option to run sequentially. various fixes. R CMD check OK
authorBenjamin Auder <benjamin.auder@somewhere>
Wed, 8 Mar 2017 00:52:24 +0000 (01:52 +0100)
committerBenjamin Auder <benjamin.auder@somewhere>
Wed, 8 Mar 2017 00:52:24 +0000 (01:52 +0100)
TODO
epclust/DESCRIPTION
epclust/R/a_NAMESPACE.R
epclust/R/clustering.R
epclust/R/de_serialize.R
epclust/R/main.R
epclust/tests/testthat/test.clustering.R
epclust/tests/testthat/test.de_serialize.R

diff --git a/TODO b/TODO
index 275c10d..1788ce6 100644 (file)
--- a/TODO
+++ b/TODO
@@ -38,3 +38,10 @@ utiliser Rcpp ?
 #      x <- x / sqrt(rowSums(x^2))
 #      x %*% t(x)
 #}
+
+#TODO: soften condition clustering.R line 37 ?
+#regarder mapply et mcmapply pour le // (pas OK pour Windows ou GUI... mais ?)
+#TODO: map-reduce more appropriate R/clustering.R ligne 88
+#Alternative: use bigmemory to share series when CSV or matrix(...)
+
+#' @importFrom synchronicity boost.mutex lock unlock
index 66d1712..304fdff 100644 (file)
@@ -18,8 +18,10 @@ Imports:
     parallel,
     cluster,
     wavelets,
+               bigmemory,
                Rwave
 Suggests:
+    synchronicity,
     devtools,
     testthat,
                MASS,
index 3453108..89aa453 100644 (file)
@@ -9,5 +9,5 @@
 #' @importFrom stats filter spline
 #' @importFrom utils tail
 #' @importFrom methods is
+#' @importFrom bigmemory as.big.matrix is.big.matrix
 NULL
-
index 6408370..74d009e 100644 (file)
@@ -1,15 +1,16 @@
 #' @name clustering
 #' @rdname clustering
-#' @aliases clusteringTask computeClusters1 computeClusters2
+#' @aliases clusteringTask1 computeClusters1 computeClusters2
 #'
-#' @title Two-stages clustering, withing one task (see \code{claws()})
+#' @title Two-stage 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)
+#' @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{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:
@@ -18,7 +19,7 @@
 #' @inheritParams computeSynchrones
 #' @inheritParams claws
 #'
-#' @return For \code{clusteringTask()} and \code{computeClusters1()}, the indices of the
+#' @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 matrix of medoids
 #'   (of size limited by nb_series_per_chunk)
@@ -26,42 +27,36 @@ NULL
 
 #' @rdname clustering
 #' @export
-clusteringTask = function(indices, getContribs, K1, nb_series_per_chunk, ncores_clust)
+clusteringTask1 = function(
+       indices, getContribs, K1, nb_series_per_chunk, ncores_clust=1, verbose=FALSE, parll=TRUE)
 {
+       if (verbose)
+               cat(paste("*** Clustering task on ",length(indices)," lines\n", sep=""))
 
-#NOTE: comment out parallel sections for debugging
-#propagate verbose arg ?!
+       wrapComputeClusters1 = function(inds) {
+               if (parll)
+                       require("epclust", quietly=TRUE)
+               if (verbose)
+                       cat(paste("   computeClusters1() on ",length(inds)," lines\n", sep=""))
+               inds[ computeClusters1(getContribs(inds), K1) ]
+       }
 
-#      cl = parallel::makeCluster(ncores_clust)
-#      parallel::clusterExport(cl, varlist=c("getContribs","K1"), envir=environment())
-       repeat
+       if (parll)
        {
-
-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
+               cl = parallel::makeCluster(ncores_clust)
+               parallel::clusterExport(cl, varlist=c("getContribs","K1","verbose"), envir=environment())
        }
-#      parallel::stopCluster(cl)
+       while (length(indices) > K1)
+       {
+               indices_workers = .spreadIndices(indices, nb_series_per_chunk)
+               if (parll)
+                       indices = unlist( parallel::parLapply(cl, indices_workers, wrapComputeClusters1) )
+               else
+                       indices = unlist( lapply(indices_workers, wrapComputeClusters1) )
+       }
+       if (parll)
+               parallel::stopCluster(cl)
+
        indices #medoids
 }
 
@@ -72,10 +67,13 @@ computeClusters1 = function(contribs, K1)
 
 #' @rdname clustering
 #' @export
-computeClusters2 = function(medoids, K2, getRefSeries, nb_series_per_chunk)
+computeClusters2 = function(medoids, K2,
+       getRefSeries, nb_ref_curves, nb_series_per_chunk, ncores_clust=1,verbose=FALSE,parll=TRUE)
 {
-       synchrones = computeSynchrones(medoids, getRefSeries, nb_series_per_chunk)
-       medoids[ cluster::pam(computeWerDists(synchrones), K2, diss=TRUE)$medoids , ]
+       synchrones = computeSynchrones(medoids,
+               getRefSeries, nb_ref_curves, nb_series_per_chunk, ncores_clust, verbose, parll)
+       distances = computeWerDists(synchrones, ncores_clust, verbose, parll)
+       medoids[ cluster::pam(distances, K2, diss=TRUE)$medoids , ]
 }
 
 #' computeSynchrones
@@ -86,34 +84,67 @@ computeClusters2 = function(medoids, K2, getRefSeries, nb_series_per_chunk)
 #' @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)
+#' @param nb_ref_curves How many reference series? (This number is known at this stage)
 #' @inheritParams claws
 #'
 #' @export
-computeSynchrones = function(medoids, getRefSeries, nb_series_per_chunk)
+computeSynchrones = function(medoids, getRefSeries,
+       nb_ref_curves, nb_series_per_chunk, ncores_clust=1,verbose=FALSE,parll=TRUE)
 {
-       K = nrow(medoids)
-       synchrones = matrix(0, nrow=K, ncol=ncol(medoids))
-       counts = rep(0,K)
-       index = 1
-       repeat
+       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)
                #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 ) )
+                       if (parll)
+                               synchronicity::lock(m)
                        synchrones[j,] = synchrones[j,] + ref_series[i,]
-                       counts[j] = counts[j] + 1
+                       counts[j,1] = counts[j,1] + 1
+                       if (parll)
+                               synchronicity::unlock(m)
+               }
+       }
+
+       K = nrow(medoids)
+       # Use bigmemory (shared==TRUE by default) + synchronicity to fill synchrones in //
+       synchrones = bigmemory::big.matrix(nrow=K,ncol=ncol(medoids),type="double",init=0.)
+       counts = bigmemory::big.matrix(nrow=K,ncol=1,type="double",init=0)
+       # Fork (// run) only on Linux & MacOS; on Windows: run sequentially
+       parll = (requireNamespace("synchronicity",quietly=TRUE)
+               && parll && Sys.info()['sysname'] != "Windows")
+       if (parll)
+               m <- synchronicity::boost.mutex()
+
+       indices_workers = .spreadIndices(seq_len(nb_ref_curves), nb_series_per_chunk)
+       for (inds in indices_workers)
+       {
+               if (verbose)
+               {
+                       cat(paste("--- Compute synchrones for indices range ",
+                               min(inds)," -> ",max(inds),"\n", sep=""))
                }
-               index = index + nb_series_per_chunk
+               if (parll)
+                       ignored <- parallel::mcparallel(computeSynchronesChunk(inds))
+               else
+                       computeSynchronesChunk(inds)
+       }
+       if (parll)
+               parallel::mccollect()
+
+       mat_syncs = matrix(nrow=K, ncol=ncol(medoids))
+       vec_count = rep(NA, K)
+       #TODO: can we avoid this loop?
+       for (i in seq_len(K))
+       {
+               mat_syncs[i,] = synchrones[i,]
+               vec_count[i] = counts[i,1]
        }
        #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,]))) , ]
+       mat_syncs = sweep(mat_syncs, 1, vec_count, '/')
+       mat_syncs[ sapply(seq_len(K), function(i) all(!is.nan(mat_syncs[i,]))) , ]
 }
 
 #' computeWerDists
@@ -122,10 +153,11 @@ computeSynchrones = function(medoids, getRefSeries, nb_series_per_chunk)
 #' 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
+#'   series in the initial dataset
+#' @inheritParams claws
 #'
 #' @export
-computeWerDists = function(synchrones)
+computeWerDists = function(synchrones, ncores_clust=1,verbose=FALSE,parll=TRUE)
 {
        n <- nrow(synchrones)
        delta <- ncol(synchrones)
@@ -143,8 +175,10 @@ computeWerDists = function(synchrones)
        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) {
+       computeCWT = function(i)
+       {
+               if (verbose)
+                       cat(paste("+++ Compute Rwave::cwt() on serie ",i,"\n", sep=""))
                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)]
@@ -152,24 +186,96 @@ computeWerDists = function(synchrones)
                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)
+       if (parll)
+       {
+               cl = parallel::makeCluster(ncores_clust)
+               parallel::clusterExport(cl, varlist=c("getContribs","K1","verbose"), envir=environment())
+       }
+
+       # (normalized) observations node with CWT
+       Xcwt4 <-
+               if (parll)
+                       parallel::parLapply(cl, seq_len(n), computeCWT)
+               else
+                       lapply(seq_len(n), computeCWT)
+
+       if (parll)
+               parallel::stopCluster(cl)
+
+       Xwer_dist <- bigmemory::big.matrix(nrow=n, ncol=n, type="double")
        fcoefs = rep(1/3, 3) #moving average on 3 values (TODO: very slow! correct?!)
-       for (i in 1:(n-1))
+       if (verbose)
+               cat("*** Compute WER distances from CWT\n")
+
+       computeDistancesLineI = function(i)
        {
+               if (verbose)
+                       cat(paste("   Line ",i,"\n", sep=""))
                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
+                       #TODO: '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)) )
+                       if (parll)
+                               synchronicity::lock(m)
                        Xwer_dist[i,j] <- sqrt(delta * ncol(Xcwt4[[1]]) * (1 - wer2))
                        Xwer_dist[j,i] <- Xwer_dist[i,j]
+                       if (parll)
+                               synchronicity::unlock(m)
+               }
+               Xwer_dist[i,i] = 0.
+       }
+
+       parll = (requireNamespace("synchronicity",quietly=TRUE)
+               && parll && Sys.info()['sysname'] != "Windows")
+       if (parll)
+               m <- synchronicity::boost.mutex()
+
+       for (i in 1:(n-1))
+       {
+               if (parll)
+                       ignored <- parallel::mcparallel(computeDistancesLineI(i))
+               else
+                       computeDistancesLineI(i)
+       }
+       Xwer_dist[n,n] = 0.
+
+       if (parll)
+               parallel::mccollect()
+
+       mat_dists = matrix(nrow=n, ncol=n)
+       #TODO: avoid this loop?
+       for (i in 1:n)
+               mat_dists[i,] = Xwer_dist[i,]
+       mat_dists
+}
+
+# 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)
+       {
+               # 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)
+               {
+                       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
 }
index 8dde258..b6684d2 100644 (file)
@@ -1,12 +1,15 @@
 #' @name de_serialize
 #' @rdname de_serialize
-#' @aliases binarize getDataInFile
+#' @aliases binarize binarizeTransform getDataInFile
 #'
-#' @title (De)Serialization of a matrix
+#' @title (De)Serialization of a [big]matrix or data stream
 #'
 #' @description \code{binarize()} serializes a matrix or CSV file with minimal overhead,
 #'   into a binary file. \code{getDataInFile()} achieves the inverse task: she retrieves
-#'   (ASCII) data rows from indices in the binary file
+#'   (ASCII) data rows from indices in the binary file. Finally,
+#'   \code{binarizeTransform()} serialize transformations of all data chunks; to use it,
+#'   a data-retrieval function must be provided, thus \code{binarize} will most likely be
+#'   used first (and then a function defined to seek in generated binary file)
 #'
 #' @param data_ascii Either a matrix or CSV file, with items in rows
 #' @param indices Indices of the lines to retrieve
 #'   or intput (\code{getDataInFile})
 #' @param nb_per_chunk Number of lines to process in one batch
 #' @inheritParams claws
+#' @param getData Function to retrieve data chunks
+#' @param transform Transformation function to apply on data chunks
 #'
 #' @return For \code{getDataInFile()}, the matrix with rows corresponding to the
-#'   requested indices
+#'   requested indices. \code{binarizeTransform} returns the number of processed lines.
+#'   \code{binarize} is designed to serialize in several calls, thus returns nothing.
 NULL
 
 #' @rdname de_serialize
@@ -28,6 +34,7 @@ binarize = function(data_ascii, data_bin_file, nb_per_chunk,
                data_ascii = file(data_ascii, open="r")
        else if (methods::is(data_ascii,"connection") && !isOpen(data_ascii))
                open(data_ascii)
+       is_matrix = !methods::is(data_ascii,"connection")
 
        first_write = (!file.exists(data_bin_file) || file.info(data_bin_file)$size == 0)
        data_bin = file(data_bin_file, open=ifelse(first_write,"wb","ab"))
@@ -37,9 +44,9 @@ binarize = function(data_ascii, data_bin_file, nb_per_chunk,
        {
                #number of items always on 8 bytes
                writeBin(0L, data_bin, size=8, endian=endian)
-               if (is.matrix(data_ascii))
+               if ( is_matrix )
                        data_length = ncol(data_ascii)
-               else #if (is(data, "connection"))
+               else #connection
                {
                        data_line = scan(data_ascii, double(), sep=sep, nlines=1, quiet=TRUE)
                        writeBin(data_line, data_bin, size=nbytes, endian=endian)
@@ -47,18 +54,17 @@ binarize = function(data_ascii, data_bin_file, nb_per_chunk,
                }
        }
 
-       if (is.matrix(data_ascii))
+       if (is_matrix)
                index = 1
        repeat
        {
-               if (is.matrix(data_ascii))
+               if ( is_matrix )
                {
-                       range = index:min(nrow(data_ascii),index+nb_per_chunk)
                        data_chunk =
-                               if (range[1] <= nrow(data_ascii))
-                                       as.double(t(data_ascii[range,]))
+                               if (index <= nrow(data_ascii))
+                                       as.double(t(data_ascii[index:min(nrow(data_ascii),index+nb_per_chunk-1),]))
                                else
-                                       integer(0)
+                                       double(0)
                        index = index + nb_per_chunk
                }
                else
@@ -70,16 +76,36 @@ binarize = function(data_ascii, data_bin_file, nb_per_chunk,
 
        if (first_write)
        {
-               #ecrire file_size-1 / (nbytes*nbWritten) en 0 dans bin_data ! ignored == file_size
+               # Write data_length, = (file_size-1) / (nbytes*nbWritten) at offset 0 in data_bin
                ignored = seek(data_bin, 0)
                writeBin(data_length, data_bin, size=8, endian=endian)
        }
        close(data_bin)
 
-       if (methods::is(data_ascii,"connection"))
+       if ( ! is_matrix )
                close(data_ascii)
 }
 
+#' @rdname de_serialize
+#' @export
+binarizeTransform = function(getData, transform, data_bin_file, nb_per_chunk,
+       nbytes=4, endian=.Platform$endian)
+{
+       nb_items = 0
+       index = 1
+       repeat
+       {
+               data_chunk = getData((index-1)+seq_len(nb_per_chunk))
+               if (is.null(data_chunk))
+                       break
+               transformed_chunk = transform(data_chunk)
+               binarize(transformed_chunk, data_bin_file, nb_per_chunk, ",", nbytes, endian)
+               index = index + nb_per_chunk
+               nb_items = nb_items + nrow(data_chunk)
+       }
+       nb_items
+}
+
 #' @rdname de_serialize
 #' @export
 getDataInFile = function(indices, data_bin_file, nbytes=4, endian=.Platform$endian)
index a982f4c..9064dfa 100644 (file)
@@ -30,6 +30,7 @@
 #' @param nbytes Number of bytes to serialize a floating-point number; 4 or 8
 #' @param endian Endianness to use for (de)serialization. Use "little" or "big" for portability
 #' @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 medoids curves (K2) in rows
 #'
@@ -104,13 +105,14 @@ claws = function(getSeries, K1, K2,
        nb_series_per_chunk=50*K1, min_series_per_chunk=5*K1, #chunk size
        sep=",", #ASCII input separator
        nbytes=4, endian=.Platform$endian, #serialization (write,read)
-       verbose=FALSE)
+       verbose=FALSE, parll=TRUE)
 {
        # Check/transform arguments
-       if (!is.matrix(getSeries) && !is.function(getSeries) &&
-               !methods::is(getSeries, "connection" && !is.character(getSeries)))
+       if (!is.matrix(getSeries) && !bigmemory::is.big.matrix(getSeries)
+               && !is.function(getSeries)
+               && !methods::is(getSeries,"connection") && !is.character(getSeries))
        {
-               stop("'getSeries': matrix, function, file or valid connection (no NA)")
+               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)
@@ -148,16 +150,9 @@ claws = function(getSeries, K1, K2,
        nb_curves = 0
        if (verbose)
                cat("...Compute contributions and serialize them\n")
-       repeat
-       {
-               series = getSeries((index-1)+seq_len(nb_series_per_chunk))
-               if (is.null(series))
-                       break
-               contribs_chunk = curvesToContribs(series, wf, ctype)
-               binarize(contribs_chunk, contribs_file, nb_series_per_chunk, sep, nbytes, endian)
-               index = index + nb_series_per_chunk
-               nb_curves = nb_curves + nrow(contribs_chunk)
-       }
+       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)
 
        if (nb_curves < min_series_per_chunk)
@@ -174,28 +169,37 @@ claws = function(getSeries, K1, K2,
        })
        if (verbose)
                cat(paste("...Run ",ntasks," x stage 1 in parallel\n",sep=""))
-#      cl = parallel::makeCluster(ncores_tasks)
-#      parallel::clusterExport(cl, varlist=c("getSeries","getContribs","K1","K2",
-#              "nb_series_per_chunk","ncores_clust","synchrones_file","sep","nbytes","endian"),
-#              envir = environment())
-       # 1000*K1 indices [if WER=="end"], or empty vector [if WER=="mix"] --> series on file
-#      indices = unlist( parallel::parLapply(cl, indices_tasks, function(inds) {
-       indices = unlist( lapply(indices_tasks, function(inds) {
-#              require("epclust", quietly=TRUE)
-
-               browser() #TODO: CONTINUE DEBUG HERE
+       if (parll)
+       {
+               cl = parallel::makeCluster(ncores_tasks)
+               parallel::clusterExport(cl, varlist=c("getSeries","getContribs","K1","K2","verbose","parll",
+                       "nb_series_per_chunk","ncores_clust","synchrones_file","sep","nbytes","endian"),
+                       envir = environment())
+       }
 
-               indices_medoids = clusteringTask(inds,getContribs,K1,nb_series_per_chunk,ncores_clust)
+       runTwoStepClustering = function(inds)
+       {
+               if (parll)
+                       require("epclust", quietly=TRUE)
+               indices_medoids = clusteringTask1(
+                       inds, getContribs, K1, nb_series_per_chunk, ncores_clust, verbose, parll)
                if (WER=="mix")
                {
-                       medoids2 = computeClusters2(
-                               getSeries(indices_medoids), K2, getSeries, nb_series_per_chunk)
+                       medoids2 = computeClusters2(getSeries(indices_medoids),
+                               K2, getSeries, nb_curves, nb_series_per_chunk, ncores_clust, verbose, parll)
                        binarize(medoids2, synchrones_file, nb_series_per_chunk, sep, nbytes, endian)
                        return (vector("integer",0))
                }
                indices_medoids
-       }) )
-#      parallel::stopCluster(cl)
+       }
+
+       # 1000*K1 indices [if WER=="end"], or empty vector [if WER=="mix"] --> series on file
+       if (parll)
+               indices = unlist( parallel::parLapply(cl, indices_tasks, runTwoStepClustering) )
+       else
+               indices = unlist( lapply(indices_tasks, runTwoStepClustering) )
+       if (parll)
+               parallel::stopCluster(cl)
 
        getRefSeries = getSeries
        synchrones_file = paste(bin_dir,"synchrones",sep="") ; unlink(synchrones_file)
@@ -209,23 +213,18 @@ claws = function(getSeries, K1, K2,
                index = 1
                if (verbose)
                        cat("...Serialize contributions computed on synchrones\n")
-               repeat
-               {
-                       series = getSeries((index-1)+seq_len(nb_series_per_chunk))
-                       if (is.null(series))
-                               break
-                       contribs_chunk = curvesToContribs(series, wf, ctype)
-                       binarize(contribs_chunk, contribs_file, nb_series_per_chunk, sep, nbytes, endian)
-                       index = index + nb_series_per_chunk
-               }
+               ignored = binarizeTransform(getSeries,
+                       function(series) curvesToContribs(series, wf, ctype),
+                       contribs_file, nb_series_per_chunk, nbytes, endian)
        }
 
        # Run step2 on resulting indices or series (from file)
        if (verbose)
                cat("...Run final // stage 1 + stage 2\n")
-       indices_medoids = clusteringTask(
-               indices, getContribs, K1, nb_series_per_chunk, ncores_tasks*ncores_clust)
-       medoids = computeClusters2(getSeries(indices_medoids),K2,getRefSeries,nb_series_per_chunk)
+       indices_medoids = clusteringTask1(
+               indices, getContribs, K1, nb_series_per_chunk, ncores_tasks*ncores_clust, verbose)
+       medoids = computeClusters2(getSeries(indices_medoids),
+               K2, getRefSeries, nb_curves, nb_series_per_chunk, ncores_tasks*ncores_clust, verbose)
 
        # Cleanup
        unlink(bin_dir, recursive=TRUE)
@@ -259,7 +258,7 @@ curvesToContribs = function(series, wf, ctype)
        }) )
 }
 
-# Helper for main function: check integer arguments with functiional conditions
+# Check integer arguments with functional conditions
 .toInteger <- function(x, condition)
 {
        if (!is.integer(x))
index 9333876..49afe60 100644 (file)
@@ -63,10 +63,11 @@ test_that("computeSynchrones behave as expected",
        for (i in seq_len(n))
                series[i,] = s[[I(i,K)]] + rnorm(L,sd=0.01)
        getRefSeries = function(indices) {
-               indices = indices[indices < n]
+               indices = indices[indices <= n]
                if (length(indices)>0) series[indices,] else NULL
        }
-       synchrones = computeSynchrones(rbind(s1,s2,s3), getRefSeries, 100)
+       synchrones = computeSynchrones(rbind(s1,s2,s3), getRefSeries, n, 100,
+               verbose=TRUE, parll=FALSE)
 
        expect_equal(dim(synchrones), c(K,L))
        for (i in 1:K)
@@ -95,23 +96,23 @@ test_that("computeClusters2 behave as expected",
        for (i in seq_len(n))
                series[i,] = s[[I(i,K1)]] + rnorm(L,sd=0.01)
        getRefSeries = function(indices) {
-               indices = indices[indices < n]
+               indices = indices[indices <= n]
                if (length(indices)>0) series[indices,] else NULL
        }
        # Artificially simulate 60 medoids - perfect situation, all equal to one of the refs
        medoids_K1 = do.call(rbind, lapply( 1:K1, function(i) s[[I(i,K1)]] ) )
-       medoids_K2 = computeClusters2(medoids_K1, K2, getRefSeries, 75)
+       medoids_K2 = computeClusters2(medoids_K1, K2, getRefSeries, n, 75,
+               verbose=TRUE, parll=FALSE)
 
        expect_equal(dim(medoids_K2), c(K2,L))
        # Not easy to evaluate result: at least we expect it to be better than random selection of
        # medoids within 1...K1 (among references)
-       
        distorGood = computeDistortion(series, medoids_K2)
        for (i in 1:3)
                expect_lte( distorGood, computeDistortion(series,medoids_K1[sample(1:K1, K2),]) )
 })
 
-test_that("clusteringTask + computeClusters2 behave as expected",
+test_that("clusteringTask1 + computeClusters2 behave as expected",
 {
        n = 900
        x = seq(0,9.5,0.1)
@@ -129,8 +130,10 @@ test_that("clusteringTask + computeClusters2 behave as expected",
        wf = "haar"
        ctype = "absolute"
        getContribs = function(indices) curvesToContribs(series[indices,],wf,ctype)
-       medoids_K1 = getSeries( clusteringTask(1:n, getContribs, K1, 75, 4) )
-       medoids_K2 = computeClusters2(medoids_K1, K2, getSeries, 120)
+       medoids_K1 = getSeries( clusteringTask1(1:n, getContribs, K1, 75,
+               verbose=TRUE, parll=FALSE) )
+       medoids_K2 = computeClusters2(medoids_K1, K2, getSeries, n, 120,
+               verbose=TRUE, parll=FALSE)
 
        expect_equal(dim(medoids_K1), c(K1,L))
        expect_equal(dim(medoids_K2), c(K2,L))
index a2fae5e..8403e6d 100644 (file)
@@ -32,6 +32,33 @@ test_that("serialization + getDataInFile retrieve original data / from matrix",
        unlink(data_bin_file)
 })
 
+test_that("serialization + transform + getDataInFile retrieve original transforms",
+{
+       data_bin_file = "/tmp/epclust_test_t.bin"
+       unlink(data_bin_file)
+
+       #dataset 200 lignes / 30 columns
+       data_ascii = matrix(runif(200*30,-10,10),ncol=30)
+       nbytes = 8
+       endian = "little"
+
+       binarize(data_ascii, data_bin_file, 500, ",", nbytes, endian)
+       # Serialize transformation (just compute range) into a new binary file
+       trans_bin_file = "/tmp/epclust_test_t_trans.bin"
+       unlink(trans_bin_file)
+       getSeries = function(inds) getDataInFile(inds, data_bin_file, nbytes, endian)
+       binarizeTransform(getSeries, function(series) t(apply(series, 1, range)),
+               trans_bin_file, 250, nbytes, endian)
+       unlink(data_bin_file)
+       expect_equal(file.info(trans_bin_file)$size, 2*nrow(data_ascii)*nbytes+8)
+       for (indices in list(c(1,3,5), 3:13, c(5,20,50), c(75,130:135), 196:200))
+       {
+               trans_lines = getDataInFile(indices, trans_bin_file, nbytes, endian)
+               expect_equal(trans_lines, t(apply(data_ascii[indices,],1,range)), tolerance=1e-6)
+       }
+       unlink(trans_bin_file)
+})
+
 test_that("serialization + getDataInFile retrieve original data / from connection",
 {
        data_bin_file = "/tmp/epclust_test_c.bin"