save state; test clustering not OK, all others OK

[epclust.git] / epclust / R / main.R

#' CLAWS: CLustering with wAvelets and Wer distanceS
#'
#' Cluster electricity power curves (or any series of similar nature) by applying a
#' two stage procedure in parallel (see details).
#' Input series must be sampled on the same time grid, no missing values.
#'
#' @details Summary of the function execution flow:
#'   \enumerate{
#'     \item Compute and serialize all contributions, obtained through discrete wavelet
#'       decomposition (see Antoniadis & al. [2013])
#'     \item Divide series into \code{ntasks} groups to process in parallel. In each task:
#'     \enumerate{
#'       \item iterate the first clustering algorithm on its aggregated outputs,
#'         on inputs of size \code{nb_items_clust1}
#'       \item optionally, if WER=="mix":
#'         a) compute the K1 synchrones curves,
#'         b) compute WER distances (K1xK1 matrix) between synchrones and
#'         c) apply the second clustering algorithm
#'     }
#'     \item Launch a final task on the aggregated outputs of all previous tasks:
#'       in the case WER=="end" this task takes indices in input, otherwise
#'       (medoid) curves
#'   }
#'   The main argument -- \code{getSeries} -- has a quite misleading name, since it can be
#'   either a [big.]matrix, a CSV file, a connection or a user function to retrieve
#'   series; the name was chosen because all types of arguments are converted to a function.
#'   When \code{getSeries} is given as a function, it must take a single argument,
#'   'indices', integer vector equal to the indices of the curves to retrieve;
#'   see SQLite example. The nature and role of other arguments should be clear
#'
#' @param getSeries Access to the (time-)series, which can be of one of the three
#'   following types:
#'   \itemize{
#'     \item [big.]matrix: each column contains the (time-ordered) values of one time-serie
#'     \item connection: any R connection object providing lines as described above
#'     \item character: name of a CSV file containing series in rows (no header)
#'     \item function: a custom way to retrieve the curves; it has only one argument:
#'       the indices of the series to be retrieved. See SQLite example
#'   }
#' @param K1 Number of clusters to be found after stage 1 (K1 << N [number of series])
#' @param K2 Number of clusters to be found after stage 2 (K2 << K1)
#' @param nb_series_per_chunk (Maximum) number of series to retrieve in one batch
#' @param algoClust1 Clustering algorithm for stage 1. A function which takes (data, K)
#'   as argument where data is a matrix in columns and K the desired number of clusters,
#'   and outputs K medoids ranks. Default: PAM.
#'   In our method, this function is called on iterated medoids during stage 1
#' @param algoClust2 Clustering algorithm for stage 2. A function which takes (dists, K)
#'   as argument where dists is a matrix of distances and K the desired number of clusters,
#'   and outputs K clusters representatives (curves). Default: PAM.
#'   In our method, this function is called on a matrix of K1 x K1 (WER) distances computed
#    between synchrones
#' @param nb_items_clust1 (~Maximum) number of items in input of the clustering algorithm
#'   for stage 1. At worst, a clustering algorithm might be called with ~2*nb_items_clust1
#'   items; but this could only happen at the last few iterations.
#' @param wav_filt Wavelet transform filter; see ?wavelets::wt.filter
#' @param contrib_type Type of contribution: "relative", "logit" or "absolute" (any prefix)
#' @param WER "end" to apply stage 2 after stage 1 has fully iterated, or "mix" to apply
#'   stage 2 at the end of each task
#' @param sync_mean TRUE to compute a synchrone as a mean curve, FALSE for a sum
#' @param random TRUE (default) for random chunks repartition
#' @param ntasks Number of tasks (parallel iterations to obtain K1 [if WER=="end"]
#'   or K2 [if WER=="mix"] medoids); default: 1.
#'   Note: ntasks << N (number of series), so that N is "roughly divisible" by ntasks
#' @param ncores_tasks Number of parallel tasks (1 to disable: sequential tasks)
#' @param ncores_clust Number of parallel clusterings in one task (4 should be a minimum)
#' @param sep Separator in CSV input file (if any provided)
#' @param nbytes Number of bytes to serialize a floating-point number; 4 or 8
#' @param endian Endianness for (de)serialization ("little" or "big")
#' @param verbose Level of verbosity (0/FALSE for nothing or 1/TRUE for all; devel stage)
#' @param parll TRUE to fully parallelize; otherwise run sequentially (debug, comparison)
#'
#' @return A matrix of the final K2 medoids curves, in columns
#'
#' @references Clustering functional data using Wavelets [2013];
#'   A. Antoniadis, X. Brossat, J. Cugliari & J.-M. Poggi.
#'   Inter. J. of Wavelets, Multiresolution and Information Procesing,
#'   vol. 11, No 1, pp.1-30. doi:10.1142/S0219691313500033
#'
#' @examples
#' \dontrun{
#' # WER distances computations are too long for CRAN (for now)
#'
#' # Random series around cos(x,2x,3x)/sin(x,2x,3x)
#' x = seq(0,500,0.05)
#' L = length(x) #10001
#' ref_series = matrix( c(cos(x),cos(2*x),cos(3*x),sin(x),sin(2*x),sin(3*x)), ncol=6 )
#' library(wmtsa)
#' series = do.call( cbind, lapply( 1:6, function(i)
#'   do.call(cbind, wmtsa::wavBootstrap(ref_series[i,], n.realization=400)) ) )
#' #dim(series) #c(2400,10001)
#' medoids_ascii = claws(series, K1=60, K2=6, 200, verbose=TRUE)
#'
#' # Same example, from CSV file
#' csv_file = "/tmp/epclust_series.csv"
#' write.table(series, csv_file, sep=",", row.names=FALSE, col.names=FALSE)
#' medoids_csv = claws(csv_file, K1=60, K2=6, 200)
#'
#' # Same example, from binary file
#' bin_file <- "/tmp/epclust_series.bin"
#' nbytes <- 8
#' endian <- "little"
#' binarize(csv_file, bin_file, 500, nbytes, endian)
#' getSeries <- function(indices) getDataInFile(indices, bin_file, nbytes, endian)
#' medoids_bin <- claws(getSeries, K1=60, K2=6, 200)
#' unlink(csv_file)
#' unlink(bin_file)
#'
#' # Same example, from SQLite database
#' library(DBI)
#' series_db <- dbConnect(RSQLite::SQLite(), "file::memory:")
#' # Prepare data.frame in DB-format
#' n <- nrow(series)
#' time_values <- data.frame(
#'   id = rep(1:n,each=L),
#'   time = rep( as.POSIXct(1800*(0:n),"GMT",origin="2001-01-01"), L ),
#'   value = as.double(t(series)) )
#' dbWriteTable(series_db, "times_values", times_values)
#' # Fill associative array, map index to identifier
#' indexToID_inDB <- as.character(
#'   dbGetQuery(series_db, 'SELECT DISTINCT id FROM time_values')[,"id"] )
#' serie_length <- as.integer( dbGetQuery(series_db,
#'   paste("SELECT COUNT * FROM time_values WHERE id == ",indexToID_inDB[1],sep="")) )
#' getSeries <- function(indices) {
#'   request <- "SELECT id,value FROM times_values WHERE id in ("
#'   for (i in indices)
#'     request <- paste(request, indexToID_inDB[i], ",", sep="")
#'   request <- paste(request, ")", sep="")
#'   df_series <- dbGetQuery(series_db, request)
#'   as.matrix(df_series[,"value"], nrow=serie_length)
#' }
#' medoids_db = claws(getSeries, K1=60, K2=6, 200))
#' dbDisconnect(series_db)
#'
#' # All computed medoids should be the same:
#' digest::sha1(medoids_ascii)
#' digest::sha1(medoids_csv)
#' digest::sha1(medoids_bin)
#' digest::sha1(medoids_db)
#' }
#' @export
claws <- function(getSeries, K1, K2, nb_series_per_chunk,
	nb_items_clust1=7*K1,
	algoClust1=function(data,K) cluster::pam(t(data),K,diss=FALSE)$id.med,
	algoClust2=function(dists,K) t( cluster::pam(dists,K,diss=TRUE)$medoids ),
	wav_filt="d8", contrib_type="absolute",
	WER="end",sync_mean=TRUE,
	random=TRUE,
	ntasks=1, ncores_tasks=1, ncores_clust=4,
	sep=",",
	nbytes=4, endian=.Platform$endian,
	verbose=FALSE, parll=TRUE)
{
	# Check/transform arguments
	if (!is.matrix(getSeries) && !bigmemory::is.big.matrix(getSeries)
		&& !is.function(getSeries)
		&& !methods::is(getSeries,"connection") && !is.character(getSeries))
	{
		stop("'getSeries': [big]matrix, function, file or valid connection (no NA)")
	}
	K1 <- .toInteger(K1, function(x) x>=2)
	K2 <- .toInteger(K2, function(x) x>=2)
	nb_series_per_chunk <- .toInteger(nb_series_per_chunk, function(x) x>=1)
	# K1 (number of clusters at step 1) cannot exceed nb_series_per_chunk, because we will need
	# to load K1 series in memory for clustering stage 2.
	if (K1 > nb_series_per_chunk)
		stop("'K1' cannot exceed 'nb_series_per_chunk'")
	nb_items_clust1 <- .toInteger(nb_items_clust1, function(x) x>K1)
	random <- .toLogical(random)
	tryCatch( {ignored <- wavelets::wt.filter(wav_filt)},
		error = function(e) stop("Invalid wavelet filter; see ?wavelets::wt.filter") )
	ctypes = c("relative","absolute","logit")
	contrib_type = ctypes[ pmatch(contrib_type,ctypes) ]
	if (is.na(contrib_type))
		stop("'contrib_type' in {'relative','absolute','logit'}")
	if (WER!="end" && WER!="mix")
		stop("'WER': in {'end','mix'}")
	sync_mean <- .toLogical(sync_mean)
	random <- .toLogical(random)
	ntasks <- .toInteger(ntasks, function(x) x>=1)
	ncores_tasks <- .toInteger(ncores_tasks, function(x) x>=1)
	ncores_clust <- .toInteger(ncores_clust, function(x) x>=1)
	if (!is.character(sep))
		stop("'sep': character")
	nbytes <- .toInteger(nbytes, function(x) x==4 || x==8)
	verbose <- .toLogical(verbose)
	parll <- .toLogical(parll)

	# Since we don't make assumptions on initial data, there is a possibility that even
	# when serialized, contributions or synchrones do not fit in RAM. For example,
	# 30e6 series of length 100,000 would lead to a +4Go contribution matrix. Therefore,
	# it's safer to place these in (binary) files, located in the following folder.
	bin_dir <- ".epclust_bin/"
	dir.create(bin_dir, showWarnings=FALSE, mode="0755")

	# Binarize series if getSeries is not a function; the aim is to always use a function,
	# to uniformize treatments. An equally good alternative would be to use a file-backed
	# bigmemory::big.matrix, but it would break the uniformity.
	if (!is.function(getSeries))
	{
		if (verbose)
			cat("...Serialize time-series\n")
		series_file = paste(bin_dir,"data",sep="") ; unlink(series_file)
		binarize(getSeries, series_file, nb_series_per_chunk, sep, nbytes, endian)
		getSeries = function(inds) getDataInFile(inds, series_file, nbytes, endian)
	}

	# Serialize all computed wavelets contributions into a file
	contribs_file = paste(bin_dir,"contribs",sep="") ; unlink(contribs_file)
	index = 1
	nb_curves = 0
	if (verbose)
		cat("...Compute contributions and serialize them\n")
	nb_curves = binarizeTransform(getSeries,
		function(series) curvesToContribs(series, wf, ctype),
		contribs_file, nb_series_per_chunk, nbytes, endian)
	getContribs = function(indices) getDataInFile(indices, contribs_file, nbytes, endian)

	# A few sanity checks: do not continue if too few data available.
	if (nb_curves < K2)
		stop("Not enough data: less series than final number of clusters")
	nb_series_per_task = round(nb_curves / ntasks)
	if (nb_series_per_task < K2)
		stop("Too many tasks: less series in one task than final number of clusters")

	# Generate a random permutation of 1:N (if random==TRUE); otherwise just use arrival
	# (storage) order.
	indices_all = if (random) sample(nb_curves) else seq_len(nb_curves)
	# Split (all) indices into ntasks groups of ~same size
	indices_tasks = lapply(seq_len(ntasks), function(i) {
		upper_bound = ifelse( i<ntasks, min(nb_series_per_task*i,nb_curves), nb_curves )
		indices_all[((i-1)*nb_series_per_task+1):upper_bound]
	})

	if (parll && ntasks>1)
	{
		# Initialize parallel runs: outfile="" allow to output verbose traces in the console
		# under Linux. All necessary variables are passed to the workers.
		cl = parallel::makeCluster(ncores_tasks, outfile="")
		varlist = c("getSeries","getContribs","K1","K2","algoClust1","algoClust2",
			"nb_series_per_chunk","nb_items_clust1","ncores_clust","sep",
			"nbytes","endian","verbose","parll")
		if (WER=="mix")
			varlist = c(varlist, "medoids_file")
		parallel::clusterExport(cl, varlist, envir = environment())
	}

	# This function achieves one complete clustering task, divided in stage 1 + stage 2.
	# stage 1: n indices  --> clusteringTask1(...) --> K1 medoids
	# stage 2: K1 medoids --> clusteringTask2(...) --> K2 medoids,
	# where n = N / ntasks, N being the total number of curves.
	runTwoStepClustering = function(inds)
	{
		# When running in parallel, the environment is blank: we need to load required
		# packages, and pass useful variables.
		if (parll && ntasks>1)
			require("epclust", quietly=TRUE)
		indices_medoids = clusteringTask1(
			inds, getContribs, K1, algoClust1, nb_series_per_chunk, ncores_clust, verbose, parll)
		if (WER=="mix")
		{
			if (parll && ntasks>1)
				require("bigmemory", quietly=TRUE)
			medoids1 = bigmemory::as.big.matrix( getSeries(indices_medoids) )
			medoids2 = clusteringTask2(medoids1, K2, algoClust2, getSeries, nb_curves,
				nb_series_per_chunk, sync_mean, nbytes, endian, ncores_clust, verbose, parll)
			binarize(medoids2, medoids_file, nb_series_per_chunk, sep, nbytes, endian)
			return (vector("integer",0))
		}
		indices_medoids
	}

	# Synchrones (medoids) need to be stored only if WER=="mix"; indeed in this case, every
	# task output is a set of new (medoids) curves. If WER=="end" however, output is just a
	# set of indices, representing some initial series.
	if (WER=="mix")
		{medoids_file = paste(bin_dir,"medoids",sep="") ; unlink(medoids_file)}

	if (verbose)
	{
		message = paste("...Run ",ntasks," x stage 1", sep="")
		if (WER=="mix")
			message = paste(message," + stage 2", sep="")
		cat(paste(message,"\n", sep=""))
	}

	# As explained above, indices will be assigned to ntasks*K1 medoids indices [if WER=="end"],
	# or nothing (empty vector) if WER=="mix"; in this case, medoids (synchrones) are stored
	# in a file.
	indices <-
		if (parll && ntasks>1)
			unlist( parallel::parLapply(cl, indices_tasks, runTwoStepClustering) )
		else
			unlist( lapply(indices_tasks, runTwoStepClustering) )
	if (parll && ntasks>1)
		parallel::stopCluster(cl)

	# Right before the final stage, two situations are possible:
	#  a. data to be processed now sit in binary format in medoids_file (if WER=="mix")
	#  b. data still is the initial set of curves, referenced by the ntasks*K1 indices
	# So, the function getSeries() will potentially change. However, computeSynchrones()
	# requires a function retrieving the initial series. Thus, the next line saves future
	# conditional instructions.
	getRefSeries = getSeries

	if (WER=="mix")
	{
		indices = seq_len(ntasks*K2)
		# Now series (synchrones) must be retrieved from medoids_file
		getSeries = function(inds) getDataInFile(inds, medoids_file, nbytes, endian)
		# Contributions must be re-computed
		unlink(contribs_file)
		index = 1
		if (verbose)
			cat("...Serialize contributions computed on synchrones\n")
		ignored = binarizeTransform(getSeries,
			function(series) curvesToContribs(series, wf, ctype),
			contribs_file, nb_series_per_chunk, nbytes, endian)
	}

#TODO: check THAT


	# Run step2 on resulting indices or series (from file)
	if (verbose)
		cat("...Run final // stage 1 + stage 2\n")
	indices_medoids = clusteringTask1(indices, getContribs, K1, algoClust1,
		nb_series_per_chunk, ncores_tasks*ncores_clust, verbose, parll)
	medoids1 = bigmemory::as.big.matrix( getSeries(indices_medoids) )
	medoids2 = clusteringTask2(medoids1, K2, algoClust2, getRefSeries, nb_curves,
		nb_series_per_chunk, sync_mean, nbytes, endian, ncores_tasks*ncores_clust, verbose, parll)

	# Cleanup: remove temporary binary files and their folder
	unlink(bin_dir, recursive=TRUE)

	# Return medoids as a standard matrix, since K2 series have to fit in RAM
	# (clustering algorithm 1 takes K1 > K2 of them as input)
	medoids2[,]
}

#' curvesToContribs
#'
#' Compute the discrete wavelet coefficients for each series, and aggregate them in
#' energy contribution across scales as described in https://arxiv.org/abs/1101.4744v2
#'
#' @param series [big.]matrix of series (in columns), of size L x n
#' @inheritParams claws
#'
#' @return A [big.]matrix of size log(L) x n containing contributions in columns
#'
#' @export
curvesToContribs = function(series, wav_filt, contrib_type, coin=FALSE)
{
	L = nrow(series)
	if (coin) browser()
	D = ceiling( log2(L) )
	nb_sample_points = 2^D
	apply(series, 2, function(x) {
		interpolated_curve = spline(1:L, x, n=nb_sample_points)$y
		W = wavelets::dwt(interpolated_curve, filter=wf, D)@W
		nrj = rev( sapply( W, function(v) ( sqrt( sum(v^2) ) ) ) )
		if (contrib_type!="absolute")
			nrj = nrj / sum(nrj)
		if (contrib_type=="logit")
			nrj = - log(1 - nrj)
		nrj
	})
}

# Check integer arguments with functional conditions
.toInteger <- function(x, condition)
{
	errWarn <- function(ignored)
		paste("Cannot convert argument' ",substitute(x),"' to integer", sep="")
	if (!is.integer(x))
		tryCatch({x = as.integer(x)[1]; if (is.na(x)) stop()},
			warning = errWarn, error = errWarn)
	if (!condition(x))
	{
		stop(paste("Argument '",substitute(x),
			"' does not verify condition ",body(condition), sep=""))
	}
	x
}

# Check logical arguments
.toLogical <- function(x)
{
	errWarn <- function(ignored)
		paste("Cannot convert argument' ",substitute(x),"' to logical", sep="")
	if (!is.logical(x))
		tryCatch({x = as.logical(x)[1]; if (is.na(x)) stop()},
			warning = errWarn, error = errWarn)
	x
}