问题
I am trying to set the seeds inside the caret's gafsControl(), but I am getting this error:
Error in { : task 1 failed - "supplied seed is not a valid integer"
I understand that seeds for trainControl() is a vector equal to the number of resamples plus one, with the number of combinations of models's tuning parameters (in my case 36, SVM with 6 Sigma and 6 Cost values) in each (resamples) entries. However, I couldn't figure out what I should use for gafsControl(). I've tried iters*popSize (100*10), iters (100), popSize (10), but none has worked.
Thanks in advance.
here is my code (with simulated data):
library(caret)
library(doMC)
library(kernlab)
registerDoMC(cores=32)
set.seed(1234)
train.set <- twoClassSim(300, noiseVars = 100, corrVar = 100, corrValue = 0.75)
mylogGA <- caretGA
mylogGA$fitness_extern <- mnLogLoss
#Index for gafsControl
set.seed(1045481)
ga_index <- createFolds(train.set$Class, k=3)
#Seed for the gafsControl()
set.seed(1056)
ga_seeds <- vector(mode = "list", length = 4)
for(i in 1:3) ga_seeds[[i]] <- sample.int(1500, 1000)
## For the last model:
ga_seeds[[4]] <- sample.int(1000, 1)
#Index for the trainControl()
set.seed(1045481)
tr_index <- createFolds(train.set$Class, k=5)
#Seeds for the trainControl()
set.seed(1056)
tr_seeds <- vector(mode = "list", length = 6)
for(i in 1:5) tr_seeds[[i]] <- sample.int(1000, 36)#
## For the last model:
tr_seeds[[6]] <- sample.int(1000, 1)
gaCtrl <- gafsControl(functions = mylogGA,
method = "cv",
number = 3,
metric = c(internal = "logLoss",
external = "logLoss"),
verbose = TRUE,
maximize = c(internal = FALSE,
external = FALSE),
index = ga_index,
seeds = ga_seeds,
allowParallel = TRUE)
tCtrl = trainControl(method = "cv",
number = 5,
classProbs = TRUE,
summaryFunction = mnLogLoss,
index = tr_index,
seeds = tr_seeds,
allowParallel = FALSE)
svmGrid <- expand.grid(sigma= 2^c(-25, -20, -15,-10, -5, 0), C= 2^c(0:5))
t1 <- Sys.time()
set.seed(1234235)
svmFuser.gafs <- gafs(x = train.set[, names(train.set) != "Class"],
y = train.set$Class,
gafsControl = gaCtrl,
trControl = tCtrl,
popSize = 10,
iters = 100,
method = "svmRadial",
preProc = c("center", "scale"),
tuneGrid = svmGrid,
metric="logLoss",
maximize = FALSE)
t2<- Sys.time()
svmFuser.gafs.time<-difftime(t2,t1)
save(svmFuser.gafs, file ="svmFuser.gafs.rda")
save(svmFuser.gafs.time, file ="svmFuser.gafs.time.rda")
Session Info:
> sessionInfo()
R version 3.2.2 (2015-08-14)
Platform: x86_64-pc-linux-gnu (64-bit)
Running under: Ubuntu 14.04.3 LTS
locale:
[1] LC_CTYPE=en_CA.UTF-8 LC_NUMERIC=C LC_TIME=en_CA.UTF-8
[4] LC_COLLATE=en_CA.UTF-8 LC_MONETARY=en_CA.UTF-8 LC_MESSAGES=en_CA.UTF-8
[7] LC_PAPER=en_CA.UTF-8 LC_NAME=C LC_ADDRESS=C
[10] LC_TELEPHONE=C LC_MEASUREMENT=en_CA.UTF-8 LC_IDENTIFICATION=C
attached base packages:
[1] stats graphics grDevices utils datasets methods base
other attached packages:
[1] kernlab_0.9-22 doMC_1.3.3 iterators_1.0.7 foreach_1.4.2 caret_6.0-52 ggplot2_1.0.1 lattice_0.20-33
loaded via a namespace (and not attached):
[1] Rcpp_0.12.0 magrittr_1.5 splines_3.2.2 MASS_7.3-43 munsell_0.4.2
[6] colorspace_1.2-6 foreach_1.4.2 minqa_1.2.4 car_2.0-26 stringr_1.0.0
[11] plyr_1.8.3 tools_3.2.2 parallel_3.2.2 pbkrtest_0.4-2 nnet_7.3-10
[16] grid_3.2.2 gtable_0.1.2 nlme_3.1-122 mgcv_1.8-7 quantreg_5.18
[21] MatrixModels_0.4-1 iterators_1.0.7 gtools_3.5.0 lme4_1.1-9 digest_0.6.8
[26] Matrix_1.2-2 nloptr_1.0.4 reshape2_1.4.1 codetools_0.2-11 stringi_0.5-5
[31] compiler_3.2.2 BradleyTerry2_1.0-6 scales_0.3.0 stats4_3.2.2 SparseM_1.7
[36] brglm_0.5-9 proto_0.3-10
>
回答1:
I am not so familiar with the gafsControl() function that you mention, but I encountered a very similar issue when setting parallel seeds using trainControl(). In the instructions, it describes how to create a list (length = number of resamples + 1), where each item is a list (length = number of parameter combinations to test). I find that doing that does not work (see topepo/caret issue #248 for info). However, if you then turn each item into a vector, e.g.
seeds <- lapply(seeds, as.vector)
then the seeds seem to work (i.e. models and predictions are entirely reproducible). I should clarify that this is using doMC as the backend. It may be different for other parallel backends.
Hope this helps
回答2:
I was able to figure out my mistake by inspecting gafs.default. The seeds inside gafsControl() takes a vector with length (n_repeats*nresampling)+1 and not a list (as in trainControl$seeds). It is actually stated in the documentation of ?gafsControl that seeds is a vector or integers that can be used to set the seed during each search. The number of seeds must be equal to the number of resamples plus one. I figured it out the hard way, this is a reminder to carefully read the documentation :D.
if (!is.null(gafsControl$seeds)) {
if (length(gafsControl$seeds) < length(gafsControl$index) +
1)
stop(paste("There must be at least", length(gafsControl$index) +
1, "random number seeds passed to gafsControl"))
}
else {
gafsControl$seeds <- sample.int(1e+05, length(gafsControl$index) +
1)
}
So, the proper way to set my ga_seeds is:
#Index for gafsControl
set.seed(1045481)
ga_index <- createFolds(train.set$Class, k=3)
#Seed for the gafsControl()
set.seed(1056)
ga_seeds <- sample.int(1500, 4)
回答3:
If that way settings seeds you can ensure each run the same feature subset is selected ? I ams asking due randominess of GA
来源：https://stackoverflow.com/questions/32494744/caret-setting-the-seeds-inside-the-gafscontrol