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 matrix: each line contains all the values for one time-serie, ordered by time
  11 #'     \item connection: any R connection object (e.g. a file) providing lines as described above
  12 #'     \item function: a custom way to retrieve the curves; it has only one argument:
  13 #'       the indices of the series to be retrieved. See examples
  14 #'   }
  15 #' @inheritParams clustering
  16 #' @param K1 Number of super-consumers to be found after stage 1 (K1 << N)
  17 #' @param K2 Number of clusters to be found after stage 2 (K2 << K1)
  18 #' @param wf Wavelet transform filter; see ?wavelets::wt.filter
  19 #' @param ctype Type of contribution: "relative" or "absolute" (or any prefix)
  20 #' @param WER "end" to apply stage 2 after stage 1 has fully iterated, or "mix" to apply stage 2
  21 #'   at the end of each task
  22 #' @param random TRUE (default) for random chunks repartition
  23 #' @param ntasks Number of tasks (parallel iterations to obtain K1 medoids); default: 1.
  24 #'   Note: ntasks << N, so that N is "roughly divisible" by N (number of series)
  25 #' @param ncores_tasks "MPI" number of parallel tasks (1 to disable: sequential tasks)
  26 #' @param ncores_clust "OpenMP" number of parallel clusterings in one task
  27 #' @param nb_series_per_chunk (~Maximum) number of series in each group, inside a task
  28 #' @param min_series_per_chunk Minimum number of series in each group
  29 #' @param sep Separator in CSV input file (if any provided)
  30 #' @param nbytes Number of bytes to serialize a floating-point number; 4 or 8
  31 #' @param endian Endianness to use for (de)serialization. Use "little" or "big" for portability
  32 #' @param verbose Level of verbosity (0/FALSE for nothing or 1/TRUE for all; devel stage)
  33 #' @param parll TRUE to fully parallelize; otherwise run sequentially (debug, comparison)
  34 #'
  35 #' @return A matrix of the final medoids curves (K2) in rows
  36 #'
  37 #' @examples
  38 #' \dontrun{
  39 #' # WER distances computations are a bit too long for CRAN (for now)
  40 #'
  41 #' # Random series around cos(x,2x,3x)/sin(x,2x,3x)
  42 #' x = seq(0,500,0.05)
  43 #' L = length(x) #10001
  44 #' ref_series = matrix( c(cos(x), cos(2*x), cos(3*x), sin(x), sin(2*x), sin(3*x)),
  45 #'   byrow=TRUE, ncol=L )
  46 #' library(wmtsa)
  47 #' series = do.call( rbind, lapply( 1:6, function(i)
  48 #'   do.call(rbind, wmtsa::wavBootstrap(ref_series[i,], n.realization=400)) ) )
  49 #' #dim(series) #c(2400,10001)
  50 #' medoids_ascii = claws(series, K1=60, K2=6, "d8", "rel", nb_series_per_chunk=500)
  51 #'
  52 #' # Same example, from CSV file
  53 #' csv_file = "/tmp/epclust_series.csv"
  54 #' write.table(series, csv_file, sep=",", row.names=FALSE, col.names=FALSE)
  55 #' medoids_csv = claws(csv_file, K1=60, K2=6, "d8", "rel", nb_series_per_chunk=500)
  56 #'
  57 #' # Same example, from binary file
  58 #' bin_file = "/tmp/epclust_series.bin"
  59 #' nbytes = 8
  60 #' endian = "little"
  61 #' epclust::binarize(csv_file, bin_file, 500, nbytes, endian)
  62 #' getSeries = function(indices) getDataInFile(indices, bin_file, nbytes, endian)
  63 #' medoids_bin = claws(getSeries, K1=60, K2=6, "d8", "rel", nb_series_per_chunk=500)
  64 #' unlink(csv_file)
  65 #' unlink(bin_file)
  66 #'
  67 #' # Same example, from SQLite database
  68 #' library(DBI)
  69 #' series_db <- dbConnect(RSQLite::SQLite(), "file::memory:")
  70 #' # Prepare data.frame in DB-format
  71 #' n = nrow(series)
  72 #' time_values = data.frame(
  73 #'   id = rep(1:n,each=L),
  74 #'   time = rep( as.POSIXct(1800*(0:n),"GMT",origin="2001-01-01"), L ),
  75 #'   value = as.double(t(series)) )
  76 #' dbWriteTable(series_db, "times_values", times_values)
  77 #' # Fill associative array, map index to identifier
  78 #' indexToID_inDB <- as.character(
  79 #'   dbGetQuery(series_db, 'SELECT DISTINCT id FROM time_values')[,"id"] )
  80 #' getSeries = function(indices) {
  81 #'   request = "SELECT id,value FROM times_values WHERE id in ("
  82 #'   for (i in indices)
  83 #'     request = paste(request, i, ",", sep="")
  84 #'   request = paste(request, ")", sep="")
  85 #'   df_series = dbGetQuery(series_db, request)
  86 #'   # Assume that all series share same length at this stage
  87 #'   ts_length = sum(df_series[,"id"] == df_series[1,"id"])
  88 #'   t( as.matrix(df_series[,"value"], nrow=ts_length) )
  89 #' }
  90 #' medoids_db = claws(getSeries, K1=60, K2=6, "d8", "rel", nb_series_per_chunk=500)
  91 #' dbDisconnect(series_db)
  92 #'
  93 #' # All computed medoids should be the same:
  94 #' digest::sha1(medoids_ascii)
  95 #' digest::sha1(medoids_csv)
  96 #' digest::sha1(medoids_bin)
  97 #' digest::sha1(medoids_db)
  98 #' }
  99 #' @export
 100 claws = function(getSeries, K1, K2,
 101     wf,ctype, #stage 1
 102     WER="end", #stage 2
 103     random=TRUE, #randomize series order?
 104     ntasks=1, ncores_tasks=1, ncores_clust=4, #control parallelism
 105     nb_series_per_chunk=50*K1, min_series_per_chunk=5*K1, #chunk size
 106     sep=",", #ASCII input separator
 107     nbytes=4, endian=.Platform$endian, #serialization (write,read)
 108     verbose=FALSE, parll=TRUE)
 109 {
 110     # Check/transform arguments
 111     if (!is.matrix(getSeries) && !bigmemory::is.big.matrix(getSeries)
 112         && !is.function(getSeries)
 113         && !methods::is(getSeries,"connection") && !is.character(getSeries))
 114     {
 115         stop("'getSeries': [big]matrix, function, file or valid connection (no NA)")
 116     }
 117     K1 = .toInteger(K1, function(x) x>=2)
 118     K2 = .toInteger(K2, function(x) x>=2)
 119     if (!is.logical(random))
 120         stop("'random': logical")
 121     tryCatch(
 122         {ignored <- wavelets::wt.filter(wf)},
 123         error = function(e) stop("Invalid wavelet filter; see ?wavelets::wt.filter"))
 124     if (WER!="end" && WER!="mix")
 125         stop("WER takes values in {'end','mix'}")
 126     ntasks = .toInteger(ntasks, function(x) x>=1)
 127     ncores_tasks = .toInteger(ncores_tasks, function(x) x>=1)
 128     ncores_clust = .toInteger(ncores_clust, function(x) x>=1)
 129     nb_series_per_chunk = .toInteger(nb_series_per_chunk, function(x) x>=K1)
 130     min_series_per_chunk = .toInteger(K1, function(x) x>=K1 && x<=nb_series_per_chunk)
 131     if (!is.character(sep))
 132         stop("'sep': character")
 133     nbytes = .toInteger(nbytes, function(x) x==4 || x==8)
 134
 135     # Serialize series if required, to always use a function
 136     bin_dir = ".epclust_bin/"
 137     dir.create(bin_dir, showWarnings=FALSE, mode="0755")
 138     if (!is.function(getSeries))
 139     {
 140         if (verbose)
 141             cat("...Serialize time-series\n")
 142         series_file = paste(bin_dir,"data",sep="") ; unlink(series_file)
 143         binarize(getSeries, series_file, nb_series_per_chunk, sep, nbytes, endian)
 144         getSeries = function(inds) getDataInFile(inds, series_file, nbytes, endian)
 145     }
 146
 147     # Serialize all computed wavelets contributions into a file
 148     contribs_file = paste(bin_dir,"contribs",sep="") ; unlink(contribs_file)
 149     index = 1
 150     nb_curves = 0
 151     if (verbose)
 152         cat("...Compute contributions and serialize them\n")
 153     nb_curves = binarizeTransform(getSeries,
 154         function(series) curvesToContribs(series, wf, ctype),
 155         contribs_file, nb_series_per_chunk, nbytes, endian)
 156     getContribs = function(indices) getDataInFile(indices, contribs_file, nbytes, endian)
 157
 158     if (nb_curves < min_series_per_chunk)
 159         stop("Not enough data: less rows than min_series_per_chunk!")
 160     nb_series_per_task = round(nb_curves / ntasks)
 161     if (nb_series_per_task < min_series_per_chunk)
 162         stop("Too many tasks: less series in one task than min_series_per_chunk!")
 163
 164     runTwoStepClustering = function(inds)
 165     {
 166         if (parll)
 167             require("epclust", quietly=TRUE)
 168         indices_medoids = clusteringTask1(
 169             inds, getContribs, K1, nb_series_per_chunk, ncores_clust, verbose, parll)
 170         if (WER=="mix")
 171         {
 172
 173
 174
 175
 176 #TODO: getSeries(indices_medoids) BAD ; il faudrait une big.matrix de medoids en entree
 177             #OK en RAM il y en aura 1000 (donc 1000*K1*17519... OK)
 178             #...mais du coup chaque process ne re-dupliquera pas medoids
 179
 180
 181             medoids2 = computeClusters2(getSeries(indices_medoids),
 182                 K2, getSeries, nb_curves, nb_series_per_chunk, ncores_clust, verbose, parll)
 183             binarize(medoids2, synchrones_file, nb_series_per_chunk, sep, nbytes, endian)
 184             return (vector("integer",0))
 185         }
 186         indices_medoids
 187     }
 188
 189     # Cluster contributions in parallel (by nb_series_per_chunk)
 190     indices_all = if (random) sample(nb_curves) else seq_len(nb_curves)
 191     indices_tasks = lapply(seq_len(ntasks), function(i) {
 192         upper_bound = ifelse( i<ntasks, min(nb_series_per_task*i,nb_curves), nb_curves )
 193         indices_all[((i-1)*nb_series_per_task+1):upper_bound]
 194     })
 195     if (verbose)
 196         cat(paste("...Run ",ntasks," x stage 1 in parallel\n",sep=""))
 197     if (WER=="mix")
 198         {synchrones_file = paste(bin_dir,"synchrones",sep="") ; unlink(synchrones_file)}
 199     if (parll)
 200     {
 201         cl = parallel::makeCluster(ncores_tasks)
 202         varlist = c("getSeries","getContribs","K1","K2","verbose","parll",
 203             "nb_series_per_chunk","ncores_clust","sep","nbytes","endian")
 204         if (WER=="mix")
 205             varlist = c(varlist, "synchrones_file")
 206         parallel::clusterExport(cl, varlist=varlist, envir = environment())
 207     }
 208
 209     # 1000*K1 indices [if WER=="end"], or empty vector [if WER=="mix"] --> series on file
 210     if (parll)
 211         indices = unlist( parallel::parLapply(cl, indices_tasks, runTwoStepClustering) )
 212     else
 213         indices = unlist( lapply(indices_tasks, runTwoStepClustering) )
 214     if (parll)
 215         parallel::stopCluster(cl)
 216
 217     getRefSeries = getSeries
 218     if (WER=="mix")
 219     {
 220         indices = seq_len(ntasks*K2)
 221         #Now series must be retrieved from synchrones_file
 222         getSeries = function(inds) getDataInFile(inds, synchrones_file, nbytes, endian)
 223         #Contributions must be re-computed
 224         unlink(contribs_file)
 225         index = 1
 226         if (verbose)
 227             cat("...Serialize contributions computed on synchrones\n")
 228         ignored = binarizeTransform(getSeries,
 229             function(series) curvesToContribs(series, wf, ctype),
 230             contribs_file, nb_series_per_chunk, nbytes, endian)
 231     }
 232
 233
 234
 235 #TODO: if ntasks==1, c'est deja terminé
 236
 237     # Run step2 on resulting indices or series (from file)
 238     if (verbose)
 239         cat("...Run final // stage 1 + stage 2\n")
 240     indices_medoids = clusteringTask1(
 241         indices, getContribs, K1, nb_series_per_chunk, ncores_tasks*ncores_clust, verbose, parll)
 242     medoids = computeClusters2(getSeries(indices_medoids), K2,
 243         getRefSeries, nb_curves, nb_series_per_chunk, ncores_tasks*ncores_clust, verbose, parll)
 244
 245     # Cleanup
 246     unlink(bin_dir, recursive=TRUE)
 247
 248     medoids
 249 }
 250
 251 #' curvesToContribs
 252 #'
 253 #' Compute the discrete wavelet coefficients for each series, and aggregate them in
 254 #' energy contribution across scales as described in https://arxiv.org/abs/1101.4744v2
 255 #'
 256 #' @param series Matrix of series (in rows), of size n x L
 257 #' @inheritParams claws
 258 #'
 259 #' @return A matrix of size n x log(L) containing contributions in rows
 260 #'
 261 #' @export
 262 curvesToContribs = function(series, wf, ctype)
 263 {
 264     L = length(series[1,])
 265     D = ceiling( log2(L) )
 266     nb_sample_points = 2^D
 267     cont_types = c("relative","absolute")
 268     ctype = cont_types[ pmatch(ctype,cont_types) ]
 269     t( apply(series, 1, function(x) {
 270         interpolated_curve = spline(1:L, x, n=nb_sample_points)$y
 271         W = wavelets::dwt(interpolated_curve, filter=wf, D)@W
 272         nrj = rev( sapply( W, function(v) ( sqrt( sum(v^2) ) ) ) )
 273         if (ctype=="relative") nrj / sum(nrj) else nrj
 274     }) )
 275 }
 276
 277 # Check integer arguments with functional conditions
 278 .toInteger <- function(x, condition)
 279 {
 280     if (!is.integer(x))
 281         tryCatch(
 282             {x = as.integer(x)[1]},
 283             error = function(e) paste("Cannot convert argument",substitute(x),"to integer")
 284         )
 285     if (!condition(x))
 286         stop(paste("Argument",substitute(x),"does not verify condition",body(condition)))
 287     x
 288 }