Error in na.fail.default and #461 fix

[farbodr]

Not sure if this is a new issue or not. I am getting this error and don't quite understand the fix that was applied for #461 as this appears to be related to the fix.

Here is my call to train. The data set contains no NA:

fitWithXGBLinear <- train(target~.,
                                data=train,  
                                method = 'xgbLinear',
                                tunrLength=3,
                                trControl = ctrl ,
                                metric="logLoss"
)

This produces the error:

Error in na.fail.default(list(target = c(10L, 8L, 8L, 12L, 1L, 1L, 4L,  : 
  missing values in object

As shown below, my data set contains no NA. This data set is from a current kaggle competition:

> lapply(train, function(x) any(is.na(x)))
$a
[1] FALSE

$target
[1] FALSE

$b
[1] FALSE

$c
[1] FALSE

This data set is from a current kaggle competition:

'data.frame':   51336 obs. of  4 variables:
 $ a     :Class 'integer64'  num [1:51336] 1.06e+05 -9.81e+71 3.76e+234 -4.97e-249 -3.98e+136 ...
 $ target: Factor w/ 12 levels "F23.","F24.26",..: 10 8 8 12 1 1 4 4 10 6 ...
 $ b     : chr  "小米" "TCL" "TCL" "小米" ...
 $ c     : chr  "MI 3" "么么哒" "么么哒" "MI 4" ...

If I change the call to train to this, it seems to run. I've never had to do this before so not sure I understand the need.

fitWithXGBLinear <- train(target~.,
                                data=train,  
                                method = 'xgbLinear',
                                tunrLength=3,
                                trControl = ctrl ,
                                na.action = na.omit,
                                metric="logLoss"
)

[topepo]

It is not a bug. Originally it was setup to stop execution when you have NA fails unless you use na.action = na.pass or other option. #461 fixed the bug (in R) that erroneously made the default na.pass.

The issue here is that people had missing data and did not realized that it was being omitted. Even when they chose an imputation method, the missing data were excluded before train was able to pre-process.

[farbodr]

Thanks for the explanation. I am still not clear why I got the missing value error when there are no NAs in the data.

[topepo]

We would need the data to figure that out.

[NBRAYKO]

@farbodr had same issue and it annoyed the hell out of me. Still not sure what exactly causes it, but found that I'd passed my object as a tibble, and it all works when I convert to data.frame. Weirdly, only got the error when I'd set up multiple cores with doMC.

[khari14]

titanicData$Pclass <- as.factor(titanicData$Pclass)
titanicData$Survived <- as.factor(titanicData$Survived)
titanicData$Title <- as.factor(titanicData$Title)
titanicData$Pclass <- as.factor(titanicData$FsizeD)
titanicData$Pclass <- as.factor(titanicData$Embarked)

index<- sample(1:nrow(titanicData),nrow(titanicData)*0.75, replace = F)
train<- titanicData[index,]
test<- titanicData[-index,]

library(randomForest)
rf_model <- randomForest(Survived ~ Pclass + Sex + Age + SibSp + Parch + Fare + Title + FsizeD, data = train, na.action = na.omit)

I am getting following error when I am fitting the rf:

Error in randomForest.default(m, y, ...) :
NA/NaN/Inf in foreign function call (arg 1)
In addition: Warning message:
In data.matrix(x) : NAs introduced by coercion

Can you please explain?

[topepo]

Please open a new issue and follow the issue template.

[DhanaLakshmi15]

str(data)
data$Selector<-as.factor(data$Selector)
table(data$Selector)
set.seed(123)
ind<-sample(2, nrow(data), replace=TRUE, prob=c(0.7,0.3))
train<-data[ind==1,]
test<-data[ind==2,]
#226 library(randomForest)
set.seed(222)
rf<-randomForest(Selector~., data=train)

i am getting the error when i am running the statement of "rf"
Error in randomForest.default(m, y, ...) : cannot handle categorical predictors with more than 53 categories.
can you explain and fix the problem?

[topepo]

@DhanaLakshmi15

This is unrelated to caret. Also, I don't think that there is a fix but you can try ranger instead.