clustering (on nb_curves / ntasks energy contributions, computed through discrete
	9	#' wavelets coefficients). \code{computeClusters1()} and \code{computeClusters2()}
	10	#' correspond to the atomic clustering procedures respectively for stage 1 and 2.
	11	#' The former applies the clustering algorithm (PAM) on a contributions matrix, while
	12	#' the latter clusters a chunk of series inside one task (~max nb_series_per_chunk)
	13	#'
	14	#' @param indices Range of series indices to cluster in parallel (initial data)
	15	#' @param getContribs Function to retrieve contributions from initial series indices:
	16	#' \code{getContribs(indices)} outpus a contributions matrix
	17	#' @param contribs matrix of contributions (e.g. output of \code{curvesToContribs()})
	18	#' @inheritParams computeSynchrones
	19	#' @inheritParams claws
	20	#'
	21	#' @return For \code{clusteringTask()} and \code{computeClusters1()}, the indices of the
	22	#' computed (K1) medoids. Indices are irrelevant for stage 2 clustering, thus
	23	#' \code{computeClusters2()} outputs a matrix of medoids
	24	#' (of size limited by nb_series_per_chunk)
	25	NULL
	26
	27	#' @rdname clustering
	28	#' @export
	29	clusteringTask = function(indices, getContribs, K1, nb_series_per_chunk, ncores_clust)
	30	{
	31
	32	#NOTE: comment out parallel sections for debugging
	33	#propagate verbose arg ?!
	34
	35	# cl = parallel::makeCluster(ncores_clust)
	36	# parallel::clusterExport(cl, varlist=c("getContribs","K1"), envir=environment())
	37	repeat
	38	{
	39
	40	print(length(indices))
	41
	42	nb_workers = max( 1, floor( length(indices) / nb_series_per_chunk ) )
	43	indices_workers = lapply( seq_len(nb_workers), function(i)
	44	indices[(nb_series_per_chunk(i-1)+1):(nb_series_per_chunki)] )
	45	# Spread the remaining load among the workers
	46	rem = length(indices) %% nb_series_per_chunk
	47	while (rem > 0)
	48	{
	49	index = rem%%nb_workers + 1
	50	indices_workers[[index]] = c(indices_workers[[index]], tail(indices,rem))
	51	rem = rem - 1
	52	}
	53	# indices = unlist( parallel::parLapply( cl, indices_workers, function(inds) {
	54	indices = unlist( lapply( indices_workers, function(inds) {
	55	# require("epclust", quietly=TRUE)
	56
	57	print(paste(" ",length(inds))) ## PROBLEME ICI : 21104 ??!
	58
	59	inds[ computeClusters1(getContribs(inds), K1) ]
	60	} ) )
	61	if (length(indices) == K1)
	62	break
	63	}
	64	# parallel::stopCluster(cl)
	65	indices #medoids
	66	}
	67
	68	#' @rdname clustering
	69	#' @export
	70	computeClusters1 = function(contribs, K1)
	71	cluster::pam(contribs, K1, diss=FALSE)$id.med
	72
	73	#' @rdname clustering
	74	#' @export
	75	computeClusters2 = function(medoids, K2, getRefSeries, nb_series_per_chunk)
	76	{
	77	synchrones = computeSynchrones(medoids, getRefSeries, nb_series_per_chunk)
	78	medoids[ cluster::pam(computeWerDists(synchrones), K2, diss=TRUE)$medoids , ]
	79	}
	80
	81	#' computeSynchrones
	82	#'
	83	#' Compute the synchrones curves (sum of clusters elements) from a matrix of medoids,
	84	#' using L2 distances.
	85	#'
	86	#' @param medoids Matrix of medoids (curves of same legnth as initial series)
	87	#' @param getRefSeries Function to retrieve initial series (e.g. in stage 2 after series
	88	#' have been replaced by stage-1 medoids)
	89	#' @inheritParams claws
	90	#'
	91	#' @export
	92	computeSynchrones = function(medoids, getRefSeries, nb_series_per_chunk)
	93	{
	94	K = nrow(medoids)
	95	synchrones = matrix(0, nrow=K, ncol=ncol(medoids))
	96	counts = rep(0,K)
	97	index = 1
	98	repeat
	99	{
	100	range = (index-1) + seq_len(nb_series_per_chunk)
	101	ref_series = getRefSeries(range)
	102	if (is.null(ref_series))
	103	break
	104	#get medoids indices for this chunk of series
	105	for (i in seq_len(nrow(ref_series)))
	106	{
	107	j = which.min( rowSums( sweep(medoids, 2, ref_series[i,], '-')^2 ) )
	108	synchrones[j,] = synchrones[j,] + ref_series[i,]
	109	counts[j] = counts[j] + 1
	110	}
	111	index = index + nb_series_per_chunk
	112	}
	113	#NOTE: odds for some clusters to be empty? (when series already come from stage 2)
	114	# ...maybe; but let's hope resulting K1' be still quite bigger than K2
	115	synchrones = sweep(synchrones, 1, counts, '/')
	116	synchrones[ sapply(seq_len(K), function(i) all(!is.nan(synchrones[i,]))) , ]
	117	}
	118
	119	#' computeWerDists
	120	#'
	121	#' Compute the WER distances between the synchrones curves (in rows), which are
	122	#' returned (e.g.) by \code{computeSynchrones()}
	123	#'
	124	#' @param synchrones A matrix of synchrones, in rows. The series have same length as the
	125	#' series in the initial dataset
	126	#'
	127	#' @export
	128	computeWerDists = function(synchrones)
	129	{
	130	n <- nrow(synchrones)
	131	delta <- ncol(synchrones)
	132	#TODO: automatic tune of all these parameters ? (for other users)
	133	nvoice <- 4
	134	# noctave = 2^13 = 8192 half hours ~ 180 days ; ~log2(ncol(synchrones))
	135	noctave = 13
	136	# 4 here represent 2^5 = 32 half-hours ~ 1 day
	137	#NOTE: default scalevector == 2^(0:(noctave * nvoice) / nvoice) * s0 (?)
	138	scalevector <- 2^(4:(noctave * nvoice) / nvoice) * 2
	139	#condition: ( log2(s0w0/(2pi)) - 1 ) * nvoice + 1.5 >= 1
	140	s0=2
	141	w0=2*pi
	142	scaled=FALSE
	143	s0log = as.integer( (log2( s0w0/(2pi) ) - 1) * nvoice + 1.5 )
	144	totnoct = noctave + as.integer(s0log/nvoice) + 1
	145
	146	# (normalized) observations node with CWT
	147	Xcwt4 <- lapply(seq_len(n), function(i) {
	148	ts <- scale(ts(synchrones[i,]), center=TRUE, scale=scaled)
	149	totts.cwt = Rwave::cwt(ts,totnoct,nvoice,w0,plot=0)
	150	ts.cwt = totts.cwt[,s0log:(s0log+noctave*nvoice)]
	151	#Normalization
	152	sqs <- sqrt(2^(0:(noctavenvoice)/nvoice)s0)
	153	sqres <- sweep(ts.cwt,MARGIN=2,sqs,'*')
	154	sqres / max(Mod(sqres))
	155	})
	156
	157	Xwer_dist <- matrix(0., n, n)
	158	fcoefs = rep(1/3, 3) #moving average on 3 values (TODO: very slow! correct?!)
	159	for (i in 1:(n-1))
	160	{
	161	for (j in (i+1):n)
	162	{
	163	#TODO: later, compute CWT here (because not enough storage space for 200k series)
	164	# 'circular=TRUE' is wrong, should just take values on the sides; to rewrite in C
	165	num <- filter(Mod(Xcwt4[[i]] * Conj(Xcwt4[[j]])), fcoefs, circular=TRUE)
	166	WX <- filter(Mod(Xcwt4[[i]] * Conj(Xcwt4[[i]])), fcoefs, circular=TRUE)
	167	WY <- filter(Mod(Xcwt4[[j]] * Conj(Xcwt4[[j]])), fcoefs, circular=TRUE)
	168	wer2 <- sum(colSums(num)^2) / sum( sum(colSums(WX) * colSums(WY)) )
	169	Xwer_dist[i,j] <- sqrt(delta * ncol(Xcwt4[[1]]) * (1 - wer2))
	170	Xwer_dist[j,i] <- Xwer_dist[i,j]
	171	}
	172	}
	173	diag(Xwer_dist) <- numeric(n)
	174	Xwer_dist
	175	}