09e1ed7feca546b476b3e69d8b9e379e535a967d
[epclust.git] / 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_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 time-serie
42 #' \item connection: any R connection object providing lines as described above
43 #' \item character: name of a CSV file containing series in rows (no header)
44 #' \item function: a custom way to retrieve the curves; it has only one argument:
45 #' the indices of the series to be retrieved. See SQLite example
46 #' }
47 #' @param K1 Number of clusters to be found after stage 1 (K1 << N)
48 #' @param K2 Number of clusters to be found after stage 2 (K2 << K1)
49 #' @param nb_series_per_chunk Number of series to retrieve in one batch
50 #' @param nb_items_clust Number of items in 1st clustering algorithm input
51 #' @param algoClust1 Clustering algorithm for stage 1. A function which takes (data, K)
52 #' as argument where data is a matrix in columns and K the desired number of clusters,
53 #' and outputs K medoids ranks. Default: PAM.
54 #' @param algoClust2 Clustering algorithm for stage 2. A function which takes (dists, K)
55 #' as argument where dists is a matrix of distances and K the desired number of clusters,
56 #' and outputs K medoids ranks. Default: PAM.
57 #' @param wav_filt Wavelet transform filter; see ?wavelets::wt.filter
58 #' @param contrib_type Type of contribution: "relative", "logit" or "absolute" (any prefix)
59 #' @param WER "end" to apply stage 2 after stage 1 has fully iterated, or "mix" to apply
60 #' stage 2 at the end of each task
61 #' @param smooth_lvl Smoothing level: odd integer, 1 == no smoothing.
62 #' @param nvoice Number of voices within each octave for CWT computations
63 #' @param random TRUE (default) for random chunks repartition
64 #' @param ntasks Number of tasks (parallel iterations to obtain K1 [if WER=="end"]
65 #' or K2 [if WER=="mix"] medoids); default: 1.\cr
66 #' Note: ntasks << N (number of series), so that N is "roughly divisible" by ntasks
67 #' @param ncores_tasks Number of parallel tasks ('1' == sequential tasks)
68 #' @param ncores_clust Number of parallel clusterings in one task
69 #' @param sep Separator in CSV input file (if any provided)
70 #' @param nbytes 4 or 8 bytes to (de)serialize a floating-point number
71 #' @param endian Endianness for (de)serialization: "little" or "big"
72 #' @param verbose FALSE: nothing printed; TRUE: some execution traces
73 #'
74 #' @return A list:
75 #' \itemize{
76 #' \item medoids: matrix of the final K2 medoids curves
77 #' \item ranks: corresponding indices in the dataset
78 #' \item synchrones: sum of series within each final group
79 #' }
80 #'
81 #' @references Clustering functional data using Wavelets [2013];
82 #' A. Antoniadis, X. Brossat, J. Cugliari & J.-M. Poggi.
83 #' Inter. J. of Wavelets, Multiresolution and Information Procesing,
84 #' vol. 11, No 1, pp.1-30. doi:10.1142/S0219691313500033
85 #'
86 #' @examples
87 #' \dontrun{
88 #' # WER distances computations are too long for CRAN (for now)
89 #' # Note: on this small example, sequential run is faster
90 #'
91 #' # Random series around cos(x,2x,3x)/sin(x,2x,3x)
92 #' x <- seq(0,50,0.05)
93 #' L <- length(x) #1001
94 #' ref_series <- matrix( c(cos(x),cos(2*x),cos(3*x),sin(x),sin(2*x),sin(3*x)), ncol=6 )
95 #' library(wmtsa)
96 #' series <- do.call( cbind, lapply( 1:6, function(i)
97 #' do.call(cbind, wmtsa::wavBootstrap(ref_series[,i], n.realization=40)) ) )
98 #' # Mix series so that all groups are evenly spread
99 #' permut <- (0:239)%%6 * 40 + (0:239)%/%6 + 1
100 #' series = series[,permut]
101 #' #dim(series) #c(240,1001)
102 #' res_ascii <- claws(series, K1=30, K2=6, nb_series_per_chunk=500,
103 #' nb_items_clust=100, random=FALSE, verbose=TRUE, ncores_clust=1)
104 #'
105 #' # Same example, from CSV file
106 #' csv_file <- tempfile(pattern="epclust_series.csv_")
107 #' write.table(t(series), csv_file, sep=",", row.names=FALSE, col.names=FALSE)
108 #' res_csv <- claws(csv_file, 30, 6, 500, 100, random=FALSE, ncores_clust=1)
109 #'
110 #' # Same example, from binary file
111 #' bin_file <- tempfile(pattern="epclust_series.bin_")
112 #' nbytes <- 8
113 #' endian <- "little"
114 #' binarize(csv_file, bin_file, 500, ",", nbytes, endian)
115 #' getSeries <- function(indices) getDataInFile(indices, bin_file, nbytes, endian)
116 #' res_bin <- claws(getSeries, 30, 6, 500, 100, random=FALSE, ncores_clust=1)
117 #' unlink(csv_file)
118 #' unlink(bin_file)
119 #'
120 #' # Same example, from SQLite database
121 #' library(DBI)
122 #' series_db <- dbConnect(RSQLite::SQLite(), "file::memory:")
123 #' # Prepare data.frame in DB-format
124 #' n <- ncol(series)
125 #' times_values <- data.frame(
126 #' id = rep(1:n,each=L),
127 #' time = rep( as.POSIXct(1800*(1:L),"GMT",origin="2001-01-01"), n ),
128 #' value = as.double(series) )
129 #' dbWriteTable(series_db, "times_values", times_values)
130 #' # Fill associative array, map index to identifier
131 #' indexToID_inDB <- as.character(
132 #' dbGetQuery(series_db, 'SELECT DISTINCT id FROM times_values')[,"id"] )
133 #' serie_length <- as.integer( dbGetQuery(series_db,
134 #' paste("SELECT COUNT(*) FROM times_values WHERE id == ",indexToID_inDB[1],sep="")) )
135 #' getSeries <- function(indices) {
136 #' indices = indices[ indices <= length(indexToID_inDB) ]
137 #' if (length(indices) == 0)
138 #' return (NULL)
139 #' request <- "SELECT id,value FROM times_values WHERE id in ("
140 #' for (i in seq_along(indices)) {
141 #' request <- paste(request, indexToID_inDB[ indices[i] ], sep="")
142 #' if (i < length(indices))
143 #' request <- paste(request, ",", sep="")
144 #' }
145 #' request <- paste(request, ")", sep="")
146 #' df_series <- dbGetQuery(series_db, request)
147 #' matrix(df_series[,"value"], nrow=serie_length)
148 #' }
149 #' res_db <- claws(getSeries, 30, 6, 500, 100, random=FALSE, ncores_clust=1)
150 #' dbDisconnect(series_db)
151 #'
152 #' # All results should be equal:
153 #' all(res_ascii$ranks == res_csv$ranks
154 #' & res_ascii$ranks == res_bin$ranks
155 #' & res_ascii$ranks == res_db$ranks)
156 #' }
157 #' @export
158 claws <- function(series, K1, K2, nb_series_per_chunk, nb_items_clust=5*K1,
159 algoClust1=function(data,K) cluster::pam(t(data),K,diss=FALSE,pamonce=1)$id.med,
160 algoClust2=function(dists,K) cluster::pam(dists,K,diss=TRUE,pamonce=1)$id.med,
161 wav_filt="d8", contrib_type="absolute", WER="end", smooth_lvl=3, nvoice=4,
162 random=TRUE, ntasks=1, ncores_tasks=1, ncores_clust=3, sep=",", nbytes=4,
163 endian=.Platform$endian, verbose=FALSE)
164 {
165 # Check/transform arguments
166 if (!is.matrix(series) && !bigmemory::is.big.matrix(series)
167 && !is.function(series)
168 && !methods::is(series,"connection") && !is.character(series))
169 {
170 stop("'series': [big]matrix, function, file or valid connection (no NA)")
171 }
172 K1 <- .toInteger(K1, function(x) x>=2)
173 K2 <- .toInteger(K2, function(x) x>=2)
174 nb_series_per_chunk <- .toInteger(nb_series_per_chunk, function(x) x>=1)
175 nb_items_clust <- .toInteger(nb_items_clust, function(x) x>K1)
176 random <- .toLogical(random)
177 tryCatch({ignored <- wavelets::wt.filter(wav_filt)},
178 error=function(e) stop("Invalid wavelet filter; see ?wavelets::wt.filter") )
179 ctypes <- c("relative","absolute","logit")
180 contrib_type <- ctypes[ pmatch(contrib_type,ctypes) ]
181 if (is.na(contrib_type))
182 stop("'contrib_type' in {'relative','absolute','logit'}")
183 if (WER!="end" && WER!="mix")
184 stop("'WER': in {'end','mix'}")
185 random <- .toLogical(random)
186 ntasks <- .toInteger(ntasks, function(x) x>=1)
187 ncores_tasks <- .toInteger(ncores_tasks, function(x) x>=1)
188 ncores_clust <- .toInteger(ncores_clust, function(x) x>=1)
189 if (!is.character(sep))
190 stop("'sep': character")
191 nbytes <- .toInteger(nbytes, function(x) x==4 || x==8)
192 verbose <- .toLogical(verbose)
193
194 # Binarize series if it is not a function; the aim is to always use a function,
195 # to uniformize treatments. An equally good alternative would be to use a file-backed
196 # bigmemory::big.matrix, but it would break the "all-is-function" pattern.
197 if (!is.function(series))
198 {
199 if (verbose)
200 cat("...Serialize time-series (or retrieve past binary file)\n")
201 series_file <- ".series.epclust.bin"
202 if (!file.exists(series_file))
203 binarize(series, series_file, nb_series_per_chunk, sep, nbytes, endian)
204 getSeries <- function(inds) getDataInFile(inds, series_file, nbytes, endian)
205 }
206 else
207 getSeries <- series
208
209 # Serialize all computed wavelets contributions into a file
210 contribs_file <- ".contribs.epclust.bin"
211 if (verbose)
212 cat("...Compute contributions and serialize them (or retrieve past binary file)\n")
213 if (!file.exists(contribs_file))
214 {
215 nb_curves <- binarizeTransform(getSeries,
216 function(curves) curvesToContribs(curves, wav_filt, contrib_type),
217 contribs_file, nb_series_per_chunk, nbytes, endian)
218 }
219 else
220 {
221 # TODO: duplicate from getDataInFile() in de_serialize.R
222 contribs_size <- file.info(contribs_file)$size #number of bytes in the file
223 contrib_length <- readBin(contribs_file, "integer", n=1, size=8, endian=endian)
224 nb_curves <- (contribs_size-8) / (nbytes*contrib_length)
225 }
226 getContribs <- function(indices) getDataInFile(indices, contribs_file, nbytes, endian)
227
228 # A few sanity checks: do not continue if too few data available.
229 if (nb_curves < K2)
230 stop("Not enough data: less series than final number of clusters")
231 nb_series_per_task <- round(nb_curves / ntasks)
232 if (nb_series_per_task < K2)
233 stop("Too many tasks: less series in one task than final number of clusters")
234
235 # Generate a random permutation of 1:N (if random==TRUE);
236 # otherwise just use arrival (storage) order.
237 indices_all <- if (random) sample(nb_curves) else seq_len(nb_curves)
238 # Split (all) indices into ntasks groups of ~same size
239 indices_tasks <- lapply(seq_len(ntasks), function(i) {
240 upper_bound <- ifelse( i<ntasks, min(nb_series_per_task*i,nb_curves), nb_curves )
241 indices_all[((i-1)*nb_series_per_task+1):upper_bound]
242 })
243
244 parll <- (ncores_tasks > 1)
245 if (parll && ntasks>1)
246 {
247 # Initialize parallel runs: outfile="" allow to output verbose traces in the console
248 # under Linux. All necessary variables are passed to the workers.
249 cl <-
250 if (verbose)
251 parallel::makeCluster(ncores_tasks, outfile="")
252 else
253 parallel::makeCluster(ncores_tasks)
254 varlist <- c("ncores_clust","verbose", #task 1 & 2
255 "K1","getContribs","algoClust1","nb_items_clust") #task 1
256 if (WER=="mix")
257 {
258 # Add variables for task 2
259 varlist <- c(varlist, "K2","getSeries","algoClust2","nb_series_per_chunk",
260 "smooth_lvl","nvoice","nbytes","endian")
261 }
262 parallel::clusterExport(cl, varlist, envir <- environment())
263 }
264
265 # This function achieves one complete clustering task, divided in stage 1 + stage 2.
266 # stage 1: n indices --> clusteringTask1(...) --> K1 medoids (indices)
267 # stage 2: K1 indices --> K1xK1 WER distances --> clusteringTask2(...) --> K2 medoids,
268 # where n == N / ntasks, N being the total number of curves.
269 runTwoStepClustering <- function(inds)
270 {
271 # When running in parallel, the environment is blank: we need to load the required
272 # packages, and pass useful variables.
273 if (parll && ntasks>1)
274 require("epclust", quietly=TRUE)
275 indices_medoids <- clusteringTask1(inds, getContribs, K1, algoClust1,
276 nb_items_clust, ncores_clust, verbose)
277 if (WER=="mix")
278 {
279 indices_medoids <- clusteringTask2(indices_medoids, getSeries, K2, algoClust2,
280 nb_series_per_chunk,smooth_lvl,nvoice,nbytes,endian,ncores_clust,verbose)
281 }
282 indices_medoids
283 }
284
285 if (verbose)
286 {
287 message <- paste("...Run ",ntasks," x stage 1", sep="")
288 if (WER=="mix")
289 message <- paste(message," + stage 2", sep="")
290 cat(paste(message,"\n", sep=""))
291 }
292
293 # As explained above, we obtain after all runs ntasks*[K1 or K2] medoids indices,
294 # depending whether WER=="end" or "mix", respectively.
295 indices_medoids_all <-
296 if (parll && ntasks>1)
297 unlist( parallel::parLapply(cl, indices_tasks, runTwoStepClustering) )
298 else
299 unlist( lapply(indices_tasks, runTwoStepClustering) )
300
301 if (parll && ntasks>1)
302 parallel::stopCluster(cl)
303
304 # For the last stage, ncores_tasks*(ncores_clusts+1) cores should be available:
305 # - ntasks for level 1 parallelism
306 # - ntasks*ncores_clust for level 2 parallelism,
307 # but since an extension MPI <--> tasks / OpenMP <--> sub-tasks is on the way,
308 # it's better to just re-use ncores_clust
309 ncores_last_stage <- ncores_clust
310
311 # Run last clustering tasks to obtain only K2 medoids indices
312 if (verbose)
313 cat("...Run final // stage 1 + stage 2\n")
314 indices_medoids <- clusteringTask1(indices_medoids_all, getContribs, K1, algoClust1,
315 nb_items_clust, ncores_tasks*ncores_clust, verbose)
316
317 indices_medoids <- clusteringTask2(indices_medoids, getSeries, K2, algoClust2,
318 nb_series_per_chunk,smooth_lvl,nvoice,nbytes,endian,ncores_last_stage,verbose)
319
320 # Compute synchrones, that is to say the cumulated power consumptions for each of the K2
321 # final groups.
322 medoids <- getSeries(indices_medoids)
323 synchrones <- computeSynchrones(medoids, getSeries, nb_curves, nb_series_per_chunk,
324 ncores_last_stage, verbose)
325
326 # NOTE: no need to use big.matrix here, since there are only K2 << K1 << N remaining curves
327 list("medoids"=medoids, "ranks"=indices_medoids, "synchrones"=synchrones)
328 }