Mercurial > repos > nicolas > oghma

comparison aggregation.R @ 62:613c81a4f7c0 draft

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Uploaded
author nicolas
date Wed, 26 Oct 2016 18:25:19 -0400
parents
children
comparison
equal deleted inserted replaced
61:339ae95a3d46 62:613c81a4f7c0
1 ########################################################
2 #
3 # creation date : 25/10/16
4 # last modification : 25/10/16
5 # author : Dr Nicolas Beaume
6 #
7 ########################################################
8
9 suppressWarnings(suppressMessages(library(GA)))
10 library("miscTools")
11 library(rpart)
12 suppressWarnings(suppressMessages(library(randomForest)))
13 library(e1071)
14 suppressWarnings(suppressMessages(library(glmnet)))
15 options(warn=-1)
16 ############################ helper functions #######################
17
18 ##### Genetic algorithm
19
20 # compute r2 by computing the classic formula
21 # compare the sum of square difference from target to prediciton
22 # to the sum of square difference from target to the mean of the target
23 r2 <- function(target, prediction) {
24 sst <- sum((target-mean(target))^2)
25 ssr <- sum((target-prediction)^2)
26 return(1-ssr/sst)
27 }
28
29 optimizeOneIndividual <- function(values, trueValue) {
30 # change the value into a function
31 f <- function(w) {sum(values * w/sum(w))}
32 fitness <- function(x) {1/abs(trueValue-f(x))}
33 resp <- ga(type = "real-valued", fitness = fitness, min = rep(0, length(values)), max = rep(1, length(values)),
34 maxiter = 1000, monitor = NULL, keepBest = T)
35 resp@solution <- resp@solution/sum(resp@solution)
36 return(resp)
37 }
38
39 optimizeWeight <- function(values, trueValue, n=1000) {
40 fitnessAll <- function(w) {
41 predicted <- apply(values, 1, weightedPrediction.vec, w)
42 return(mean(r2(trueValue, predicted)))
43 #return(mean(1/abs(trueValue-predicted)))
44 }
45 resp <- ga(type = "real-valued", fitness = fitnessAll, min = rep(0, ncol(values)), max = rep(1, ncol(values)),
46 maxiter = n, monitor = NULL, keepBest = T)
47 resp@solution <- resp@solution/sum(resp@solution)
48 return(resp)
49 }
50
51 weightedPrediction <- function(classifiers, w) {
52 if(length(w) > ncol(classifiers)) {
53 warning("more weights than classifiers, extra weigths are ignored")
54 w <- w[1:ncol(classifiers)]
55 } else if(length(w) < ncol(classifiers)) {
56 warning("less weights than classifiers, extra classifiers are ignored")
57 classifiers <- classifiers[,1:length(w)]
58 }
59 prediction <- NULL
60 prediction <- c(prediction, apply(classifiers, 1, weightedPrediction.vec, w))
61 return(prediction)
62 }
63
64 weightedPrediction.vec <- function(values, w) {
65 return(sum(values * w/sum(w)))
66 }
67
68 ##### meta-decision tree
69
70 tuneTree <- function(data, target) {
71 data <- data.frame(data, target=target)
72 size <- nrow(data)
73 xerror <- NULL
74 split <- 1:ceiling(size/5)
75 leafSize <- 1:ceiling(size/10)
76 xerror <- matrix(rep(-1, length(split)*length(leafSize)), ncol=length(leafSize))
77 cp <- matrix(rep(-1, length(split)*length(leafSize)), ncol=length(leafSize))
78 for(i in 1:length(split)) {
79 for(j in 1:length(leafSize)) {
80 op <- list(minsplit=split[i], minbucket=leafSize[j])
81 tree <- rpart(target ~., data=data, control=op, method="anova")
82 xerror[i,j] <- tree$cptable[which.min(tree$cptable[,"xerror"]),"xerror"]
83 cp[i,j] <- tree$cptable[which.min(tree$cptable[,"xerror"]),"CP"]
84 }
85 }
86 index <- which(xerror==min(xerror), arr.ind = T)
87 op <- list(minsplit=split[index[1]], minbucket=leafSize[index[2]], cp=cp[index[1], index[2]])
88 return(op)
89 }
90
91 ###### meta-LASSO
92 # create fold by picking at random row indexes
93 createFolds <- function(nbObs, n) {
94 # pick indexes
95 index <- sample(1:n, size=nbObs, replace = T)
96 # populate folds
97 folds <- NULL
98 for(i in 1:n) {
99 folds <- c(folds, list(which(index==i)))
100 }
101 return(folds)
102 }
103
104 searchParamLASSO <- function(genotype, phenotype, alpha=seq(0,1,0.1), n=7) {
105 folds <- createFolds(nrow(genotype), n = n)
106 acc <- NULL
107 indexAlpha <- 1
108 for(a in alpha) {
109 curAcc <- NULL
110 for(i in 1:n) {
111 train <- genotype[-folds[[i]],]
112 test <- genotype[folds[[i]],]
113 phenoTrain <- phenotype[-folds[[i]]]
114 phenoTest <- phenotype[folds[[i]]]
115 cv <- cv.glmnet(x=as.matrix(train), y=phenoTrain, alpha=a)
116 model <- glmnet(x=as.matrix(train), y=phenoTrain, alpha=a, lambda = cv$lambda.1se)
117 pred <- predict(model, test, type = "response")
118 curAcc <- c(curAcc, r2(phenoTest, pred))
119 }
120 acc <- c(acc, mean(curAcc))
121 }
122 names(acc) <- alpha
123 return(as.numeric(names(acc)[which.max(acc)]))
124 }
125
126 ###### meta-random forest
127
128 searchParamRF <- function(genotype, phenotype, rangeNtree, mtry=ncol(genotype)) {
129 n <- ceiling(nrow(genotype)/3)
130 indexTest <- sample(1:nrow(genotype), size=n)
131 train <- genotype[-indexTest,]
132 test <- genotype[indexTest,]
133 phenoTrain <- phenotype[-indexTest]
134 phenoTest <- phenotype[indexTest]
135 acc <- NULL
136 indexNtree <- 1
137 for(ntree in rangeNtree) {
138 model <- randomForest(x=train, y=phenoTrain, ntree = ntree, mtry = mtry)
139 pred <- predict(model, test)
140 acc <- c(acc, r2(phenoTest, pred))
141 }
142 names(acc) <- rangeNtree
143 best <- which.max(acc)
144 return(as.numeric(names(acc)[best]))
145 }
146
147 ###### meta-SVM
148 searchParamSVM <- function(train, target, kernel="radial") {
149 # tuning parameters then train
150 model <- NULL
151 switch(kernel,
152 sigmoid={
153 tune <- tune.svm(train, target, gamma = 10^(-6:-1), cost = 10^(0:2), kernel="sigmoid");
154 g <- tune$best.parameters[[1]];
155 c <- tune$best.parameters[[2]];
156 model <- svm(x=train, y=target, gamma = g, cost = c, kernel = "sigmoid")},
157 linear={
158 tune <- tune.svm(train, target, cost = 10^(0:2), kernel="linear");
159 c <- tune$best.parameters[[1]];
160 model <- svm(x=train, y=target, cost = c, kernel = "linear")},
161 polynomial={
162 tune <- tune.svm(train, target, gamma = 10^(-6:-1), cost = 10^(0:2), degree = 0:4, coef0 = 0:3, kernel="polynomial");
163 d <- tune$best.parameters[[1]];
164 g <- tune$best.parameters[[2]];
165 coef <- tune$best.parameters[[3]];
166 c <- tune$best.parameters[[4]];
167 model <- svm(x=train, y=target, gamma = g, cost = c, kernel = "polynomial", degree = d, coef0 = coef)},
168 {
169 tune <- tune.svm(train, target, gamma = 10^(-6:-1), cost = 10^(0:3), kernel="radial");
170 g <- tune$best.parameters[[1]];
171 c <- tune$best.parameters[[2]];
172 model <- svm(x=train, y=target, gamma = g, cost = c, kernel = "radial")}
173 )
174 return(model)
175 }
176
177 #################### upper level functions #####################
178
179 aggregateDT <- function(classifiers, target=NULL, prediction=F, model=NULL, out) {
180 if(!prediction) {
181 treeParam <- tuneTree(classifiers, target)
182 data <- data.frame(classifiers, target)
183 model <- rpart(target ~., data=data, method = "anova", control = treeParam)
184 model <- prune(model, cp=treeParam["cp"])
185 out <- paste(out, ".rds", sep = "")
186 saveRDS(model, out)
187 return(out)
188 } else {
189 return(predict(model, classifiers))
190 }
191 }
192
193 aggregateGeneticMean <- function(classifiers, target=NULL, prediction=F, model=NULL, out){
194 if(!prediction) {
195 opt <- optimizeWeight(values = classifiers, trueValue = target)
196 out <- paste(out, ".rds", sep = "")
197 saveRDS(opt@solution, out)
198 return(out)
199 } else {
200 return(weightedPrediction.vec(classifiers, model))
201 }
202 }
203
204 aggregateLASSO <- function(classifiers, target=NULL, prediction=F, model=NULL, alpha=NULL, out) {
205 if(!prediction) {
206 alpha <- searchParamLASSO(classifiers, target)
207 cv <- cv.glmnet(x=as.matrix(classifiers), y=target, alpha=alpha)
208 model <- glmnet(x=as.matrix(classifiers), y=target, alpha=alpha, lambda = cv$lambda.1se)
209 out <- paste(out, ".rds", sep = "")
210 saveRDS(model, out)
211 return(out)
212 } else {
213 return(predict(model, classifiers))
214 }
215 }
216
217 aggregateRF <- function(classifiers, target=NULL, model=NULL, ntree=NULL, prediction=F, out) {
218 if(!prediction) {
219 ntree <- searchParamRF(genotype = classifiers, phenotype = target,
220 rangeNtree = seq(100, 1000, 100))
221 model <- randomForest(x=classifiers, y=target, ntree = ntree, mtry = ncol(classifiers))
222 out <- paste(out, ".rds", sep = "")
223 saveRDS(model, out)
224 return(out)
225 } else {
226 return(predict(model, classifiers))
227 }
228 }
229
230 aggregateSVM <- function(classifiers, target=NULL, prediction=F,
231 model=NULL, c=NULL, g=NULL, d=NULL, coef=NULL, kernel="radial", out) {
232 if(!prediction) {
233 model <- searchParamSVM(train = classifiers, target = target, kernel = kernel)
234 out <- paste(out, ".rds", sep = "")
235 saveRDS(model, out)
236 return(out)
237 } else {
238 return(predict(model, classifiers))
239 }
240 }
241
242 ################################### main #############################
243 # # load argument
244 cmd <- commandArgs(T)
245 source(cmd[1])
246 # check if evaluation is required
247 evaluation <- F
248 if(as.integer(doEvaluation) == 1) {
249 evaluation <- T
250 con = file(folds)
251 folds <- readLines(con = con, n = 1, ok=T)
252 close(con)
253 folds <- readRDS(folds)
254 }
255 # check for model
256 if(model == "None") {
257 model <- NULL
258 prediction <- F
259 } else {
260 prediction <- T
261 con = file(model)
262 model <- readLines(con = con, n = 1, ok=T)
263 close(con)
264 model <- readRDS(model)
265 }
266 # load classifiers and phenotype
267 classifiers <- NULL
268 classifNames <- NULL
269 if(lassoPred !="None"){
270 classifiers <- c(classifiers, lassoPred)
271 classifNames <- c(classifNames, "lasso")
272 }
273 if(rrBLUPPred !="None"){
274 classifiers <- c(classifiers, rrBLUPPred)
275 classifNames <- c(classifNames, "rrBLUP")
276 }
277 if(rfPred !="None"){
278 classifiers <- c(classifiers, rfPred)
279 classifNames <- c(classifNames, "rf")
280 }
281 if(svmPred !="None"){
282 classifiers <- c(classifiers, svmPred)
283 classifNames <- c(classifNames, "svm")
284 }
285 classifPrediction <- NULL
286 for(classif in classifiers) {
287 classifPrediction <- c(classifPrediction, list(read.table(classif, sep="\t", h=T)))
288 }
289 classifPrediction <- data.frame(classifPrediction)
290 colnames(classifPrediction) <- classifNames
291 # phenotype is written as a table (in columns) but it must be sent as a vector for mixed.solve
292 phenotype <- read.table(phenotype, sep="\t", h=T)[,1]
293 # aggregate !
294 switch(method,
295 geneticMean={
296 res <- aggregateGeneticMean(classifiers = classifPrediction, target = phenotype,
297 out = out, prediction = prediction, model=model)
298 },
299 dt={
300 res <- aggregateDT(classifiers = classifPrediction, target = phenotype,
301 out = out, prediction = prediction, model=model)
302 },
303 lasso={
304 res <- aggregateLASSO(classifiers = data.matrix(classifPrediction), target = phenotype,
305 out = out, prediction = prediction, model=model)
306 },
307 rf={
308 res <- aggregateRF(classifiers = classifPrediction, target = phenotype,
309 out = out, prediction = prediction, model=model)
310 },
311 # svm
312 {res <- aggregateSVM(classifiers = classifPrediction, target = phenotype, kernel = kernel,
313 out = out, prediction = prediction, model = model)}
314 )
315 if(prediction) {
316 write.table(data.frame(lines=rownames(classifPrediction), res), paste(out,".csv", sep=""),
317 sep="\t", row.names = F)
318 } else {
319 cat(paste(paste(out, ".rds", sep = ""), "\n", sep=""))
320 }