| 1 | #TODO: setRefClass... to avoid copy data !! |
| 2 | #http://stackoverflow.com/questions/2603184/r-pass-by-reference |
| 3 | |
| 4 | #fields: data (can be NULL or provided by user), coeffs (will be computed |
| 5 | #con can be a character string naming a file; see readLines() |
| 6 | #data can be in DB format, on one column : TODO: guess (from header, or col. length...) |
| 7 | |
| 8 | |
| 9 | writeTmp(curves [uncompressed coeffs, limited number - nbSeriesPerChunk], last=FALSE) #if last=TRUE, close the conn |
| 10 | readTmp(..., from index, n curves) #careful: connection must remain open |
| 11 | #TODO: write read/write tmp reference ( on file in .tmp/ folder ... ) |
| 12 | |
| 13 | #data: |
| 14 | #stop("Unrecognizable 'data' argument (must be numeric, functional or connection)") |
| 15 | |
| 16 | #WER: "end" to apply stage 2 after stage 1 iterated, "mix" (or anything else...?!) to apply it after every stage 1 |
| 17 | epclust = function(data, K, nbPerChunk, WER="end", ncores=NULL, writeTmp=ref_writeTmp, readTmp=ref_readTmp) #where to put/retrieve intermediate results; if not provided, use file on disk |
| 18 | { |
| 19 | |
| 20 | |
| 21 | #on input: can be data or con; data handled by writing it to file (ascii or bin ?!), |
| 22 | #data: con or matrix or DB |
| 23 | |
| 24 | #1) acquire data (process curves, get as coeffs) |
| 25 | if (is.numeric(data)) |
| 26 | { |
| 27 | #full data matrix |
| 28 | index = 1 |
| 29 | n = nrow(data) |
| 30 | while (index < n) |
| 31 | { |
| 32 | writeTmp( getCoeffs(data) ) |
| 33 | index = index + nbSeriesPerChunk |
| 34 | } |
| 35 | } else if (is.function(data)) |
| 36 | { |
| 37 | #custom user function to retrieve next n curves, probably to read from DB |
| 38 | writeTmp( getCoeffs( data(nbPerChunk) ) ) |
| 39 | } else |
| 40 | { |
| 41 | #incremental connection |
| 42 | #read it one by one and get coeffs until nbSeriesPerChunk |
| 43 | #then launch a clustering task............ |
| 44 | #TODO: find a better way to parse than using a temp file |
| 45 | ascii_lines = readLines(data, nbSeriesPerChunk) |
| 46 | seriesChunkFile = ".tmp/seriesChunk" |
| 47 | writeLines(ascii_lines, seriesChunkFile) |
| 48 | writeTmp( getCoeffs( read.csv(seriesChunkFile) ) ) |
| 49 | } |
| 50 | |
| 51 | library(parallel) |
| 52 | ncores = ifelse(is.numeric(ncores), ncores, parallel::detectCores()) |
| 53 | cl = parallel::makeCluster(ncores) |
| 54 | 115 parallel::clusterExport(cl=cl, varlist=c("X", "Y", "K", "p"), envir=environment()) |
| 55 | 116 li = parallel::parLapply(cl, 1:B, getParamsAtIndex) |
| 56 | |
| 57 | #2) process coeffs (by nbSeriesPerChunk) and cluster in parallel (just launch async task, wait for them to complete, and re-do if necessary) |
| 58 | repeat |
| 59 | { |
| 60 | completed = rep(FALSE, ............) |
| 61 | #while there is jobs to do (i.e. size of tmp "file" is greater than nbSeriesPerChunk), |
| 62 | #A) determine which tasks which processor will do (OK) |
| 63 | #B) send each (sets of) tasks in parallel |
| 64 | #C) flush tmp file (current parallel processes will write in it) |
| 65 | #always check "complete" flag (array, as I did in MPI) to know if "slaves" finished |
| 66 | } |
| 67 | |
| 68 | parallel::stopCluster(cl) |
| 69 | |
| 70 | #3) readTmp last results, apply PAM on it, and return medoids + identifiers |
| 71 | |
| 72 | #4) apply stage 2 (in parallel ? inside task 2) ?) |
| 73 | if (WER == "end") |
| 74 | { |
| 75 | #from center curves, apply stage 2... |
| 76 | } |
| 77 | } |
| 78 | |
| 79 | getCoeffs = function(series) |
| 80 | { |
| 81 | #... return wavelets coeffs : compute in parallel ! |
| 82 | } |