I came across the same thing, doing multinomial cross validation with cv.glmnet but also doing it with a for loop with subsets on the X matrix and y response categories. I've tested it out various ways and I think the problem occurs because in one of the folds there are no codes for at least one of the responses. From what I gather, this trips up glmnet. See in the table code below where in the first case no zeroes appear, but in the second a zero appears.
rand <- sample(3,dim(alldata)[1], replace=T) # alldata is a dataframe; allcodes is factor response variables obj1 <- glmnet(x=alldata[rand!=2,],y=allcodes[rand!=2], family="multinomial",maxit=500) #Worked obj2 <- glmnet(x=alldata[rand!=3,],y=allcodes[rand!=3], family="multinomial",maxit=500) #doesn't work > table(allcodes[rand!=2]) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 84 31 14 67 8 9 8 16 31 5 11 3 35 3 9 7 2 17 18 12 3 1 4 1 > table(allcodes[rand!=3]) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 85 20 14 72 12 7 13 15 32 4 13 3 26 3 15 5 6 13 23 16 1 0 3 1 I've looked at this with various sequences and it always seems to work when there's no zeroes, and crashes when there are zeroes. I'm working on a small data frame here (because of memory issues) so I don't think in general I would have 0s in nfold code categories. -Loren
______________________________________________ R-help@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide http://www.R-project.org/posting-guide.html and provide commented, minimal, self-contained, reproducible code.
