Error in if (r == 0) stop("Results from FUN have 0 variance, cannot build GP.") :

[GerardYu]

Hi i got the above error when using the bayesOpt() function. I've checked that both the data and labels are valid.

here's the full ouput

Running initial scoring function 8 times in 4 thread(s)...  2130.505 seconds

Starting Epoch 1 
  1) Fitting Gaussian Process...
Error in if (r == 0) stop("Results from FUN have 0 variance, cannot build GP.") : 
  missing value where TRUE/FALSE needed
In addition: Warning messages:
1: In .Internal(gc(verbose, reset, full)) :
  closing unused connection 7 (<-localhost:11016)
2: In .Internal(gc(verbose, reset, full)) :
  closing unused connection 6 (<-localhost:11016)
3: In .Internal(gc(verbose, reset, full)) :
  closing unused connection 5 (<-localhost:11016)
4: In .Internal(gc(verbose, reset, full)) :
  closing unused connection 4 (<-localhost:11016)

and here's the code that i've run

library(ParBayesianOptimization)
library(xgboost)
library(doParallel)
nthread<-16
##data prep

  load("label.rdata")
  load("dat.rdata")

##optimizer

obj_func <- function(eta, max_depth, min_child_weight, subsample, lambda, alpha) {
  xgb_train_dat<- xgb.DMatrix(data = train_feat_noNA, label = train_outcome_noNA)
  param <- list(

    # Hyper parameters 
    eta = eta,
    max_depth = max_depth,
    min_child_weight = min_child_weight,
    subsample = subsample,
    lambda = lambda,
    alpha = alpha,

    # Tree model 
    booster = "gbtree",

    # Regression problem 
    objective = "reg:squarederror",

    # Use the Mean Absolute Percentage Error
    eval_metric = "mae")

  xgbcv <- xgb.cv(params = param,
                  data =   xgb_train_dat,
                  nround = 500,
                  nfold = 5,
                  early_stopping_rounds = 5,
                  verbose = 0,
                  maximize = F)

  lst <- list(

    # First argument must be named as "Score"
    # Function finds maxima so inverting the output
    Score = -min(xgbcv$evaluation_log$test_mape_mean),

    # Get number of trees for the best performing model
    nrounds = xgbcv$best_iteration
  )

  return(lst)
}

bounds <- list(eta = c(0.0001, 0.1),
               max_depth = c(1L, 110L),
               min_child_weight = c(1, 50),
               subsample = c(0.1, 1),
               lambda = c(0.001, 1000),
               alpha = c(0.001, 1000))
set.seed(123)

#setting up parallel processing

cl <- makeCluster(nthread)
registerDoParallel(cl)
clusterExport(cl,c('train_feat_noNA','train_outcome_noNA'))
clusterEvalQ(cl,expr= {
  library(xgboost)
})

bayes_out <- bayesOpt(FUN = obj_func, bounds = bounds,iters.n = nthread,iters.k=nthread,initPoints = length(bounds) + 2,parallel = T)

[GerardYu]

So i discovered that the error disappears if eval_metric = "mape" instead of eval_metric = "mae"

[AnotherSamWilson]

This error means that all of the outputs from the tuning function have the same value. This causes singularity issues when trying to train a Gaussian Process.

[crypto-gabe]

I am running a binary classification problem and getting a similar issue. From my understanding the error is coming from the zeroOneScale function:

Scale a vector between 0-1
zeroOneScale <- function(vec) {

  r <- max(vec) - min(vec)

  # If the scoring function returned the same results
  # this results in the function a vector of 1s.
  if(r==0) stop("Results from FUN have 0 variance, cannot build GP.")

  vec <- (vec - min(vec))/r

  return(vec)

I think this is trying to scale my binary data which I do not want to happen.

Similar to OP this problem goes away when I use 'auc' as my evaluation metric but my dataset is highly skewed and I want to test with different metrics to see if this affects the tuning.

Heres my code for refernece

obj_func <- function(eta, max_depth, min_child_weight, subsample, lambda, alpha, nfolds) {

  dtrain <- xgb.DMatrix(data = as.matrix(train_data[, -3]), label = train_data$hidden_hypoxemia, missing = NA)

  param <- list(

    eta = eta,
    max_depth = max_depth,
    min_child_weight = min_child_weight,
    subsample = subsample,
    lambda = lambda,
    alpha = alpha,

    # Tree model 
    booster = "gbtree",

    objective = "binary:logistic",

    eval_metric = "logloss"

    )

  xgbcv <- xgb.cv(params = param,
                  data = dtrain,

                  nround = 50,
                  nfold = nfolds,
                  prediction = TRUE,
                  early_stopping_rounds = 10,
                  verbose = 2,
                  maximize = TRUE,
                  stratified = TRUE
                  )

  lst <- list(

    # First argument must be named as "Score"
    # Function finds maxima so inverting the output
    Score = suppressWarnings(min(xgbcv$evaluation_log$test_auc_mean)),

    # Get number of trees for the best performing model
    nrounds = xgbcv$best_iteration
  )

  return(lst)

param_bounds <- list(eta = c(0.001, 0.15),
                     max_depth = c(1L, 10L),
                     min_child_weight = c(1, 50),
                     subsample = c(0.1, 1),
                     lambda = c(1, 10),
                     alpha = c(1,10),
                     nfold = c(3L, 10L))
}

bayes_out <- bayesOpt(FUN = obj_func, bounds = param_bounds, initPoints = length(param_bounds) + 2, iters.n = 3)

Any help with this would be greatly appreciated