First commit
[epclust.git] / pkg / R / main.R
CommitLineData
b7cd987d
BA
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_clust}\cr
15#' -> K1 medoids indices
16#' \item optionally, if WER=="mix":\cr
17#' a. compute WER distances (K1xK1) between medoids\cr
18#' b. apply the 2nd clustering algorithm\cr
19#' -> K2 medoids indices
20#' }
21#' \item Launch a final task on the aggregated outputs of all previous tasks:
22#' ntasks*K1 if WER=="end", ntasks*K2 otherwise
23#' \item Compute synchrones (sum of series within each final group)
24#' }
25#'
26#' The main argument -- \code{series} -- has a quite misleading name, since it can be
27#' either a [big.]matrix, a CSV file, a connection or a user function to retrieve series.
28#' When \code{series} 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.
31#' WARNING: the return value must be a matrix (in columns), or NULL if no matches.
32#'
33#' Note: Since we don't make assumptions on initial data, there is a possibility that
34#' even when serialized, contributions do not fit in RAM. For example,
35#' 30e6 series of length 100,000 would lead to a +4Go contribution matrix. Therefore,
36#' it's safer to place these in (binary) files; that's what we do.
37#'
38#' @param series Access to the N (time-)series, which can be of one of the four
39#' following types:
40#' \itemize{
41#' \item [big.]matrix: each column contains the (time-ordered) values of one
42#' time-serie
43#' \item connection: any R connection object providing lines as described above
44#' \item character: name of a CSV file containing series in rows (no header)
45#' \item function: a custom way to retrieve the curves; it has only one argument:
46#' the indices of the series to be retrieved. See SQLite example
47#' }
48#' @param K1 Number of clusters to be found after stage 1 (K1 << N)
49#' @param K2 Number of clusters to be found after stage 2 (K2 << K1)
50#' @param nb_series_per_chunk Number of series to retrieve in one batch
51#' @param nb_items_clust Number of items in 1st clustering algorithm input
52#' @param algoClust1 Clustering algorithm for stage 1. A function which takes (data, K)
53#' as argument where data is a matrix in columns and K the desired number of clusters,
54#' and outputs K medoids ranks. Default: PAM.
55#' @param algoClust2 Clustering algorithm for stage 2. A function which takes (dists, K)
56#' as argument where dists is a matrix of distances and K the desired number of
57#' clusters, and outputs K medoids ranks. Default: PAM.
58#' @param wav_filt Wavelet transform filter, as a string "Family:FilterNumber"; see
59#' ?wavethresh::wd
60#' @param contrib_type Type of contribution: "relative", "logit" or "absolute" (any
61#' 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 smooth_lvl Smoothing level: odd integer, 1 == no smoothing.
65#' @param nvoice Number of voices within each octave for CWT computations
66#' @param random TRUE (default) for random chunks repartition
67#' @param ntasks Number of tasks (parallel iterations to obtain K1 [if WER=="end"]
68#' or K2 [if WER=="mix"] medoids); default: 1.\cr
69#' Note: ntasks << N (number of series), so that N is "roughly divisible" by ntasks
70#' @param ncores_tasks Number of parallel tasks ('1' == sequential tasks)
71#' @param ncores_clust Number of parallel clusterings in one task
72#' @param sep Separator in CSV input file (if any provided)
73#' @param nbytes 4 or 8 bytes to (de)serialize a floating-point number
74#' @param endian Endianness for (de)serialization: "little" or "big"
75#' @param verbose FALSE: nothing printed; TRUE: some execution traces
76#'
77#' @return A list:
78#' \itemize{
79#' \item medoids: matrix of the final K2 medoids curves
80#' \item ranks: corresponding indices in the dataset
81#' \item synchrones: sum of series within each final group
82#' }
83#'
84#' @references Clustering functional data using Wavelets [2013];
85#' A. Antoniadis, X. Brossat, J. Cugliari & J.-M. Poggi.
86#' Inter. J. of Wavelets, Multiresolution and Information Procesing,
87#' vol. 11, No 1, pp.1-30. doi:10.1142/S0219691313500033
88#'
89#' @examples
90#' \dontrun{
91#' # WER distances computations are too long for CRAN (for now)
92#' # Note: on this small example, sequential run is faster
93#'
94#' # Random series around cos(x,2x,3x)/sin(x,2x,3x)
95#' x <- seq(0,50,0.05)
96#' L <- length(x) #1001
97#' ref_series <- matrix( c(cos(x),cos(2*x),cos(3*x),sin(x),sin(2*x),sin(3*x)), ncol=6 )
98#' library(wmtsa)
99#' series <- do.call( cbind, lapply( 1:6, function(i)
100#' do.call(cbind, wmtsa::wavBootstrap(ref_series[,i], n.realization=40)) ) )
101#' # Mix series so that all groups are evenly spread
102#' permut <- (0:239)%%6 * 40 + (0:239)%/%6 + 1
103#' series = series[,permut]
104#' #dim(series) #c(240,1001)
105#' res_ascii <- claws(series, K1=30, K2=6, nb_series_per_chunk=500,
106#' nb_items_clust=100, random=FALSE, verbose=TRUE, ncores_clust=1)
107#'
108#' # Same example, from CSV file
109#' csv_file <- tempfile(pattern="epclust_series.csv_")
110#' write.table(t(series), csv_file, sep=",", row.names=FALSE, col.names=FALSE)
111#' res_csv <- claws(csv_file, 30, 6, 500, 100, random=FALSE, ncores_clust=1)
112#'
113#' # Same example, from binary file
114#' bin_file <- tempfile(pattern="epclust_series.bin_")
115#' nbytes <- 8
116#' endian <- "little"
117#' binarize(csv_file, bin_file, 500, ",", nbytes, endian)
118#' getSeries <- function(indices) getDataInFile(indices, bin_file, nbytes, endian)
119#' res_bin <- claws(getSeries, 30, 6, 500, 100, random=FALSE, ncores_clust=1)
120#' unlink(csv_file)
121#' unlink(bin_file)
122#'
123#' # Same example, from SQLite database
124#' library(DBI)
125#' series_db <- dbConnect(RSQLite::SQLite(), "file::memory:")
126#' # Prepare data.frame in DB-format
127#' n <- ncol(series)
128#' times_values <- data.frame(
129#' id = rep(1:n,each=L),
130#' time = rep( as.POSIXct(1800*(1:L),"GMT",origin="2001-01-01"), n ),
131#' value = as.double(series) )
132#' dbWriteTable(series_db, "times_values", times_values)
133#' # Fill associative array, map index to identifier
134#' indexToID_inDB <- as.character(
135#' dbGetQuery(series_db, 'SELECT DISTINCT id FROM times_values')[,"id"] )
136#' serie_length <- as.integer( dbGetQuery(series_db,
137#' paste("SELECT COUNT(*) FROM times_values WHERE id == ",indexToID_inDB[1],sep="")) )
138#' getSeries <- function(indices) {
139#' indices = indices[ indices <= length(indexToID_inDB) ]
140#' if (length(indices) == 0)
141#' return (NULL)
142#' request <- "SELECT id,value FROM times_values WHERE id in ("
143#' for (i in seq_along(indices)) {
144#' request <- paste(request, indexToID_inDB[ indices[i] ], sep="")
145#' if (i < length(indices))
146#' request <- paste(request, ",", sep="")
147#' }
148#' request <- paste(request, ")", sep="")
149#' df_series <- dbGetQuery(series_db, request)
150#' matrix(df_series[,"value"], nrow=serie_length)
151#' }
152#' res_db <- claws(getSeries, 30, 6, 500, 100, random=FALSE, ncores_clust=1)
153#' dbDisconnect(series_db)
154#'
155#' # All results should be equal:
156#' all(res_ascii$ranks == res_csv$ranks
157#' & res_ascii$ranks == res_bin$ranks
158#' & res_ascii$ranks == res_db$ranks)
159#' }
160#' @export
161claws <- function(series, K1, K2, nb_series_per_chunk, nb_items_clust=5*K1,
162 algoClust1=function(data,K) cluster::pam(t(data),K,diss=FALSE,pamonce=1)$id.med,
163 algoClust2=function(dists,K) cluster::pam(dists,K,diss=TRUE,pamonce=1)$id.med,
164 wav_filt="Coiflets:1", contrib_type="absolute", WER="end", smooth_lvl=3, nvoice=4,
165 random=TRUE, ntasks=1, ncores_tasks=1, ncores_clust=3, sep=",", nbytes=4,
166 endian=.Platform$endian, verbose=FALSE)
167{
168 # Check/transform arguments
169 if (!is.matrix(series) && !bigmemory::is.big.matrix(series)
170 && !is.function(series)
171 && !methods::is(series,"connection") && !is.character(series))
172 {
173 stop("'series': [big]matrix, function, file or valid connection (no NA)")
174 }
175 K1 <- .toInteger(K1, function(x) x>=2)
176 K2 <- .toInteger(K2, function(x) x>=2)
177 nb_series_per_chunk <- .toInteger(nb_series_per_chunk, function(x) x>=1)
178 nb_items_clust <- .toInteger(nb_items_clust, function(x) x>K1)
179 random <- .toLogical(random)
180 wav_filt <- strsplit(wav_filt, ':')[[1]]
181 wav_family <- wav_filt[1]
182 wav_number <- wav_filt[2]
183 tryCatch({ignored <- wavethresh::filter.select(wav_number, wav_family)},
184 error=function(e) stop("Invalid wavelet filter; see ?wavethresh::filter.select") )
185 ctypes <- c("relative","absolute","logit")
186 contrib_type <- ctypes[ pmatch(contrib_type,ctypes) ]
187 if (is.na(contrib_type))
188 stop("'contrib_type' in {'relative','absolute','logit'}")
189 if (WER!="end" && WER!="mix")
190 stop("'WER': in {'end','mix'}")
191 random <- .toLogical(random)
192 ntasks <- .toInteger(ntasks, function(x) x>=1)
193 ncores_tasks <- .toInteger(ncores_tasks, function(x) x>=1)
194 ncores_clust <- .toInteger(ncores_clust, function(x) x>=1)
195 if (!is.character(sep))
196 stop("'sep': character")
197 nbytes <- .toInteger(nbytes, function(x) x==4 || x==8)
198 verbose <- .toLogical(verbose)
199
200 # Binarize series if it is not a function; the aim is to always use a function,
201 # to uniformize treatments. An equally good alternative would be to use a file-backed
202 # bigmemory::big.matrix, but it would break the "all-is-function" pattern.
203 if (!is.function(series))
204 {
205 if (verbose)
206 cat("...Serialize time-series (or retrieve past binary file)\n")
207 series_file <- ".series.epclust.bin"
208 if (!file.exists(series_file))
209 binarize(series, series_file, nb_series_per_chunk, sep, nbytes, endian)
210 getSeries <- function(inds) getDataInFile(inds, series_file, nbytes, endian)
211 }
212 else
213 getSeries <- series
214
215 # Serialize all computed wavelets contributions into a file
216 contribs_file <- ".contribs.epclust.bin"
217 if (verbose)
218 cat("...Compute contributions and serialize them (or retrieve past binary file)\n")
219 if (!file.exists(contribs_file))
220 {
221 nb_curves <- binarizeTransform(getSeries,
222 function(curves) curvesToContribs(curves, wav_filt, contrib_type),
223 contribs_file, nb_series_per_chunk, nbytes, endian)
224 }
225 else
226 {
227 # TODO: duplicate from getDataInFile() in de_serialize.R
228 contribs_size <- file.info(contribs_file)$size #number of bytes in the file
229 contrib_length <- readBin(contribs_file, "integer", n=1, size=8, endian=endian)
230 nb_curves <- (contribs_size-8) / (nbytes*contrib_length)
231 }
232 getContribs <- function(indices) getDataInFile(indices, contribs_file, nbytes, endian)
233
234 # A few sanity checks: do not continue if too few data available.
235 if (nb_curves < K2)
236 stop("Not enough data: less series than final number of clusters")
237 nb_series_per_task <- round(nb_curves / ntasks)
238 if (nb_series_per_task < K2)
239 stop("Too many tasks: less series in one task than final number of clusters")
240
241 # Generate a random permutation of 1:N (if random==TRUE);
242 # otherwise just use arrival (storage) order.
243 indices_all <- if (random) sample(nb_curves) else seq_len(nb_curves)
244 # Split (all) indices into ntasks groups of ~same size
245 indices_tasks <- lapply(seq_len(ntasks), function(i) {
246 upper_bound <- ifelse( i<ntasks, min(nb_series_per_task*i,nb_curves), nb_curves )
247 indices_all[((i-1)*nb_series_per_task+1):upper_bound]
248 })
249
250 parll <- (ncores_tasks > 1)
251 if (parll && ntasks>1)
252 {
253 # Initialize parallel runs: outfile="" allow to output verbose traces in the console
254 # under Linux. All necessary variables are passed to the workers.
255 cl <-
256 if (verbose)
257 parallel::makeCluster(ncores_tasks, outfile="")
258 else
259 parallel::makeCluster(ncores_tasks)
260 varlist <- c("ncores_clust","verbose", #task 1 & 2
261 "K1","getContribs","algoClust1","nb_items_clust") #task 1
262 if (WER=="mix")
263 {
264 # Add variables for task 2
265 varlist <- c(varlist, "K2","getSeries","algoClust2","nb_series_per_chunk",
266 "smooth_lvl","nvoice","nbytes","endian")
267 }
268 parallel::clusterExport(cl, varlist, envir <- environment())
269 }
270
271 # This function achieves one complete clustering task, divided in stage 1 + stage 2.
272 # stage 1: n indices --> clusteringTask1(...) --> K1 medoids (indices)
273 # stage 2: K1 indices --> K1xK1 WER distances --> clusteringTask2(...) --> K2 medoids,
274 # where n == N / ntasks, N being the total number of curves.
275 runTwoStepClustering <- function(inds)
276 {
277 # When running in parallel, the environment is blank: we need to load the required
278 # packages, and pass useful variables.
279 if (parll && ntasks>1)
280 require("epclust", quietly=TRUE)
281 indices_medoids <- clusteringTask1(inds, getContribs, K1, algoClust1,
282 nb_items_clust, ncores_clust, verbose)
283 if (WER=="mix")
284 {
285 indices_medoids <- clusteringTask2(indices_medoids, getSeries, K2, algoClust2,
286 nb_series_per_chunk,smooth_lvl,nvoice,nbytes,endian,ncores_clust,verbose)
287 }
288 indices_medoids
289 }
290
291 if (verbose)
292 {
293 message <- paste("...Run ",ntasks," x stage 1", sep="")
294 if (WER=="mix")
295 message <- paste(message," + stage 2", sep="")
296 cat(paste(message,"\n", sep=""))
297 }
298
299 # As explained above, we obtain after all runs ntasks*[K1 or K2] medoids indices,
300 # depending whether WER=="end" or "mix", respectively.
301 indices_medoids_all <-
302 if (parll && ntasks>1)
303 unlist( parallel::parLapply(cl, indices_tasks, runTwoStepClustering) )
304 else
305 unlist( lapply(indices_tasks, runTwoStepClustering) )
306
307 if (parll && ntasks>1)
308 parallel::stopCluster(cl)
309
310 # For the last stage, ncores_tasks*(ncores_clusts+1) cores should be available:
311 # - ntasks for level 1 parallelism
312 # - ntasks*ncores_clust for level 2 parallelism,
313 # but since an extension MPI <--> tasks / OpenMP <--> sub-tasks is on the way,
314 # it's better to just re-use ncores_clust
315 ncores_last_stage <- ncores_clust
316
317 # Run last clustering tasks to obtain only K2 medoids indices
318 if (verbose)
319 cat("...Run final // stage 1 + stage 2\n")
320 indices_medoids <- clusteringTask1(indices_medoids_all, getContribs, K1, algoClust1,
321 nb_items_clust, ncores_tasks*ncores_clust, verbose)
322
323 indices_medoids <- clusteringTask2(indices_medoids, getSeries, K2, algoClust2,
324 nb_series_per_chunk,smooth_lvl,nvoice,nbytes,endian,ncores_last_stage,verbose)
325
326 # Compute synchrones, that is to say the cumulated power consumptions for each of the
327 # K2 final groups.
328 medoids <- getSeries(indices_medoids)
329 synchrones <- computeSynchrones(medoids, getSeries, nb_curves, nb_series_per_chunk,
330 ncores_last_stage, verbose)
331
332 # NOTE: no need to use big.matrix here, because only K2 << K1 << N remaining curves
333 list("medoids"=medoids, "ranks"=indices_medoids, "synchrones"=synchrones)
334}