epclust/R/main.R

   1 #' CLAWS: CLustering with wAvelets and Wer distanceS
   2 #'
   3 #' Groups electricity power curves (or any series of similar nature) by applying PAM
   4 #' algorithm in parallel to chunks of size \code{nb_series_per_chunk}. Input series
   5 #' must be sampled on the same time grid, no missing values.
   6 #'
   7 #' @param getSeries Access to the (time-)series, which can be of one of the three
   8 #'   following types:
   9 #'   \itemize{
  10 #'     \item [big.]matrix: each line contains all the values for one time-serie, ordered by time
  11 #'     \item connection: any R connection object providing lines as described above
  12 #'     \item character: name of a CSV file containing series in rows (no header)
  13 #'     \item function: a custom way to retrieve the curves; it has only one argument:
  14 #'       the indices of the series to be retrieved. See examples
  15 #'   }
  16 #' @inheritParams clustering
  17 #' @param K1 Number of super-consumers to be found after stage 1 (K1 << N)
  18 #' @param K2 Number of clusters to be found after stage 2 (K2 << K1)
  19 #' @param wf Wavelet transform filter; see ?wavelets::wt.filter
  20 #' @param ctype Type of contribution: "relative" or "absolute" (or any prefix)
  21 #' @param WER "end" to apply stage 2 after stage 1 has fully iterated, or "mix" to apply stage 2
  22 #'   at the end of each task
  23 #' @param random TRUE (default) for random chunks repartition
  24 #' @param ntasks Number of tasks (parallel iterations to obtain K1 medoids); default: 1.
  25 #'   Note: ntasks << N, so that N is "roughly divisible" by N (number of series)
  26 #' @param ncores_tasks "MPI" number of parallel tasks (1 to disable: sequential tasks)
  27 #' @param ncores_clust "OpenMP" number of parallel clusterings in one task
  28 #' @param nb_series_per_chunk (~Maximum) number of series in each group, inside a task
  29 #' @param min_series_per_chunk Minimum number of series in each group
  30 #' @param sep Separator in CSV input file (if any provided)
  31 #' @param nbytes Number of bytes to serialize a floating-point number; 4 or 8
  32 #' @param endian Endianness to use for (de)serialization. Use "little" or "big" for portability
  33 #' @param verbose Level of verbosity (0/FALSE for nothing or 1/TRUE for all; devel stage)
  34 #' @param parll TRUE to fully parallelize; otherwise run sequentially (debug, comparison)
  35 #'
  36 #' @return A big.matrix of the final medoids curves (K2) in rows
  37 #'
  38 #' @examples
  39 #' \dontrun{
  40 #' # WER distances computations are a bit too long for CRAN (for now)
  41 #'
  42 #' # Random series around cos(x,2x,3x)/sin(x,2x,3x)
  43 #' x = seq(0,500,0.05)
  44 #' L = length(x) #10001
  45 #' ref_series = matrix( c(cos(x), cos(2*x), cos(3*x), sin(x), sin(2*x), sin(3*x)),
  46 #'   byrow=TRUE, ncol=L )
  47 #' library(wmtsa)
  48 #' series = do.call( rbind, lapply( 1:6, function(i)
  49 #'   do.call(rbind, wmtsa::wavBootstrap(ref_series[i,], n.realization=400)) ) )
  50 #' #dim(series) #c(2400,10001)
  51 #' medoids_ascii = claws(series, K1=60, K2=6, "d8", "rel", nb_series_per_chunk=500)
  52 #'
  53 #' # Same example, from CSV file
  54 #' csv_file = "/tmp/epclust_series.csv"
  55 #' write.table(series, csv_file, sep=",", row.names=FALSE, col.names=FALSE)
  56 #' medoids_csv = claws(csv_file, K1=60, K2=6, "d8", "rel", nb_series_per_chunk=500)
  57 #'
  58 #' # Same example, from binary file
  59 #' bin_file = "/tmp/epclust_series.bin"
  60 #' nbytes = 8
  61 #' endian = "little"
  62 #' epclust::binarize(csv_file, bin_file, 500, nbytes, endian)
  63 #' getSeries = function(indices) getDataInFile(indices, bin_file, nbytes, endian)
  64 #' medoids_bin = claws(getSeries, K1=60, K2=6, "d8", "rel", nb_series_per_chunk=500)
  65 #' unlink(csv_file)
  66 #' unlink(bin_file)
  67 #'
  68 #' # Same example, from SQLite database
  69 #' library(DBI)
  70 #' series_db <- dbConnect(RSQLite::SQLite(), "file::memory:")
  71 #' # Prepare data.frame in DB-format
  72 #' n = nrow(series)
  73 #' time_values = data.frame(
  74 #'   id = rep(1:n,each=L),
  75 #'   time = rep( as.POSIXct(1800*(0:n),"GMT",origin="2001-01-01"), L ),
  76 #'   value = as.double(t(series)) )
  77 #' dbWriteTable(series_db, "times_values", times_values)
  78 #' # Fill associative array, map index to identifier
  79 #' indexToID_inDB <- as.character(
  80 #'   dbGetQuery(series_db, 'SELECT DISTINCT id FROM time_values')[,"id"] )
  81 #' getSeries = function(indices) {
  82 #'   request = "SELECT id,value FROM times_values WHERE id in ("
  83 #'   for (i in indices)
  84 #'     request = paste(request, i, ",", sep="")
  85 #'   request = paste(request, ")", sep="")
  86 #'   df_series = dbGetQuery(series_db, request)
  87 #'   # Assume that all series share same length at this stage
  88 #'   ts_length = sum(df_series[,"id"] == df_series[1,"id"])
  89 #'   t( as.matrix(df_series[,"value"], nrow=ts_length) )
  90 #' }
  91 #' medoids_db = claws(getSeries, K1=60, K2=6, "d8", "rel", nb_series_per_chunk=500)
  92 #' dbDisconnect(series_db)
  93 #'
  94 #' # All computed medoids should be the same:
  95 #' digest::sha1(medoids_ascii)
  96 #' digest::sha1(medoids_csv)
  97 #' digest::sha1(medoids_bin)
  98 #' digest::sha1(medoids_db)
  99 #' }
 100 #' @export
 101 claws = function(getSeries, K1, K2,
 102     wf,ctype, #stage 1
 103     WER="end", #stage 2
 104     random=TRUE, #randomize series order?
 105     ntasks=1, ncores_tasks=1, ncores_clust=4, #control parallelism
 106     nb_series_per_chunk=50*K1, min_series_per_chunk=5*K1, #chunk size
 107     sep=",", #ASCII input separator
 108     nbytes=4, endian=.Platform$endian, #serialization (write,read)
 109     verbose=FALSE, parll=TRUE)
 110 {
 111     # Check/transform arguments
 112     if (!is.matrix(getSeries) && !bigmemory::is.big.matrix(getSeries)
 113         && !is.function(getSeries)
 114         && !methods::is(getSeries,"connection") && !is.character(getSeries))
 115     {
 116         stop("'getSeries': [big]matrix, function, file or valid connection (no NA)")
 117     }
 118     K1 = .toInteger(K1, function(x) x>=2)
 119     K2 = .toInteger(K2, function(x) x>=2)
 120     if (!is.logical(random))
 121         stop("'random': logical")
 122     tryCatch(
 123         {ignored <- wavelets::wt.filter(wf)},
 124         error = function(e) stop("Invalid wavelet filter; see ?wavelets::wt.filter"))
 125     if (WER!="end" && WER!="mix")
 126         stop("WER takes values in {'end','mix'}")
 127     ntasks = .toInteger(ntasks, function(x) x>=1)
 128     ncores_tasks = .toInteger(ncores_tasks, function(x) x>=1)
 129     ncores_clust = .toInteger(ncores_clust, function(x) x>=1)
 130     nb_series_per_chunk = .toInteger(nb_series_per_chunk, function(x) x>=K1)
 131     min_series_per_chunk = .toInteger(K1, function(x) x>=K1 && x<=nb_series_per_chunk)
 132     if (!is.character(sep))
 133         stop("'sep': character")
 134     nbytes = .toInteger(nbytes, function(x) x==4 || x==8)
 135
 136     # Serialize series if required, to always use a function
 137     bin_dir = ".epclust_bin/"
 138     dir.create(bin_dir, showWarnings=FALSE, mode="0755")
 139     if (!is.function(getSeries))
 140     {
 141         if (verbose)
 142             cat("...Serialize time-series\n")
 143         series_file = paste(bin_dir,"data",sep="") ; unlink(series_file)
 144         binarize(getSeries, series_file, nb_series_per_chunk, sep, nbytes, endian)
 145         getSeries = function(inds) getDataInFile(inds, series_file, nbytes, endian)
 146     }
 147
 148     # Serialize all computed wavelets contributions into a file
 149     contribs_file = paste(bin_dir,"contribs",sep="") ; unlink(contribs_file)
 150     index = 1
 151     nb_curves = 0
 152     if (verbose)
 153         cat("...Compute contributions and serialize them\n")
 154     nb_curves = binarizeTransform(getSeries,
 155         function(series) curvesToContribs(series, wf, ctype),
 156         contribs_file, nb_series_per_chunk, nbytes, endian)
 157     getContribs = function(indices) getDataInFile(indices, contribs_file, nbytes, endian)
 158
 159     if (nb_curves < min_series_per_chunk)
 160         stop("Not enough data: less rows than min_series_per_chunk!")
 161     nb_series_per_task = round(nb_curves / ntasks)
 162     if (nb_series_per_task < min_series_per_chunk)
 163         stop("Too many tasks: less series in one task than min_series_per_chunk!")
 164
 165     runTwoStepClustering = function(inds)
 166     {
 167         if (parll && ntasks>1)
 168             require("epclust", quietly=TRUE)
 169         indices_medoids = clusteringTask1(
 170             inds, getContribs, K1, nb_series_per_chunk, ncores_clust, verbose, parll)
 171         if (WER=="mix")
 172         {
 173             require("bigmemory", quietly=TRUE)
 174             medoids1 = bigmemory::as.big.matrix( getSeries(indices_medoids) )
 175             medoids2 = clusteringTask2(medoids1,
 176                 K2, getSeries, nb_curves, nb_series_per_chunk, ncores_clust, verbose, parll)
 177             binarize(medoids2, synchrones_file, nb_series_per_chunk, sep, nbytes, endian)
 178             return (vector("integer",0))
 179         }
 180         indices_medoids
 181     }
 182
 183     # Cluster contributions in parallel (by nb_series_per_chunk)
 184     indices_all = if (random) sample(nb_curves) else seq_len(nb_curves)
 185     indices_tasks = lapply(seq_len(ntasks), function(i) {
 186         upper_bound = ifelse( i<ntasks, min(nb_series_per_task*i,nb_curves), nb_curves )
 187         indices_all[((i-1)*nb_series_per_task+1):upper_bound]
 188     })
 189     if (verbose)
 190     {
 191         message = paste("...Run ",ntasks," x stage 1", sep="")
 192         if (WER=="mix")
 193             message = paste(message," + stage 2", sep="")
 194         cat(paste(message,"\n", sep=""))
 195     }
 196     if (WER=="mix")
 197         {synchrones_file = paste(bin_dir,"synchrones",sep="") ; unlink(synchrones_file)}
 198     if (parll && ntasks>1)
 199     {
 200         cl = parallel::makeCluster(ncores_tasks)
 201         varlist = c("getSeries","getContribs","K1","K2","verbose","parll",
 202             "nb_series_per_chunk","ntasks","ncores_clust","sep","nbytes","endian")
 203         if (WER=="mix")
 204             varlist = c(varlist, "synchrones_file")
 205         parallel::clusterExport(cl, varlist=varlist, envir = environment())
 206     }
 207
 208     # 1000*K1 indices [if WER=="end"], or empty vector [if WER=="mix"] --> series on file
 209     if (parll && ntasks>1)
 210         indices = unlist( parallel::parLapply(cl, indices_tasks, runTwoStepClustering) )
 211     else
 212         indices = unlist( lapply(indices_tasks, runTwoStepClustering) )
 213     if (parll && ntasks>1)
 214         parallel::stopCluster(cl)
 215
 216     getRefSeries = getSeries
 217     if (WER=="mix")
 218     {
 219         indices = seq_len(ntasks*K2)
 220         #Now series must be retrieved from synchrones_file
 221         getSeries = function(inds) getDataInFile(inds, synchrones_file, nbytes, endian)
 222         #Contributions must be re-computed
 223         unlink(contribs_file)
 224         index = 1
 225         if (verbose)
 226             cat("...Serialize contributions computed on synchrones\n")
 227         ignored = binarizeTransform(getSeries,
 228             function(series) curvesToContribs(series, wf, ctype),
 229             contribs_file, nb_series_per_chunk, nbytes, endian)
 230     }
 231
 232     # Run step2 on resulting indices or series (from file)
 233     if (verbose)
 234         cat("...Run final // stage 1 + stage 2\n")
 235     indices_medoids = clusteringTask1(
 236         indices, getContribs, K1, nb_series_per_chunk, ncores_tasks*ncores_clust, verbose, parll)
 237     medoids1 = bigmemory::as.big.matrix( getSeries(indices_medoids) )
 238     medoids2 = clusteringTask2(medoids1, K2,
 239         getRefSeries, nb_curves, nb_series_per_chunk, ncores_tasks*ncores_clust, verbose, parll)
 240
 241     # Cleanup
 242     unlink(bin_dir, recursive=TRUE)
 243
 244     medoids2
 245 }
 246
 247 #' curvesToContribs
 248 #'
 249 #' Compute the discrete wavelet coefficients for each series, and aggregate them in
 250 #' energy contribution across scales as described in https://arxiv.org/abs/1101.4744v2
 251 #'
 252 #' @param series Matrix of series (in rows), of size n x L
 253 #' @inheritParams claws
 254 #'
 255 #' @return A matrix of size n x log(L) containing contributions in rows
 256 #'
 257 #' @export
 258 curvesToContribs = function(series, wf, ctype)
 259 {
 260     L = length(series[1,])
 261     D = ceiling( log2(L) )
 262     nb_sample_points = 2^D
 263     cont_types = c("relative","absolute")
 264     ctype = cont_types[ pmatch(ctype,cont_types) ]
 265     t( apply(series, 1, function(x) {
 266         interpolated_curve = spline(1:L, x, n=nb_sample_points)$y
 267         W = wavelets::dwt(interpolated_curve, filter=wf, D)@W
 268         nrj = rev( sapply( W, function(v) ( sqrt( sum(v^2) ) ) ) )
 269         if (ctype=="relative") nrj / sum(nrj) else nrj
 270     }) )
 271 }
 272
 273 # Check integer arguments with functional conditions
 274 .toInteger <- function(x, condition)
 275 {
 276     if (!is.integer(x))
 277         tryCatch(
 278             {x = as.integer(x)[1]},
 279             error = function(e) paste("Cannot convert argument",substitute(x),"to integer")
 280         )
 281     if (!condition(x))
 282         stop(paste("Argument",substitute(x),"does not verify condition",body(condition)))
 283     x
 284 }