fixes: TODO, debug test.clustering.R and finish writing clustering.R
[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 #' @details 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 #' The main argument -- \code{getSeries} -- has a quite misleading name, since it can be
25 #' either a [big.]matrix, a CSV file, a connection or a user function to retrieve
26 #' series; the name was chosen because all types of arguments are converted to a function.
27 #' When \code{getSeries} is given as a function, it must take a single argument,
28 #' 'indices', integer vector equal to the indices of the curves to retrieve;
29 #' see SQLite example. The nature and role of other arguments should be clear
30 #'
31 #' @param getSeries Access to the (time-)series, which can be of one of the three
32 #' following types:
33 #' \itemize{
34 #' \item [big.]matrix: each column contains the (time-ordered) values of one time-serie
35 #' \item connection: any R connection object providing lines as described above
36 #' \item character: name of a CSV file containing series in rows (no header)
37 #' \item function: a custom way to retrieve the curves; it has only one argument:
38 #' the indices of the series to be retrieved. See SQLite example
39 #' }
40 #' @param K1 Number of clusters to be found after stage 1 (K1 << N [number of series])
41 #' @param K2 Number of clusters to be found after stage 2 (K2 << K1)
42 #' @param nb_series_per_chunk (Maximum) number of series to retrieve in one batch
43 #' @param algoClust1 Clustering algorithm for stage 1. A function which takes (data, K)
44 #' as argument where data is a matrix in columns and K the desired number of clusters,
45 #' and outputs K medoids ranks. Default: PAM.
46 #' In our method, this function is called on iterated medoids during stage 1
47 #' @param algoClust2 Clustering algorithm for stage 2. A function which takes (dists, K)
48 #' as argument where dists is a matrix of distances and K the desired number of clusters,
49 #' and outputs K clusters representatives (curves). Default: PAM.
50 #' In our method, this function is called on a matrix of K1 x K1 (WER) distances computed
51 # between synchrones
52 #' @param nb_items_clust1 (~Maximum) number of items in input of the clustering algorithm
53 #' for stage 1. At worst, a clustering algorithm might be called with ~2*nb_items_clust1
54 #' items; but this could only happen at the last few iterations.
55 #' @param wav_filt Wavelet transform filter; see ?wavelets::wt.filter
56 #' @param contrib_type Type of contribution: "relative", "logit" or "absolute" (any prefix)
57 #' @param WER "end" to apply stage 2 after stage 1 has fully iterated, or "mix" to apply
58 #' stage 2 at the end of each task
59 #' @param sync_mean TRUE to compute a synchrone as a mean curve, FALSE for a sum
60 #' @param random TRUE (default) for random chunks repartition
61 #' @param ntasks Number of tasks (parallel iterations to obtain K1 [if WER=="end"]
62 #' or K2 [if WER=="mix"] medoids); default: 1.
63 #' Note: ntasks << N (number of series), so that N is "roughly divisible" by ntasks
64 #' @param ncores_tasks Number of parallel tasks (1 to disable: sequential tasks)
65 #' @param ncores_clust Number of parallel clusterings in one task (4 should be a minimum)
66 #' @param sep Separator in CSV input file (if any provided)
67 #' @param nbytes Number of bytes to serialize a floating-point number; 4 or 8
68 #' @param endian Endianness for (de)serialization ("little" or "big")
69 #' @param verbose Level of verbosity (0/FALSE for nothing or 1/TRUE for all; devel stage)
70 #' @param parll TRUE to fully parallelize; otherwise run sequentially (debug, comparison)
71 #'
72 #' @return A matrix of the final K2 medoids curves, in columns
73 #'
74 #' @references Clustering functional data using Wavelets [2013];
75 #' A. Antoniadis, X. Brossat, J. Cugliari & J.-M. Poggi.
76 #' Inter. J. of Wavelets, Multiresolution and Information Procesing,
77 #' vol. 11, No 1, pp.1-30. doi:10.1142/S0219691313500033
78 #'
79 #' @examples
80 #' \dontrun{
81 #' # WER distances computations are too long for CRAN (for now)
82 #'
83 #' # Random series around cos(x,2x,3x)/sin(x,2x,3x)
84 #' x = seq(0,500,0.05)
85 #' L = length(x) #10001
86 #' ref_series = matrix( c(cos(x),cos(2*x),cos(3*x),sin(x),sin(2*x),sin(3*x)), ncol=6 )
87 #' library(wmtsa)
88 #' series = do.call( cbind, lapply( 1:6, function(i)
89 #' do.call(cbind, wmtsa::wavBootstrap(ref_series[i,], n.realization=400)) ) )
90 #' #dim(series) #c(2400,10001)
91 #' medoids_ascii = claws(series, K1=60, K2=6, 200, verbose=TRUE)
92 #'
93 #' # Same example, from CSV file
94 #' csv_file = "/tmp/epclust_series.csv"
95 #' write.table(series, csv_file, sep=",", row.names=FALSE, col.names=FALSE)
96 #' medoids_csv = claws(csv_file, K1=60, K2=6, 200)
97 #'
98 #' # Same example, from binary file
99 #' bin_file <- "/tmp/epclust_series.bin"
100 #' nbytes <- 8
101 #' endian <- "little"
102 #' binarize(csv_file, bin_file, 500, nbytes, endian)
103 #' getSeries <- function(indices) getDataInFile(indices, bin_file, nbytes, endian)
104 #' medoids_bin <- claws(getSeries, K1=60, K2=6, 200)
105 #' unlink(csv_file)
106 #' unlink(bin_file)
107 #'
108 #' # Same example, from SQLite database
109 #' library(DBI)
110 #' series_db <- dbConnect(RSQLite::SQLite(), "file::memory:")
111 #' # Prepare data.frame in DB-format
112 #' n <- nrow(series)
113 #' time_values <- data.frame(
114 #' id = rep(1:n,each=L),
115 #' time = rep( as.POSIXct(1800*(0:n),"GMT",origin="2001-01-01"), L ),
116 #' value = as.double(t(series)) )
117 #' dbWriteTable(series_db, "times_values", times_values)
118 #' # Fill associative array, map index to identifier
119 #' indexToID_inDB <- as.character(
120 #' dbGetQuery(series_db, 'SELECT DISTINCT id FROM time_values')[,"id"] )
121 #' serie_length <- as.integer( dbGetQuery(series_db,
122 #' paste("SELECT COUNT * FROM time_values WHERE id == ",indexToID_inDB[1],sep="")) )
123 #' getSeries <- function(indices) {
124 #' request <- "SELECT id,value FROM times_values WHERE id in ("
125 #' for (i in indices)
126 #' request <- paste(request, indexToID_inDB[i], ",", sep="")
127 #' request <- paste(request, ")", sep="")
128 #' df_series <- dbGetQuery(series_db, request)
129 #' as.matrix(df_series[,"value"], nrow=serie_length)
130 #' }
131 #' medoids_db = claws(getSeries, K1=60, K2=6, 200))
132 #' dbDisconnect(series_db)
133 #'
134 #' # All computed medoids should be the same:
135 #' digest::sha1(medoids_ascii)
136 #' digest::sha1(medoids_csv)
137 #' digest::sha1(medoids_bin)
138 #' digest::sha1(medoids_db)
139 #' }
140 #' @export
141 claws <- function(getSeries, K1, K2, nb_series_per_chunk,
142 nb_items_clust1=7*K1,
143 algoClust1=function(data,K) cluster::pam(t(data),K,diss=FALSE)$id.med,
144 algoClust2=function(dists,K) t( cluster::pam(dists,K,diss=TRUE)$medoids ),
145 wav_filt="d8", contrib_type="absolute",
146 WER="end",sync_mean=TRUE,
147 random=TRUE,
148 ntasks=1, ncores_tasks=1, ncores_clust=4,
149 sep=",",
150 nbytes=4, endian=.Platform$endian,
151 verbose=FALSE, parll=TRUE)
152 {
153 # Check/transform arguments
154 if (!is.matrix(getSeries) && !bigmemory::is.big.matrix(getSeries)
155 && !is.function(getSeries)
156 && !methods::is(getSeries,"connection") && !is.character(getSeries))
157 {
158 stop("'getSeries': [big]matrix, function, file or valid connection (no NA)")
159 }
160 K1 <- .toInteger(K1, function(x) x>=2)
161 K2 <- .toInteger(K2, function(x) x>=2)
162 nb_series_per_chunk <- .toInteger(nb_series_per_chunk, function(x) x>=1)
163 # K1 (number of clusters at step 1) cannot exceed nb_series_per_chunk, because we will need
164 # to load K1 series in memory for clustering stage 2.
165 if (K1 > nb_series_per_chunk)
166 stop("'K1' cannot exceed 'nb_series_per_chunk'")
167 nb_items_clust1 <- .toInteger(nb_items_clust1, function(x) x>K1)
168 random <- .toLogical(random)
169 tryCatch( {ignored <- wavelets::wt.filter(wav_filt)},
170 error = function(e) stop("Invalid wavelet filter; see ?wavelets::wt.filter") )
171 ctypes = c("relative","absolute","logit")
172 contrib_type = ctypes[ pmatch(contrib_type,ctypes) ]
173 if (is.na(contrib_type))
174 stop("'contrib_type' in {'relative','absolute','logit'}")
175 if (WER!="end" && WER!="mix")
176 stop("'WER': in {'end','mix'}")
177 sync_mean <- .toLogical(sync_mean)
178 random <- .toLogical(random)
179 ntasks <- .toInteger(ntasks, function(x) x>=1)
180 ncores_tasks <- .toInteger(ncores_tasks, function(x) x>=1)
181 ncores_clust <- .toInteger(ncores_clust, function(x) x>=1)
182 if (!is.character(sep))
183 stop("'sep': character")
184 nbytes <- .toInteger(nbytes, function(x) x==4 || x==8)
185 verbose <- .toLogical(verbose)
186 parll <- .toLogical(parll)
187
188 # Since we don't make assumptions on initial data, there is a possibility that even
189 # when serialized, contributions or synchrones do not fit in RAM. For example,
190 # 30e6 series of length 100,000 would lead to a +4Go contribution matrix. Therefore,
191 # it's safer to place these in (binary) files, located in the following folder.
192 bin_dir <- ".epclust_bin/"
193 dir.create(bin_dir, showWarnings=FALSE, mode="0755")
194
195 # Binarize series if getSeries is not a function; the aim is to always use a function,
196 # to uniformize treatments. An equally good alternative would be to use a file-backed
197 # bigmemory::big.matrix, but it would break the uniformity.
198 if (!is.function(getSeries))
199 {
200 if (verbose)
201 cat("...Serialize time-series\n")
202 series_file = paste(bin_dir,"data",sep="") ; unlink(series_file)
203 binarize(getSeries, series_file, nb_series_per_chunk, sep, nbytes, endian)
204 getSeries = function(inds) getDataInFile(inds, series_file, nbytes, endian)
205 }
206
207 # Serialize all computed wavelets contributions into a file
208 contribs_file = paste(bin_dir,"contribs",sep="") ; unlink(contribs_file)
209 index = 1
210 nb_curves = 0
211 if (verbose)
212 cat("...Compute contributions and serialize them\n")
213 nb_curves = binarizeTransform(getSeries,
214 function(series) curvesToContribs(series, wf, ctype),
215 contribs_file, nb_series_per_chunk, nbytes, endian)
216 getContribs = function(indices) getDataInFile(indices, contribs_file, nbytes, endian)
217
218 # A few sanity checks: do not continue if too few data available.
219 if (nb_curves < K2)
220 stop("Not enough data: less series than final number of clusters")
221 nb_series_per_task = round(nb_curves / ntasks)
222 if (nb_series_per_task < K2)
223 stop("Too many tasks: less series in one task than final number of clusters")
224
225 # Generate a random permutation of 1:N (if random==TRUE); otherwise just use arrival
226 # (storage) order.
227 indices_all = if (random) sample(nb_curves) else seq_len(nb_curves)
228 # Split (all) indices into ntasks groups of ~same size
229 indices_tasks = lapply(seq_len(ntasks), function(i) {
230 upper_bound = ifelse( i<ntasks, min(nb_series_per_task*i,nb_curves), nb_curves )
231 indices_all[((i-1)*nb_series_per_task+1):upper_bound]
232 })
233
234 if (parll && ntasks>1)
235 {
236 # Initialize parallel runs: outfile="" allow to output verbose traces in the console
237 # under Linux. All necessary variables are passed to the workers.
238 cl = parallel::makeCluster(ncores_tasks, outfile="")
239 varlist = c("getSeries","getContribs","K1","K2","algoClust1","algoClust2",
240 "nb_series_per_chunk","nb_items_clust1","ncores_clust","sep",
241 "nbytes","endian","verbose","parll")
242 if (WER=="mix")
243 varlist = c(varlist, "medoids_file")
244 parallel::clusterExport(cl, varlist, envir = environment())
245 }
246
247 # This function achieves one complete clustering task, divided in stage 1 + stage 2.
248 # stage 1: n indices --> clusteringTask1(...) --> K1 medoids
249 # stage 2: K1 medoids --> clusteringTask2(...) --> K2 medoids,
250 # where n = N / ntasks, N being the total number of curves.
251 runTwoStepClustering = function(inds)
252 {
253 # When running in parallel, the environment is blank: we need to load required
254 # packages, and pass useful variables.
255 if (parll && ntasks>1)
256 require("epclust", quietly=TRUE)
257 indices_medoids = clusteringTask1(
258 inds, getContribs, K1, algoClust1, nb_series_per_chunk, ncores_clust, verbose, parll)
259 if (WER=="mix")
260 {
261 if (parll && ntasks>1)
262 require("bigmemory", quietly=TRUE)
263 medoids1 = bigmemory::as.big.matrix( getSeries(indices_medoids) )
264 medoids2 = clusteringTask2(medoids1, K2, algoClust2, getSeries, nb_curves,
265 nb_series_per_chunk, sync_mean, nbytes, endian, ncores_clust, verbose, parll)
266 binarize(medoids2, medoids_file, nb_series_per_chunk, sep, nbytes, endian)
267 return (vector("integer",0))
268 }
269 indices_medoids
270 }
271
272 # Synchrones (medoids) need to be stored only if WER=="mix"; indeed in this case, every
273 # task output is a set of new (medoids) curves. If WER=="end" however, output is just a
274 # set of indices, representing some initial series.
275 if (WER=="mix")
276 {medoids_file = paste(bin_dir,"medoids",sep="") ; unlink(medoids_file)}
277
278 if (verbose)
279 {
280 message = paste("...Run ",ntasks," x stage 1", sep="")
281 if (WER=="mix")
282 message = paste(message," + stage 2", sep="")
283 cat(paste(message,"\n", sep=""))
284 }
285
286 # As explained above, indices will be assigned to ntasks*K1 medoids indices [if WER=="end"],
287 # or nothing (empty vector) if WER=="mix"; in this case, medoids (synchrones) are stored
288 # in a file.
289 indices <-
290 if (parll && ntasks>1)
291 unlist( parallel::parLapply(cl, indices_tasks, runTwoStepClustering) )
292 else
293 unlist( lapply(indices_tasks, runTwoStepClustering) )
294 if (parll && ntasks>1)
295 parallel::stopCluster(cl)
296
297 # Right before the final stage, two situations are possible:
298 # a. data to be processed now sit in binary format in medoids_file (if WER=="mix")
299 # b. data still is the initial set of curves, referenced by the ntasks*K1 indices
300 # So, the function getSeries() will potentially change. However, computeSynchrones()
301 # requires a function retrieving the initial series. Thus, the next line saves future
302 # conditional instructions.
303 getRefSeries = getSeries
304
305 if (WER=="mix")
306 {
307 indices = seq_len(ntasks*K2)
308 # Now series (synchrones) must be retrieved from medoids_file
309 getSeries = function(inds) getDataInFile(inds, medoids_file, nbytes, endian)
310 # Contributions must be re-computed
311 unlink(contribs_file)
312 index = 1
313 if (verbose)
314 cat("...Serialize contributions computed on synchrones\n")
315 ignored = binarizeTransform(getSeries,
316 function(series) curvesToContribs(series, wf, ctype),
317 contribs_file, nb_series_per_chunk, nbytes, endian)
318 }
319
320 #TODO: check THAT
321
322
323 # Run step2 on resulting indices or series (from file)
324 if (verbose)
325 cat("...Run final // stage 1 + stage 2\n")
326 indices_medoids = clusteringTask1(indices, getContribs, K1, algoClust1,
327 nb_series_per_chunk, ncores_tasks*ncores_clust, verbose, parll)
328 medoids1 = bigmemory::as.big.matrix( getSeries(indices_medoids) )
329 medoids2 = clusteringTask2(medoids1, K2, algoClust2, getRefSeries, nb_curves,
330 nb_series_per_chunk, sync_mean, nbytes, endian, ncores_tasks*ncores_clust, verbose, parll)
331
332 # Cleanup: remove temporary binary files and their folder
333 unlink(bin_dir, recursive=TRUE)
334
335 # Return medoids as a standard matrix, since K2 series have to fit in RAM
336 # (clustering algorithm 1 takes K1 > K2 of them as input)
337 medoids2[,]
338 }
339
340 #' curvesToContribs
341 #'
342 #' Compute the discrete wavelet coefficients for each series, and aggregate them in
343 #' energy contribution across scales as described in https://arxiv.org/abs/1101.4744v2
344 #'
345 #' @param series [big.]matrix of series (in columns), of size L x n
346 #' @inheritParams claws
347 #'
348 #' @return A [big.]matrix of size log(L) x n containing contributions in columns
349 #'
350 #' @export
351 curvesToContribs = function(series, wav_filt, contrib_type)
352 {
353 L = nrow(series)
354 D = ceiling( log2(L) )
355 nb_sample_points = 2^D
356 apply(series, 2, function(x) {
357 interpolated_curve = spline(1:L, x, n=nb_sample_points)$y
358 W = wavelets::dwt(interpolated_curve, filter=wf, D)@W
359 nrj = rev( sapply( W, function(v) ( sqrt( sum(v^2) ) ) ) )
360 if (contrib_type!="absolute")
361 nrj = nrj / sum(nrj)
362 if (contrib_type=="logit")
363 nrj = - log(1 - nrj)
364 nrj
365 })
366 }
367
368 # Check integer arguments with functional conditions
369 .toInteger <- function(x, condition)
370 {
371 errWarn <- function(ignored)
372 paste("Cannot convert argument' ",substitute(x),"' to integer", sep="")
373 if (!is.integer(x))
374 tryCatch({x = as.integer(x)[1]; if (is.na(x)) stop()},
375 warning = errWarn, error = errWarn)
376 if (!condition(x))
377 {
378 stop(paste("Argument '",substitute(x),
379 "' does not verify condition ",body(condition), sep=""))
380 }
381 x
382 }
383
384 # Check logical arguments
385 .toLogical <- function(x)
386 {
387 errWarn <- function(ignored)
388 paste("Cannot convert argument' ",substitute(x),"' to logical", sep="")
389 if (!is.logical(x))
390 tryCatch({x = as.logical(x)[1]; if (is.na(x)) stop()},
391 warning = errWarn, error = errWarn)
392 x
393 }