parallelization of FLR

[marchtaylor]

Some colleagues and I are having some difficulty setting up parallelization using FLR. Specifically, we are trying to apply a genetic algorithm (GA::ga), which I believe uses the parallel package. As long as we leave parallel=FALSE in that function, it runs without errors. With parallelization, we receive the following error

 Error in { : 
  task 1 failed - "error in evaluating the argument 'j' in selecting a method for function '[<-': Error in Summary.factor(1L, na.rm = FALSE) : 
  'min' not meaningful for factors
Calls: ac -> Summary.factor
"

This code is not part of our "fitness" function, nor do I see it within ga, so my guess is that it may be somewhere in FLCore. I had a previous problem with parallelization using the parallel package, which I believe dealt with conflicting iter() functions. I solved this by using snow for parallelization. I'm an not sure if this is a related issue, but was hoping someone might have some ideas.

Cheers

[iagomosqueira]

Hi,

'[' is overloaded for FLQuant, and it appears to call it with a factor for the year (j) dimension.

Can you please give us some code, for example based on ple4, so I can track it down?

[marchtaylor]

Hi Iago,
Thanks for your quick response. By making a minimal example, we realized that the problem is somewhere in our "fitness" function - so it's not a FLR-specific issue. We will head back through our code and try to figure out the issue. Here is the minimal example in case anyone is interested:

# required packages
library(FLCore)
library(FLash)
library(FLAssess)
library(ggplotFL)
library(GA)


# load data
data(ple4)
plot(ple4)


# SRR
ple4SR <- as.FLSR(ple4)
model(ple4SR) <- bevholt
ple4SR <- fmle(ple4SR) 
plot(ple4SR)


# fitness function
stffun <- function(Ftarget){
nyear <- 10
# Make the control object
ctrl_target <- data.frame(
  year = seq(from = range(ple4)["maxyear"], length.out = nyear),
  quantity = "f",
  val = Ftarget,
  max = NA, min=NA
)
ctrl_obj <- fwdControl(ctrl_target)
ple4_stf <- stf(ple4, nyear)
ple4_stf <- fwd(ple4_stf, ctrl = ctrl_obj, sr = ple4SR) 
# plot(ple4_stf)
# fitness "score" = total harvest in final year
return(sum(ple4_stf@catch[, ac(range(ple4)["maxyear"]+nyear-1)])) 
}


# Genetic algorithm search for Fmsy ---------------------------------------

# works without parallelization
system.time(
res1 <- ga(
  type = "real-valued",
  fitness = stffun,
  min = 0.05,
  max = 1.5,
  popSize = 20,
  maxiter = 5,
  seed = 1,
  parallel = FALSE
)
)
res1@fitnessValue
res1@solution # best F
plot(res1)

# also works in parallel
system.time(
res2 <- ga(
  type = "real-valued",
  fitness = stffun,
  min = 0.05,
  max = 1.5,
  popSize = 20,
  maxiter = 5,
  seed = 1,
  parallel = TRUE # 4
)
)
res2@fitnessValue
res2@solution # best F
plot(res2)

[iagomosqueira]

OK, did it have to do with a factor being used for subsetting? Glad to hear it works now.

[marchtaylor]

We still no not know where the error is coming from, but it is deeper in our model than in the functions called by this basic example. I will update here if I find any issues with FLR-related functions.

[iagomosqueira]

Perfect. Let me know and I will reopen it and investigate.