epclust/R/main.R

   1 #' CLAWS: CLustering with wAvelets and Wer distanceS
   2 #'
   3 #' Cluster electricity power curves (or any series of similar nature) by applying a
   4 #' two stage procedure in parallel (see details).
   5 #' Input series must be sampled on the same time grid, no missing values.
   6 #'
   7 #' Summary of the function execution flow:
   8 #' \enumerate{
   9 #'   \item Compute and serialize all contributions, obtained through discrete wavelet
  10 #'     decomposition (see Antoniadis & al. [2013])
  11 #'   \item Divide series into \code{ntasks} groups to process in parallel. In each task:
  12 #'   \enumerate{
  13 #'     \item iterate the first clustering algorithm on its aggregated outputs,
  14 #'       on inputs of size \code{nb_items_clust1}
  15 #'     \item optionally, if WER=="mix":
  16 #'       a) compute the K1 synchrones curves,
  17 #'       b) compute WER distances (K1xK1 matrix) between synchrones and
  18 #'       c) apply the second clustering algorithm
  19 #'   }
  20 #'   \item Launch a final task on the aggregated outputs of all previous tasks:
  21 #'     in the case WER=="end" this task takes indices in input, otherwise
  22 #'     (medoid) curves
  23 #' }
  24 #' \cr
  25 #' The main argument -- \code{getSeries} -- has a quite misleading name, since it can be
  26 #' either a [big.]matrix, a CSV file, a connection or a user function to retrieve
  27 #' series; the name was chosen because all types of arguments are converted to a function.
  28 #' When \code{getSeries} is given as a function, it must take a single argument,
  29 #' 'indices', integer vector equal to the indices of the curves to retrieve;
  30 #' see SQLite example. The nature and role of other arguments should be clear
  31 #' \cr
  32 #' Note: Since we don't make assumptions on initial data, there is a possibility that
  33 #' even when serialized, contributions or synchrones do not fit in RAM. For example,
  34 #' 30e6 series of length 100,000 would lead to a +4Go contribution matrix. Therefore,
  35 #' it's safer to place these in (binary) files; that's what we do.
  36 #'
  37 #' @param getSeries Access to the (time-)series, which can be of one of the three
  38 #'   following types:
  39 #'   \itemize{
  40 #'     \item [big.]matrix: each column contains the (time-ordered) values of one time-serie
  41 #'     \item connection: any R connection object providing lines as described above
  42 #'     \item character: name of a CSV file containing series in rows (no header)
  43 #'     \item function: a custom way to retrieve the curves; it has only one argument:
  44 #'       the indices of the series to be retrieved. See SQLite example
  45 #'   }
  46 #' @param K1 Number of clusters to be found after stage 1 (K1 << N [number of series])
  47 #' @param K2 Number of clusters to be found after stage 2 (K2 << K1)
  48 #' @param nb_series_per_chunk (Maximum) number of series to retrieve in one batch
  49 #' @param algoClust1 Clustering algorithm for stage 1. A function which takes (data, K)
  50 #'   as argument where data is a matrix in columns and K the desired number of clusters,
  51 #'   and outputs K medoids ranks. Default: PAM. In our method, this function is called
  52 #'   on iterated medoids during stage 1
  53 #' @param algoClust2 Clustering algorithm for stage 2. A function which takes (dists, K)
  54 #'   as argument where dists is a matrix of distances and K the desired number of clusters,
  55 #'   and outputs K medoids ranks. Default: PAM.  In our method, this function is called
  56 #'   on a matrix of K1 x K1 (WER) distances computed between synchrones
  57 #' @param nb_items_clust1 (~Maximum) number of items in input of the clustering algorithm
  58 #'   for stage 1. At worst, a clustering algorithm might be called with ~2*nb_items_clust1
  59 #'   items; but this could only happen at the last few iterations.
  60 #' @param wav_filt Wavelet transform filter; see ?wavelets::wt.filter
  61 #' @param contrib_type Type of contribution: "relative", "logit" or "absolute" (any prefix)
  62 #' @param WER "end" to apply stage 2 after stage 1 has fully iterated, or "mix" to apply
  63 #'   stage 2 at the end of each task
  64 #' @param random TRUE (default) for random chunks repartition
  65 #' @param ntasks Number of tasks (parallel iterations to obtain K1 [if WER=="end"]
  66 #'   or K2 [if WER=="mix"] medoids); default: 1.
  67 #'   Note: ntasks << N (number of series), so that N is "roughly divisible" by ntasks
  68 #' @param ncores_tasks Number of parallel tasks (1 to disable: sequential tasks)
  69 #' @param ncores_clust Number of parallel clusterings in one task (4 should be a minimum)
  70 #' @param sep Separator in CSV input file (if any provided)
  71 #' @param nbytes Number of bytes to serialize a floating-point number; 4 or 8
  72 #' @param endian Endianness for (de)serialization ("little" or "big")
  73 #' @param verbose Level of verbosity (0/FALSE for nothing or 1/TRUE for all; devel stage)
  74 #' @param parll TRUE to fully parallelize; otherwise run sequentially (debug, comparison)
  75 #'
  76 #' @return A matrix of the final K2 medoids curves, in columns
  77 #'
  78 #' @references Clustering functional data using Wavelets [2013];
  79 #'   A. Antoniadis, X. Brossat, J. Cugliari & J.-M. Poggi.
  80 #'   Inter. J. of Wavelets, Multiresolution and Information Procesing,
  81 #'   vol. 11, No 1, pp.1-30. doi:10.1142/S0219691313500033
  82 #'
  83 #' @examples
  84 #' \dontrun{
  85 #' # WER distances computations are too long for CRAN (for now)
  86 #'
  87 #' # Random series around cos(x,2x,3x)/sin(x,2x,3x)
  88 #' x = seq(0,500,0.05)
  89 #' L = length(x) #10001
  90 #' ref_series = matrix( c(cos(x),cos(2*x),cos(3*x),sin(x),sin(2*x),sin(3*x)), ncol=6 )
  91 #' library(wmtsa)
  92 #' series = do.call( cbind, lapply( 1:6, function(i)
  93 #'   do.call(cbind, wmtsa::wavBootstrap(ref_series[i,], n.realization=400)) ) )
  94 #' #dim(series) #c(2400,10001)
  95 #' medoids_ascii = claws(series, K1=60, K2=6, 200, verbose=TRUE)
  96 #'
  97 #' # Same example, from CSV file
  98 #' csv_file = "/tmp/epclust_series.csv"
  99 #' write.table(series, csv_file, sep=",", row.names=FALSE, col.names=FALSE)
 100 #' medoids_csv = claws(csv_file, K1=60, K2=6, 200)
 101 #'
 102 #' # Same example, from binary file
 103 #' bin_file <- "/tmp/epclust_series.bin"
 104 #' nbytes <- 8
 105 #' endian <- "little"
 106 #' binarize(csv_file, bin_file, 500, nbytes, endian)
 107 #' getSeries <- function(indices) getDataInFile(indices, bin_file, nbytes, endian)
 108 #' medoids_bin <- claws(getSeries, K1=60, K2=6, 200)
 109 #' unlink(csv_file)
 110 #' unlink(bin_file)
 111 #'
 112 #' # Same example, from SQLite database
 113 #' library(DBI)
 114 #' series_db <- dbConnect(RSQLite::SQLite(), "file::memory:")
 115 #' # Prepare data.frame in DB-format
 116 #' n <- nrow(series)
 117 #' time_values <- data.frame(
 118 #'   id = rep(1:n,each=L),
 119 #'   time = rep( as.POSIXct(1800*(0:n),"GMT",origin="2001-01-01"), L ),
 120 #'   value = as.double(t(series)) )
 121 #' dbWriteTable(series_db, "times_values", times_values)
 122 #' # Fill associative array, map index to identifier
 123 #' indexToID_inDB <- as.character(
 124 #'   dbGetQuery(series_db, 'SELECT DISTINCT id FROM time_values')[,"id"] )
 125 #' serie_length <- as.integer( dbGetQuery(series_db,
 126 #'   paste("SELECT COUNT * FROM time_values WHERE id == ",indexToID_inDB[1],sep="")) )
 127 #' getSeries <- function(indices) {
 128 #'   request <- "SELECT id,value FROM times_values WHERE id in ("
 129 #'   for (i in indices)
 130 #'     request <- paste(request, indexToID_inDB[i], ",", sep="")
 131 #'   request <- paste(request, ")", sep="")
 132 #'   df_series <- dbGetQuery(series_db, request)
 133 #'   as.matrix(df_series[,"value"], nrow=serie_length)
 134 #' }
 135 #' medoids_db = claws(getSeries, K1=60, K2=6, 200))
 136 #' dbDisconnect(series_db)
 137 #'
 138 #' # All computed medoids should be the same:
 139 #' digest::sha1(medoids_ascii)
 140 #' digest::sha1(medoids_csv)
 141 #' digest::sha1(medoids_bin)
 142 #' digest::sha1(medoids_db)
 143 #' }
 144 #' @export
 145 claws <- function(getSeries, K1, K2, nb_series_per_chunk, nb_items_clust1=7*K1,
 146     algoClust1=function(data,K) cluster::pam(t(data),K,diss=FALSE)$id.med,
 147     algoClust2=function(dists,K) cluster::pam(dists,K,diss=TRUE)$id.med,
 148     wav_filt="d8", contrib_type="absolute", WER="end", random=TRUE,
 149     ntasks=1, ncores_tasks=1, ncores_clust=4, sep=",", nbytes=4,
 150     endian=.Platform$endian, verbose=FALSE, parll=TRUE)
 151 {
 152     # Check/transform arguments
 153     if (!is.matrix(getSeries) && !bigmemory::is.big.matrix(getSeries)
 154         && !is.function(getSeries)
 155         && !methods::is(getSeries,"connection") && !is.character(getSeries))
 156     {
 157         stop("'getSeries': [big]matrix, function, file or valid connection (no NA)")
 158     }
 159     K1 <- .toInteger(K1, function(x) x>=2)
 160     K2 <- .toInteger(K2, function(x) x>=2)
 161     nb_series_per_chunk <- .toInteger(nb_series_per_chunk, function(x) x>=1)
 162     # K1 (number of clusters at step 1) cannot exceed nb_series_per_chunk, because we will need
 163     # to load K1 series in memory for clustering stage 2.
 164     if (K1 > nb_series_per_chunk)
 165         stop("'K1' cannot exceed 'nb_series_per_chunk'")
 166     nb_items_clust1 <- .toInteger(nb_items_clust1, function(x) x>K1)
 167     random <- .toLogical(random)
 168     tryCatch( {ignored <- wavelets::wt.filter(wav_filt)},
 169         error = function(e) stop("Invalid wavelet filter; see ?wavelets::wt.filter") )
 170     ctypes = c("relative","absolute","logit")
 171     contrib_type = ctypes[ pmatch(contrib_type,ctypes) ]
 172     if (is.na(contrib_type))
 173         stop("'contrib_type' in {'relative','absolute','logit'}")
 174     if (WER!="end" && WER!="mix")
 175         stop("'WER': in {'end','mix'}")
 176     random <- .toLogical(random)
 177     ntasks <- .toInteger(ntasks, function(x) x>=1)
 178     ncores_tasks <- .toInteger(ncores_tasks, function(x) x>=1)
 179     ncores_clust <- .toInteger(ncores_clust, function(x) x>=1)
 180     if (!is.character(sep))
 181         stop("'sep': character")
 182     nbytes <- .toInteger(nbytes, function(x) x==4 || x==8)
 183     verbose <- .toLogical(verbose)
 184     parll <- .toLogical(parll)
 185
 186     # Binarize series if getSeries is not a function; the aim is to always use a function,
 187     # to uniformize treatments. An equally good alternative would be to use a file-backed
 188     # bigmemory::big.matrix, but it would break the "all-is-function" pattern.
 189     if (!is.function(getSeries))
 190     {
 191         if (verbose)
 192             cat("...Serialize time-series\n")
 193         series_file = ".series.bin" ; unlink(series_file)
 194         binarize(getSeries, series_file, nb_series_per_chunk, sep, nbytes, endian)
 195         getSeries = function(inds) getDataInFile(inds, series_file, nbytes, endian)
 196     }
 197
 198     # Serialize all computed wavelets contributions into a file
 199     contribs_file = ".contribs.bin" ; unlink(contribs_file)
 200     index = 1
 201     nb_curves = 0
 202     if (verbose)
 203         cat("...Compute contributions and serialize them\n")
 204     nb_curves = binarizeTransform(getSeries,
 205         function(series) curvesToContribs(series, wav_filt, contrib_type),
 206         contribs_file, nb_series_per_chunk, nbytes, endian)
 207     getContribs = function(indices) getDataInFile(indices, contribs_file, nbytes, endian)
 208
 209     # A few sanity checks: do not continue if too few data available.
 210     if (nb_curves < K2)
 211         stop("Not enough data: less series than final number of clusters")
 212     nb_series_per_task = round(nb_curves / ntasks)
 213     if (nb_series_per_task < K2)
 214         stop("Too many tasks: less series in one task than final number of clusters")
 215
 216     # Generate a random permutation of 1:N (if random==TRUE);
 217     # otherwise just use arrival (storage) order.
 218     indices_all = if (random) sample(nb_curves) else seq_len(nb_curves)
 219     # Split (all) indices into ntasks groups of ~same size
 220     indices_tasks = lapply(seq_len(ntasks), function(i) {
 221         upper_bound = ifelse( i<ntasks, min(nb_series_per_task*i,nb_curves), nb_curves )
 222         indices_all[((i-1)*nb_series_per_task+1):upper_bound]
 223     })
 224
 225     if (parll && ntasks>1)
 226     {
 227         # Initialize parallel runs: outfile="" allow to output verbose traces in the console
 228         # under Linux. All necessary variables are passed to the workers.
 229         cl = parallel::makeCluster(ncores_tasks, outfile="")
 230         varlist = c("getSeries","getContribs","K1","K2","algoClust1","algoClust2",
 231             "nb_series_per_chunk","nb_items_clust1","ncores_clust",
 232             "sep","nbytes","endian","verbose","parll")
 233         if (WER=="mix" && ntasks>1)
 234             varlist = c(varlist, "medoids_file")
 235         parallel::clusterExport(cl, varlist, envir = environment())
 236     }
 237
 238     # This function achieves one complete clustering task, divided in stage 1 + stage 2.
 239     # stage 1: n indices  --> clusteringTask1(...) --> K1 medoids
 240     # stage 2: K1 medoids --> clusteringTask2(...) --> K2 medoids,
 241     # where n = N / ntasks, N being the total number of curves.
 242     runTwoStepClustering = function(inds)
 243     {
 244         # When running in parallel, the environment is blank: we need to load the required
 245         # packages, and pass useful variables.
 246         if (parll && ntasks>1)
 247             require("epclust", quietly=TRUE)
 248         indices_medoids = clusteringTask1(
 249             inds, getContribs, K1, algoClust1, nb_series_per_chunk, ncores_clust, verbose, parll)
 250         if (WER=="mix" && ntasks>1)
 251         {
 252             if (parll)
 253                 require("bigmemory", quietly=TRUE)
 254             medoids1 = bigmemory::as.big.matrix( getSeries(indices_medoids) )
 255             medoids2 = clusteringTask2(medoids1, K2, algoClust2, getSeries, nb_curves,
 256                 nb_series_per_chunk, nbytes, endian, ncores_clust, verbose, parll)
 257             binarize(medoids2, medoids_file, nb_series_per_chunk, sep, nbytes, endian)
 258             return (vector("integer",0))
 259         }
 260         indices_medoids
 261     }
 262
 263     # Synchrones (medoids) need to be stored only if WER=="mix"; indeed in this case, every
 264     # task output is a set of new (medoids) curves. If WER=="end" however, output is just a
 265     # set of indices, representing some initial series.
 266     if (WER=="mix" && ntasks>1)
 267         {medoids_file = ".medoids.bin" ; unlink(medoids_file)}
 268
 269     if (verbose)
 270     {
 271         message = paste("...Run ",ntasks," x stage 1", sep="")
 272         if (WER=="mix")
 273             message = paste(message," + stage 2", sep="")
 274         cat(paste(message,"\n", sep=""))
 275     }
 276
 277     # As explained above, indices will be assigned to ntasks*K1 medoids indices [if WER=="end"],
 278     # or nothing (empty vector) if WER=="mix"; in this case, synchrones are stored in a file.
 279     indices <-
 280         if (parll && ntasks>1)
 281             unlist( parallel::parLapply(cl, indices_tasks, runTwoStepClustering) )
 282         else
 283             unlist( lapply(indices_tasks, runTwoStepClustering) )
 284     if (parll && ntasks>1)
 285         parallel::stopCluster(cl)
 286
 287     # Right before the final stage, two situations are possible:
 288     #  a. data to be processed now sit in a binary format in medoids_file (if WER=="mix")
 289     #  b. data still is the initial set of curves, referenced by the ntasks*K1 indices
 290     # So, the function getSeries() will potentially change. However, computeSynchrones()
 291     # requires a function retrieving the initial series. Thus, the next line saves future
 292     # conditional instructions.
 293     getRefSeries = getSeries
 294
 295     if (WER=="mix" && ntasks>1)
 296     {
 297         indices = seq_len(ntasks*K2)
 298         # Now series (synchrones) must be retrieved from medoids_file
 299         getSeries = function(inds) getDataInFile(inds, medoids_file, nbytes, endian)
 300         # Contributions must be re-computed
 301         unlink(contribs_file)
 302         index = 1
 303         if (verbose)
 304             cat("...Serialize contributions computed on synchrones\n")
 305         ignored = binarizeTransform(getSeries,
 306             function(series) curvesToContribs(series, wav_filt, contrib_type),
 307             contribs_file, nb_series_per_chunk, nbytes, endian)
 308     }
 309
 310     # Run step2 on resulting indices or series (from file)
 311     if (verbose)
 312         cat("...Run final // stage 1 + stage 2\n")
 313     indices_medoids = clusteringTask1(indices, getContribs, K1, algoClust1,
 314         nb_series_per_chunk, ncores_tasks*ncores_clust, verbose, parll)
 315     medoids1 = bigmemory::as.big.matrix( getSeries(indices_medoids) )
 316     medoids2 = clusteringTask2(medoids1, K2, algoClust2, getRefSeries, nb_curves,
 317         nb_series_per_chunk, nbytes, endian, ncores_tasks*ncores_clust, verbose, parll)
 318
 319     # Cleanup: remove temporary binary files
 320     tryCatch(
 321         {unlink(series_file); unlink(contribs_file); unlink(medoids_file)},
 322         error = function(e) {})
 323
 324     # Return medoids as a standard matrix, since K2 series have to fit in RAM
 325     # (clustering algorithm 1 takes K1 > K2 of them as input)
 326     medoids2[,]
 327 }
 328
 329 #' curvesToContribs
 330 #'
 331 #' Compute the discrete wavelet coefficients for each series, and aggregate them in
 332 #' energy contribution across scales as described in https://arxiv.org/abs/1101.4744v2
 333 #'
 334 #' @param series [big.]matrix of series (in columns), of size L x n
 335 #' @inheritParams claws
 336 #'
 337 #' @return A [big.]matrix of size log(L) x n containing contributions in columns
 338 #'
 339 #' @export
 340 curvesToContribs = function(series, wav_filt, contrib_type, coin=FALSE)
 341 {
 342     series = as.matrix(series) #1D serie could occur
 343     L = nrow(series)
 344     D = ceiling( log2(L) )
 345     # Series are interpolated to all have length 2^D
 346     nb_sample_points = 2^D
 347     apply(series, 2, function(x) {
 348         interpolated_curve = spline(1:L, x, n=nb_sample_points)$y
 349         W = wavelets::dwt(interpolated_curve, filter=wav_filt, D)@W
 350         # Compute the sum of squared discrete wavelet coefficients, for each scale
 351         nrj = rev( sapply( W, function(v) ( sqrt( sum(v^2) ) ) ) )
 352         if (contrib_type!="absolute")
 353             nrj = nrj / sum(nrj)
 354         if (contrib_type=="logit")
 355             nrj = - log(1 - nrj)
 356         nrj
 357     })
 358 }
 359
 360 # Check integer arguments with functional conditions
 361 .toInteger <- function(x, condition)
 362 {
 363     errWarn <- function(ignored)
 364         paste("Cannot convert argument' ",substitute(x),"' to integer", sep="")
 365     if (!is.integer(x))
 366         tryCatch({x = as.integer(x)[1]; if (is.na(x)) stop()},
 367             warning = errWarn, error = errWarn)
 368     if (!condition(x))
 369     {
 370         stop(paste("Argument '",substitute(x),
 371             "' does not verify condition ",body(condition), sep=""))
 372     }
 373     x
 374 }
 375
 376 # Check logical arguments
 377 .toLogical <- function(x)
 378 {
 379     errWarn <- function(ignored)
 380         paste("Cannot convert argument' ",substitute(x),"' to logical", sep="")
 381     if (!is.logical(x))
 382         tryCatch({x = as.logical(x)[1]; if (is.na(x)) stop()},
 383             warning = errWarn, error = errWarn)
 384     x
 385 }