de_serialize works. Variables names improved. Code beautified. TODO: clustering tests
[epclust.git] / epclust / R / clustering.R
1 # Cluster one full task (nb_curves / ntasks series); only step 1
2 clusteringTask = function(indices, getCoefs, K1, nb_series_per_chunk, ncores)
3 {
4 cl = parallel::makeCluster(ncores)
5 repeat
6 {
7 nb_workers = max( 1, round( length(indices) / nb_series_per_chunk ) )
8 indices_workers = lapply(seq_len(nb_workers), function(i) {
9 upper_bound = ifelse( i<nb_workers,
10 min(nb_series_per_chunk*i,length(indices)), length(indices) )
11 indices[(nb_series_per_chunk*(i-1)+1):upper_bound]
12 })
13 indices = unlist( parallel::parLapply(cl, indices_workers, function(inds)
14 computeClusters1(getCoefs(inds), K1)) )
15 if (length(indices) == K1)
16 break
17 }
18 parallel::stopCluster(cl)
19 indices #medoids
20 }
21
22 # Apply the clustering algorithm (PAM) on a coeffs or distances matrix
23 computeClusters1 = function(coefs, K1)
24 indices[ cluster::pam(coefs, K1, diss=FALSE)$id.med ]
25
26 # Cluster a chunk of series inside one task (~max nb_series_per_chunk)
27 computeClusters2 = function(medoids, K2, getRefSeries, nb_series_per_chunk)
28 {
29 synchrones = computeSynchrones(medoids, getRefSeries, nb_series_per_chunk)
30 cluster::pam(computeWerDists(synchrones), K2, diss=TRUE)$medoids
31 }
32
33 # Compute the synchrones curves (sum of clusters elements) from a clustering result
34 computeSynchrones = function(medoids, getRefSeries, nb_series_per_chunk)
35 {
36 #les getSeries(indices) sont les medoides --> init vect nul pour chacun, puis incr avec les
37 #courbes (getSeriesForSynchrones) les plus proches... --> au sens de la norme L2 ?
38 K = nrow(medoids)
39 synchrones = matrix(0, nrow=K, ncol=ncol(medoids))
40 counts = rep(0,K)
41 index = 1
42 repeat
43 {
44 range = (index-1) + seq_len(nb_series_per_chunk)
45 ref_series = getRefSeries(range)
46 if (is.null(ref_series))
47 break
48 #get medoids indices for this chunk of series
49 for (i in seq_len(nrow(ref_series)))
50 {
51 j = which.min( rowSums( sweep(medoids, 2, series[i,], '-')^2 ) )
52 synchrones[j,] = synchrones[j,] + series[i,]
53 counts[j] = counts[j] + 1
54 }
55 index = index + nb_series_per_chunk
56 }
57 #NOTE: odds for some clusters to be empty? (when series already come from stage 2)
58 sweep(synchrones, 1, counts, '/')
59 }
60
61 # Compute the WER distance between the synchrones curves (in rows)
62 computeWerDist = function(curves)
63 {
64 if (!require("Rwave", quietly=TRUE))
65 stop("Unable to load Rwave library")
66 n <- nrow(curves)
67 delta <- ncol(curves)
68 #TODO: automatic tune of all these parameters ? (for other users)
69 nvoice <- 4
70 # noctave = 2^13 = 8192 half hours ~ 180 days ; ~log2(ncol(curves))
71 noctave = 13
72 # 4 here represent 2^5 = 32 half-hours ~ 1 day
73 #NOTE: default scalevector == 2^(0:(noctave * nvoice) / nvoice) * s0 (?)
74 scalevector <- 2^(4:(noctave * nvoice) / nvoice) * 2
75 #condition: ( log2(s0*w0/(2*pi)) - 1 ) * nvoice + 1.5 >= 1
76 s0=2
77 w0=2*pi
78 scaled=FALSE
79 s0log = as.integer( (log2( s0*w0/(2*pi) ) - 1) * nvoice + 1.5 )
80 totnoct = noctave + as.integer(s0log/nvoice) + 1
81
82 # (normalized) observations node with CWT
83 Xcwt4 <- lapply(seq_len(n), function(i) {
84 ts <- scale(ts(curves[i,]), center=TRUE, scale=scaled)
85 totts.cwt = Rwave::cwt(ts,totnoct,nvoice,w0,plot=0)
86 ts.cwt = totts.cwt[,s0log:(s0log+noctave*nvoice)]
87 #Normalization
88 sqs <- sqrt(2^(0:(noctave*nvoice)/nvoice)*s0)
89 sqres <- sweep(ts.cwt,MARGIN=2,sqs,'*')
90 sqres / max(Mod(sqres))
91 })
92
93 Xwer_dist <- matrix(0., n, n)
94 fcoefs = rep(1/3, 3) #moving average on 3 values (TODO: very slow! correct?!)
95 for (i in 1:(n-1))
96 {
97 for (j in (i+1):n)
98 {
99 #TODO: later, compute CWT here (because not enough storage space for 200k series)
100 # 'circular=TRUE' is wrong, should just take values on the sides; to rewrite in C
101 num <- filter(Mod(Xcwt4[[i]] * Conj(Xcwt4[[j]])), fcoefs, circular=TRUE)
102 WX <- filter(Mod(Xcwt4[[i]] * Conj(Xcwt4[[i]])), fcoefs, circular=TRUE)
103 WY <- filter(Mod(Xcwt4[[j]] * Conj(Xcwt4[[j]])), fcoefs, circular=TRUE)
104 wer2 <- sum(colSums(num)^2) / sum( sum(colSums(WX) * colSums(WY)) )
105 Xwer_dist[i,j] <- sqrt(delta * ncol(Xcwt4[[1]]) * (1 - wer2))
106 Xwer_dist[j,i] <- Xwer_dist[i,j]
107 }
108 }
109 diag(Xwer_dist) <- numeric(n)
110 Xwer_dist
111 }