fixes: TODO, debug test.clustering.R and finish writing clustering.R
authorBenjamin Auder <benjamin.auder@somewhere>
Fri, 10 Mar 2017 18:04:45 +0000 (19:04 +0100)
committerBenjamin Auder <benjamin.auder@somewhere>
Fri, 10 Mar 2017 18:04:45 +0000 (19:04 +0100)
epclust/R/clustering.R
epclust/R/de_serialize.R
epclust/R/main.R
epclust/tests/testthat/helper.clustering.R [new file with mode: 0644]
epclust/tests/testthat/helper.computeMedoidsIndices.R [new file with mode: 0644]
epclust/tests/testthat/test.clustering.R
epclust/tests/testthat/test.computeMedoidsIndices.R
epclust/tests/testthat/test.de_serialize.R
epclust/tests/testthat/test.filter.R

index 36b4769..b09c1bc 100644 (file)
 #'   \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_clust)
+#'   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
-#' @param contribs matrix of contributions (e.g. output of \code{curvesToContribs()})
-#' @param distances matrix of K1 x K1 (WER) distances between synchrones
 #' @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)
+#' @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, nb_items_clust1,
+clusteringTask1 = function(indices, getContribs, K1, algoClust1, nb_items_clust1,
        ncores_clust=1, verbose=FALSE, parll=TRUE)
 {
-       if (verbose)
-               cat(paste("*** Clustering task 1 on ",length(indices)," lines\n", sep=""))
-
        if (parll)
        {
                cl = parallel::makeCluster(ncores_clust, outfile = "")
@@ -43,19 +37,21 @@ clusteringTask1 = function(indices, getContribs, K1, nb_items_clust1,
        }
        while (length(indices) > K1)
        {
-               indices_workers = .spreadIndices(indices, nb_items_clust1, K1+1)
+               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[ computeClusters1(getContribs(inds), K1, verbose) ]
+                                       inds[ algoClust1(getContribs(inds), K1) ]
                                }) )
                        }
                        else
                        {
                                unlist( lapply(indices_workers, function(inds)
-                                       inds[ computeClusters1(getContribs(inds), K1, verbose) ]
+                                       inds[ algoClust1(getContribs(inds), K1) ]
                                ) )
                        }
        }
@@ -67,36 +63,20 @@ clusteringTask1 = function(indices, getContribs, K1, nb_items_clust1,
 
 #' @rdname clustering
 #' @export
-clusteringTask2 = function(medoids, K2, getRefSeries, nb_ref_curves,
-       nb_series_per_chunk, nbytes,endian,ncores_clust=1,verbose=FALSE,parll=TRUE)
+clusteringTask2 = function(medoids, K2, algoClust2, getRefSeries, nb_ref_curves,
+       nb_series_per_chunk, sync_mean, nbytes,endian,ncores_clust=1,verbose=FALSE,parll=TRUE)
 {
        if (verbose)
-               cat(paste("*** Clustering task 2 on ",nrow(medoids)," lines\n", sep=""))
+               cat(paste("*** Clustering task 2 on ",ncol(medoids)," synchrones\n", sep=""))
 
-       if (nrow(medoids) <= K2)
+       if (ncol(medoids) <= K2)
                return (medoids)
-       synchrones = computeSynchrones(medoids,
-               getRefSeries, nb_ref_curves, nb_series_per_chunk, ncores_clust, verbose, parll)
+       synchrones = computeSynchrones(medoids, getRefSeries, nb_ref_curves,
+               nb_series_per_chunk, sync_mean, ncores_clust, verbose, parll)
        distances = computeWerDists(synchrones, nbytes, endian, ncores_clust, verbose, parll)
-       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(        t(contribs)       , K1, diss=FALSE)$id.med
-}
-
-#' @rdname clustering
-#' @export
-computeClusters2 = function(distances, K2, verbose=FALSE)
-{
-       if (verbose)
-               cat(paste("   computeClusters2() on ",nrow(distances)," lines\n", sep=""))
-       cluster::pam(       distances        , K2, diss=TRUE)$id.med
+               cat(paste("   algoClust2() on ",nrow(distances)," items\n", sep=""))
+       medoids[ algoClust2(distances,K2), ]
 }
 
 #' computeSynchrones
@@ -113,12 +93,9 @@ computeClusters2 = function(distances, K2, verbose=FALSE)
 #' @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)
+computeSynchrones = function(medoids, getRefSeries, nb_ref_curves,
+       nb_series_per_chunk, sync_mean, ncores_clust=1,verbose=FALSE,parll=TRUE)
 {
-       if (verbose)
-               cat(paste("--- Compute synchrones\n", sep=""))
-
        computeSynchronesChunk = function(indices)
        {
                if (parll)
@@ -127,7 +104,8 @@ computeSynchrones = function(medoids, getRefSeries,
                        requireNamespace("synchronicity", quietly=TRUE)
                        require("epclust", quietly=TRUE)
                        synchrones <- bigmemory::attach.big.matrix(synchrones_desc)
-                       counts <- bigmemory::attach.big.matrix(counts_desc)
+                       if (sync_mean)
+                               counts <- bigmemory::attach.big.matrix(counts_desc)
                        medoids <- bigmemory::attach.big.matrix(medoids_desc)
                        m <- synchronicity::attach.mutex(m_desc)
                }
@@ -135,7 +113,7 @@ computeSynchrones = function(medoids, getRefSeries,
                ref_series = getRefSeries(indices)
                nb_series = nrow(ref_series)
 
-               #get medoids indices for this chunk of series
+               # Get medoids indices for this chunk of series
                mi = computeMedoidsIndices(medoids@address, ref_series)
 
                for (i in seq_len(nb_series))
@@ -143,17 +121,19 @@ computeSynchrones = function(medoids, getRefSeries,
                        if (parll)
                                synchronicity::lock(m)
                        synchrones[, mi[i] ] = synchrones[, mi[i] ] + ref_series[,i]
-                       counts[ mi[i] ] = counts[ mi[i] ] + 1 #TODO: remove counts? ...or as arg?!
+                       if (sync_mean)
+                               counts[ mi[i] ] = counts[ mi[i] ] + 1
                        if (parll)
                                synchronicity::unlock(m)
                }
        }
 
-       K = nrow(medoids) ; L = ncol(medoids)
+       K = ncol(medoids) ; L = nrow(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=L, ncol=K, type="double", init=0.)
-       counts = bigmemory::big.matrix(nrow=K, ncol=1, type="double", init=0)
+       if (sync_mean)
+               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")
@@ -162,13 +142,21 @@ computeSynchrones = function(medoids, getRefSeries,
                m <- synchronicity::boost.mutex()
                m_desc <- synchronicity::describe(m)
                synchrones_desc = bigmemory::describe(synchrones)
-               counts_desc = bigmemory::describe(counts)
+               if (sync_mean)
+                       counts_desc = bigmemory::describe(counts)
                medoids_desc = bigmemory::describe(medoids)
                cl = parallel::makeCluster(ncores_clust)
-               parallel::clusterExport(cl, varlist=c("synchrones_desc","counts_desc","counts",
-                       "verbose","m_desc","medoids_desc","getRefSeries"), envir=environment())
+               varlist=c("synchrones_desc","sync_mean","m_desc","medoids_desc","getRefSeries")
+               if (sync_mean)
+                       varlist = c(varlist, "counts_desc")
+               parallel::clusterExport(cl, varlist, envir=environment())
        }
 
+       if (verbose)
+       {
+               if (verbose)
+                       cat(paste("--- Compute ",K," synchrones with ",nb_ref_curves," series\n", sep=""))
+       }
        indices_workers = .spreadIndices(seq_len(nb_ref_curves), nb_series_per_chunk)
        ignored <-
                if (parll)
@@ -179,7 +167,10 @@ computeSynchrones = function(medoids, getRefSeries,
        if (parll)
                parallel::stopCluster(cl)
 
-       #TODO: can we avoid this loop? ( synchrones = sweep(synchrones, 1, counts, '/') )
+       if (!sync_mean)
+               return (synchrones)
+
+       #TODO: can we avoid this loop? ( synchrones = sweep(synchrones, 2, counts, '/') )
        for (i in seq_len(K))
                synchrones[,i] = synchrones[,i] / counts[i]
        #NOTE: odds for some clusters to be empty? (when series already come from stage 2)
@@ -205,9 +196,6 @@ computeSynchrones = function(medoids, getRefSeries,
 #' @export
 computeWerDists = function(synchrones, nbytes,endian,ncores_clust=1,verbose=FALSE,parll=TRUE)
 {
-       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)
@@ -260,7 +248,12 @@ computeWerDists = function(synchrones, nbytes,endian,ncores_clust=1,verbose=FALS
                parallel::clusterExport(cl, varlist=c("synchrones_desc","Xwer_dist_desc","totnoct",
                        "nvoice","w0","s0log","noctave","s0","verbose","getCWT"), envir=environment())
        }
-
+       
+       if (verbose)
+       {
+               cat(paste("--- Compute WER dists\n", sep=""))
+       #       precompute save all CWT........
+       }
        #precompute and serialize all CWT
        ignored <-
                if (parll)
@@ -301,6 +294,10 @@ computeWerDists = function(synchrones, nbytes,endian,ncores_clust=1,verbose=FALS
                Xwer_dist[i,i] = 0.
        }
 
+       if (verbose)
+       {
+               cat(paste("--- Compute WER dists\n", sep=""))
+       }
        ignored <-
                if (parll)
                        parallel::parLapply(cl, pairs, computeDistancesIJ)
@@ -317,11 +314,11 @@ computeWerDists = function(synchrones, nbytes,endian,ncores_clust=1,verbose=FALS
 }
 
 # Helper function to divide indices into balanced sets
-.spreadIndices = function(indices, max_per_set, min_nb_per_set = 1)
+.spreadIndices = function(indices, nb_per_set)
 {
        L = length(indices)
-       min_nb_workers = floor( L / max_per_set )
-       rem = L %% max_per_set
+       nb_workers = floor( L / nb_per_set )
+       rem = L %% max_nb_per_set
        if (nb_workers == 0 || (nb_workers==1 && rem==0))
        {
                # L <= max_nb_per_set, simple case
@@ -330,19 +327,9 @@ computeWerDists = function(synchrones, nbytes,endian,ncores_clust=1,verbose=FALS
        else
        {
                indices_workers = lapply( seq_len(nb_workers), function(i)
-                       indices[(max_nb_per_set*(i-1)+1):(max_per_set*i)] )
-               # Two cases: remainder is >= min_per_set (easy)...
-               if (rem >= min_nb_per_set)
-                       indices_workers = c( indices_workers, list(tail(indices,rem)) )
-               #...or < min_per_set: harder, need to remove indices from current sets to feed
-               # the too-small remainder. It may fail: then fallback to "slightly bigger sets"
-               else
-               {
-                       save_indices_workers = indices_workers
-                       small_set = tail(indices,rem)
-                       # Try feeding small_set until it reaches min_per_set, whle keeping the others big enough
-                       # Spread the remaining load among the workers
-               rem = L %% nb_per_chunk
+                       indices[(nb_per_chunk*(i-1)+1):(nb_per_set*i)] )
+               # Spread the remaining load among the workers
+               rem = L %% nb_per_set
                while (rem > 0)
                {
                        index = rem%%nb_workers + 1
index f04c13a..fd3a5db 100644 (file)
@@ -44,7 +44,7 @@ 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 )
+               if (is_matrix)
                        data_length = nrow(data_ascii)
                else #connection
                {
@@ -58,11 +58,11 @@ binarize = function(data_ascii, data_bin_file, nb_per_chunk,
                index = 1
        repeat
        {
-               if ( is_matrix )
+               if (is_matrix)
                {
                        data_chunk =
                                if (index <= ncol(data_ascii))
-                                       as.double(data_ascii[,index:min(nrow(data_ascii),index+nb_per_chunk-1)])
+                                       as.double(data_ascii[,index:min(ncol(data_ascii),index+nb_per_chunk-1)])
                                else
                                        double(0)
                        index = index + nb_per_chunk
@@ -113,13 +113,13 @@ getDataInFile = function(indices, data_bin_file, nbytes=4, endian=.Platform$endi
        data_bin = file(data_bin_file, "rb")
        data_size = file.info(data_bin_file)$size
        data_length = readBin(data_bin, "integer", n=1, size=8, endian=endian)
-       data_ascii = sapply( indices, function(i) {
+       data_ascii = do.call( cbind, lapply( indices, function(i) {
                offset = 8+(i-1)*data_length*nbytes
                if (offset > data_size)
-                       return (vector("double",0))
+                       return (NULL)
                ignored = seek(data_bin, offset)
                readBin(data_bin, "double", n=data_length, size=nbytes, endian=endian)
-       } )
+       } ) )
        close(data_bin)
-       if (ncol(data_ascii)>0) data_ascii else NULL
+       data_ascii
 }
index a039d1c..9e9b641 100644 (file)
@@ -11,7 +11,7 @@
 #'     \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_clust}
+#'         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
 #' @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 algo_clust1 Clustering algorithm for stage 1. A function which takes (data, K)
+#' @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
-#' @param algo_clust2 Clustering algorithm for stage 2. A function which takes (dists, K)
+#'   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: k-means
+#'   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.
@@ -53,6 +56,7 @@
 #' @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.
 #' @export
 claws <- function(getSeries, K1, K2, nb_series_per_chunk,
        nb_items_clust1=7*K1,
-       algo_clust1=function(data,K) cluster::pam(data,K,diss=FALSE),
-       algo_clust2=function(dists,K) stats::kmeans(dists,K,iter.max=50,nstart=3),
+       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",
+       WER="end",sync_mean=TRUE,
        random=TRUE,
        ntasks=1, ncores_tasks=1, ncores_clust=4,
        sep=",",
@@ -162,17 +166,15 @@ claws <- function(getSeries, K1, K2, 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")
-       )
+       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)
@@ -234,8 +236,8 @@ claws <- function(getSeries, K1, K2, nb_series_per_chunk,
                # 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","algo_clust1","algo_clust2",
-                       "nb_series_per_chunk","nb_items_clust","ncores_clust","sep",
+               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")
@@ -253,14 +255,14 @@ claws <- function(getSeries, K1, K2, nb_series_per_chunk,
                if (parll && ntasks>1)
                        require("epclust", quietly=TRUE)
                indices_medoids = clusteringTask1(
-                       inds, getContribs, K1, nb_series_per_chunk, ncores_clust, verbose, parll)
+                       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, getSeries, nb_curves, nb_series_per_chunk,
-                               nbytes, endian, ncores_clust, verbose, parll)
+                       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))
                }
@@ -321,11 +323,11 @@ claws <- function(getSeries, K1, K2, nb_series_per_chunk,
        # 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, nb_series_per_chunk, ncores_tasks*ncores_clust, verbose, parll)
+       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, getRefSeries, nb_curves, nb_series_per_chunk,
-               nbytes, endian, ncores_tasks*ncores_clust, verbose, parll)
+       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)
diff --git a/epclust/tests/testthat/helper.clustering.R b/epclust/tests/testthat/helper.clustering.R
new file mode 100644 (file)
index 0000000..21a791a
--- /dev/null
@@ -0,0 +1,15 @@
+#shorthand: map 1->1, 2->2, 3->3, 4->1, ..., 149->2, 150->3, ... (is base==3)
+I = function(i, base)
+       (i-1) %% base + 1
+
+# NOTE: medoids can be a big.matrix
+computeDistortion = function(series, medoids)
+{
+       n = nrow(series) ; L = ncol(series)
+       distortion = 0.
+       if (bigmemory::is.big.matrix(medoids))
+               medoids = medoids[,]
+       for (i in seq_len(n))
+               distortion = distortion + min( colSums( sweep(medoids,1,series[,i],'-')^2 ) / L )
+       distortion / n
+}
diff --git a/epclust/tests/testthat/helper.computeMedoidsIndices.R b/epclust/tests/testthat/helper.computeMedoidsIndices.R
new file mode 100644 (file)
index 0000000..9342feb
--- /dev/null
@@ -0,0 +1,10 @@
+#R-equivalent of computeMedoidsIndices, requiring a matrix
+#(thus potentially breaking "fit-in-memory" hope)
+R_computeMedoidsIndices <- function(medoids, series)
+{
+       nb_series = ncol(series)
+       mi = rep(NA,nb_series)
+       for (i in 1:nb_series)
+               mi[i] <- which.min( colSums( sweep(medoids, 1, series[,i], '-')^2 ) )
+       mi
+}
index 7116e73..a5dc3bd 100644 (file)
@@ -1,61 +1,5 @@
 context("clustering")
 
-#shorthand: map 1->1, 2->2, 3->3, 4->1, ..., 149->2, 150->3, ... (is base==3)
-I = function(i, base)
-       (i-1) %% base + 1
-
-test_that("computeClusters1&2 behave as expected",
-{
-       require("MASS", quietly=TRUE)
-       if (!require("clue", quietly=TRUE))
-               skip("'clue' package not available")
-
-       # 3 gaussian clusters, 300 items; and then 7 gaussian clusters, 490 items
-       n = 300
-       d = 5
-       K = 3
-       for (ndK in list( c(300,5,3), c(490,10,7) ))
-       {
-               n = ndK[1] ; d = ndK[2] ; K = ndK[3]
-               cs = n/K #cluster size
-               Id = diag(d)
-               coefs = sapply(1:K, function(i) MASS::mvrnorm(cs, c(rep(0,(i-1)),5,rep(0,d-i)), Id))
-               indices_medoids1 = computeClusters1(coefs, K, verbose=TRUE)
-               indices_medoids2 = computeClusters2(dist(coefs), K, verbose=TRUE)
-               # Get coefs assignments (to medoids)
-               assignment1 = sapply(seq_len(n), function(i)
-                       which.min( colSums( sweep(coefs[,indices_medoids1],1,coefs[,i],'-')^2 ) ) )
-               assignment2 = sapply(seq_len(n), function(i)
-                       which.min( colSums( sweep(coefs[,indices_medoids2],1,coefs[,i],'-')^2 ) ) )
-               for (i in 1:K)
-               {
-                       expect_equal(sum(assignment1==i), cs, tolerance=5)
-                       expect_equal(sum(assignment2==i), cs, tolerance=5)
-               }
-
-               costs_matrix1 = matrix(nrow=K,ncol=K)
-               costs_matrix2 = matrix(nrow=K,ncol=K)
-               for (i in 1:K)
-               {
-                       for (j in 1:K)
-                       {
-                               # assign i (in result) to j (order 1,2,3)
-                               costs_matrix1[i,j] = abs( mean(assignment1[((i-1)*cs+1):(i*cs)]) - j )
-                               costs_matrix2[i,j] = abs( mean(assignment2[((i-1)*cs+1):(i*cs)]) - j )
-                       }
-               }
-               permutation1 = as.integer( clue::solve_LSAP(costs_matrix1) )
-               permutation2 = as.integer( clue::solve_LSAP(costs_matrix2) )
-               for (i in 1:K)
-               {
-                       expect_equal(
-                               mean(assignment1[((i-1)*cs+1):(i*cs)]), permutation1[i], tolerance=0.05)
-                       expect_equal(
-                               mean(assignment2[((i-1)*cs+1):(i*cs)]), permutation2[i], tolerance=0.05)
-               }
-       }
-})
-
 test_that("computeSynchrones behave as expected",
 {
        n = 300
@@ -77,24 +21,39 @@ test_that("computeSynchrones behave as expected",
                if (length(indices)>0) series[,indices] else NULL
        }
        synchrones = computeSynchrones(bigmemory::as.big.matrix(cbind(s1,s2,s3)), getRefSeries,
-               n, 100, verbose=TRUE, parll=FALSE)
+               n, 100, sync_mean=TRUE, verbose=TRUE, parll=FALSE)
 
        expect_equal(dim(synchrones), c(L,K))
        for (i in 1:K)
                expect_equal(synchrones[,i], s[[i]], tolerance=0.01)
 })
 
-# NOTE: medoids can be a big.matrix
-computeDistortion = function(series, medoids)
+# Helper function to divide indices into balanced sets
+test_that("Helper function to spread indices work properly",
 {
-       n = nrow(series) ; L = ncol(series)
-       distortion = 0.
-       if (bigmemory::is.big.matrix(medoids))
-               medoids = medoids[,]
-       for (i in seq_len(n))
-               distortion = distortion + min( colSums( sweep(medoids,1,series[,i],'-')^2 ) / L )
-       distortion / n
-}
+       indices <- 1:400
+
+       # bigger nb_per_set than length(indices)
+       expect_equal(epclust:::.spreadIndices(indices,500), list(indices))
+
+       # nb_per_set == length(indices)
+       expect_equal(epclust:::.spreadIndices(indices,400), list(indices))
+
+       # length(indices) %% nb_per_set == 0
+       expect_equal(epclust:::.spreadIndices(indices,200),
+               c( list(indices[1:200]), list(indices[201:400]) ))
+       expect_equal(epclust:::.spreadIndices(indices,100),
+               c( list(indices[1:100]), list(indices[101:200]),
+                       list(indices[201:300]), list(indices[301:400]) ))
+
+       # length(indices) / nb_per_set == 1, length(indices) %% nb_per_set == 100
+       expect_equal(epclust:::.spreadIndices(indices,300), list(indices))
+       # length(indices) / nb_per_set == 2, length(indices) %% nb_per_set == 42
+       repartition <- epclust:::.spreadIndices(indices,179)
+       expect_equal(length(repartition), 2)
+       expect_equal(length(repartition[[1]]), 179 + 21)
+       expect_equal(length(repartition[[1]]), 179 + 21)
+})
 
 test_that("clusteringTask1 behave as expected",
 {
@@ -151,36 +110,3 @@ test_that("clusteringTask2 behave as expected",
        for (i in 1:3)
                expect_lte( distorGood, computeDistortion(series,medoids_K1[,sample(1:K1, K2)]) )
 })
-
-#NOTE: rather redundant test
-#test_that("clusteringTask1 + clusteringTask2 behave as expected",
-#{
-#      n = 900
-#      x = seq(0,9.5,0.1)
-#      L = length(x) #96 1/4h
-#      K1 = 60
-#      K2 = 3
-#      s = lapply( seq_len(K1), function(i) x^(1+i/30)*cos(x+i) )
-#      series = matrix(nrow=n, ncol=L)
-#      for (i in seq_len(n))
-#              series[i,] = s[[I(i,K1)]] + rnorm(L,sd=0.01)
-#      getSeries = function(indices) {
-#              indices = indices[indices <= n]
-#              if (length(indices)>0) series[indices,] else NULL
-#      }
-#      wf = "haar"
-#      ctype = "absolute"
-#      getContribs = function(indices) curvesToContribs(series[indices,],wf,ctype)
-#      require("bigmemory", quietly=TRUE)
-#      indices1 = clusteringTask1(1:n, getContribs, K1, 75, verbose=TRUE, parll=FALSE)
-#      medoids_K1 = bigmemory::as.big.matrix( getSeries(indices1) )
-#      medoids_K2 = clusteringTask2(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))
-#      # 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),]) )
-#})
index be5b2b4..efd6af9 100644 (file)
@@ -1,42 +1,24 @@
 context("computeMedoidsIndices")
 
-test_that("serialization + getDataInFile retrieve original data / from matrix",
+test_that("computeMedoidsIndices behave as expected",
 {
-       data_bin_file = "/tmp/epclust_test_m.bin"
-       unlink(data_bin_file)
+       # Generate a gaussian mixture
+       n = 999
+       L = 7
+       medoids = cbind( rep(0,L), rep(-5,L), rep(5,L) )
+       # short series...
+       series = t( rbind( MASS::mvrnorm(n/3, medoids[,1], diag(L)),
+               MASS::mvrnorm(n/3, medoids[,2], diag(L),
+               MASS::mvrnorm(n/3, medoids[,3], diag(L))) ) )
 
-       #dataset 200 lignes / 30 columns
-       data_ascii = matrix(runif(200*30,-10,10),ncol=30)
-       nbytes = 4 #lead to a precision of 1e-7 / 1e-8
-       endian = "little"
+       # With high probability, medoids indices should resemble 1,1,1,...,2,2,2,...,3,3,3,...
+       mi = epclust:::.computeMedoidsIndices(medoids, series)
+       mi_ref = rep(1:3, each=n/3)
+       expect_that( mean(mi != mi_ref) < 0.01 )
 
-       #Simulate serialization in one single call
-       binarize(data_ascii, data_bin_file, 500, ",", nbytes, endian)
-       expect_equal(file.info(data_bin_file)$size, length(data_ascii)*nbytes+8)
-       for (indices in list(c(1,3,5), 3:13, c(5,20,50), c(75,130:135), 196:200))
-       {
-               data_lines = getDataInFile(indices, data_bin_file, nbytes, endian)
-               expect_equal(data_lines, data_ascii[indices,], tolerance=1e-6)
-       }
-       unlink(data_bin_file)
-
-       #...in several calls (last call complete, next call NULL)
-       for (i in 1:20)
-               binarize(data_ascii[((i-1)*10+1):(i*10),], data_bin_file, 20, ",", nbytes, endian)
-       expect_equal(file.info(data_bin_file)$size, length(data_ascii)*nbytes+8)
-       for (indices in list(c(1,3,5), 3:13, c(5,20,50), c(75,130:135), 196:200))
-       {
-               data_lines = getDataInFile(indices, data_bin_file, nbytes, endian)
-               expect_equal(data_lines, data_ascii[indices,], tolerance=1e-6)
-       }
-       unlink(data_bin_file)
+       # Now with a random matrix, compare with (trusted) R version
+       series = matrix(runif(n*L, min=-7, max=7), nrow=L)
+       mi = epclust:::.computeMedoidsIndices(medoids, series)
+       mi_ref = R_computeMedoidsIndices(medoids, series)
+       expect_equal(mi, mi_ref)
 })
-
-TODO: test computeMedoids + filter
-#              #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()
-
index 25f2c2b..be49020 100644 (file)
@@ -22,7 +22,7 @@ test_that("serialization + getDataInFile retrieve original data / from matrix",
 
        #...in several calls (last call complete, next call NULL)
        for (i in 1:20)
-               binarize(data_ascii[((i-1)*10+1):(i*10),], data_bin_file, 20, ",", nbytes, endian)
+               binarize(data_ascii[,((i-1)*10+1):(i*10)], data_bin_file, 20, ",", nbytes, endian)
        expect_equal(file.info(data_bin_file)$size, length(data_ascii)*nbytes+8)
        for (indices in list(c(1,3,5), 3:13, c(5,20,50), c(75,130:135), 196:200))
        {
@@ -50,11 +50,11 @@ test_that("serialization + transform + getDataInFile retrieve original transform
        binarizeTransform(getSeries, function(series) apply(series, 2, range),
                trans_bin_file, 250, nbytes, endian)
        unlink(data_bin_file)
-       expect_equal(file.info(trans_bin_file)$size, 2*nrow(data_ascii)*nbytes+8)
+       expect_equal(file.info(trans_bin_file)$size, 2*ncol(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_cols = getDataInFile(indices, trans_bin_file, nbytes, endian)
-               expect_equal(trans_cols, apply(data_ascii[indices,],2,range), tolerance=1e-6)
+               expect_equal(trans_cols, apply(data_ascii[,indices],2,range), tolerance=1e-6)
        }
        unlink(trans_bin_file)
 })
@@ -68,15 +68,16 @@ test_that("serialization + getDataInFile retrieve original data / from connectio
        data_csv = system.file("testdata","de_serialize.csv",package="epclust")
        nbytes = 8
        endian = "big"
-       data_ascii = as.matrix(read.csv(data_csv, sep=";", header=FALSE)) #for ref
+       data_ascii = t( as.matrix(read.table(data_csv, sep=";", header=FALSE)) ) #for ref
 
        #Simulate serialization in one single call
        binarize(data_csv, data_bin_file, 350, ";", nbytes, endian)
        expect_equal(file.info(data_bin_file)$size, 300*50*8+8)
        for (indices in list(c(1,3,5), 3:13, c(5,20,50), c(75,130:135), 196:200))
        {
-               #HACK: as.matrix(as.data.frame( )) required to match (ref) data structure
-               data_cols = as.matrix(as.data.frame( getDataInFile(indices,data_bin_file,nbytes,endian) ))
+               data_cols = getDataInFile(indices,data_bin_file,nbytes,endian)
+               #HACK: rows naming required to match (ref) data structure
+               rownames(data_cols) <- paste("V",seq_len(nrow(data_ascii)), sep="")
                expect_equal(data_cols, data_ascii[,indices])
        }
        unlink(data_bin_file)
@@ -87,7 +88,8 @@ test_that("serialization + getDataInFile retrieve original data / from connectio
        expect_equal(file.info(data_bin_file)$size, 300*50*8+8)
        for (indices in list(c(1,3,5), 3:13, c(5,20,50), c(75,130:135), 196:200))
        {
-               data_cols = as.matrix(as.data.frame( getDataInFile(indices,data_bin_file,nbytes,endian) ))
+               data_cols = getDataInFile(indices,data_bin_file,nbytes,endian)
+               rownames(data_cols) <- paste("V",seq_len(nrow(data_ascii)), sep="")
                expect_equal(data_cols, data_ascii[,indices])
        }
        unlink(data_bin_file)
index d94a5ac..9d1916d 100644 (file)
@@ -1,8 +1,18 @@
-TODO: test computeMedoids + filter
-#              #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()
+context("epclustFilter")
 
+#TODO: find a better name
+test_that("[time-]serie filtering behave as expected",
+{
+       # Currently just a mean of 3 values
+       M = matrix(runif(1000,min=-7,max=7), ncol=10)
+       ref_fM = stats::filter(M, c(1/3,1/3,1/3), circular=FALSE)
+       fM = epclust:::.epclustFilter(M)
+
+       #Expect an agreement on all inner values
+       expect_equal(dim(fM), c(100,10))
+       expect_equal(fM[2:99,], ref_fM[,2:99])
+
+       #For border values, just apply a "by-hand" mean
+       expect_equal(fM[1,], colMeans(rbind(M[1,],M[2,],M[100,])))
+       expect_equal(fM[100,], colMeans(rbind(M[100,],M[99,],M[1,])))
+}