Understanding multinomial nnet

问题

I am trying to understand the code behind nnet. I am currently getting different results when I split a multinomial factor in to the binary columns instead of using the formula method.

library(nnet)

set.seed(123)
y <- class.ind(iris$Species)
x <- as.matrix(iris[,1:4])
fit1 <- nnet(x, y, size = 3, decay = .1)

# weights:  27
#initial  value 164.236516 
#iter  10 value 102.567531
#iter  20 value 58.229722
#iter  30 value 39.720137
#iter  40 value 25.049530
#iter  50 value 23.671837
#iter  60 value 23.602392
#iter  70 value 23.601927
#final  value 23.601926 
#converged

pred1 <- predict(fit1, iris[,1:4])
rowSums(head(pred1))
[1] 1.032197661 1.033700173 1.032750746 1.034229149 1.032052937 1.032539980

set.seed(123)
fit2 <- nnet(Species ~ ., data = iris, size = 3, decay = .1)

# weights:  27
#initial  value 158.508573 
#iter  10 value 37.167558
#iter  20 value 26.815839
#iter  30 value 23.746418
#iter  40 value 23.698182
#iter  50 value 23.697907
#final  value 23.697907 
#converged

pred2 <- predict(fit2, iris[,1:4])
rowSums(head(pred2))
1 2 3 4 5 6 
1 1 1 1 1 1

I know I can just use the latter approach (formula method) but I want to understand why the results are different when it appears the same method of splitting the factor is in the source code nnet.formula.

回答1:

As noted by @user20650, the softmax argument is different. Inside nnet.formula there is the section:

if (length(lev) == 2L) {
    y <- as.vector(unclass(y)) - 1
    res <- nnet.default(x, y, w, entropy = TRUE, ...)
    res$lev <- lev
}
else {
    y <- class.ind(y)
    res <- nnet.default(x, y, w, softmax = TRUE, ...)
    res$lev <- lev
}

Here the softmax is set to TRUE. Setting it in the nnet call fixes the problem and they match now.

fit <- nnet(x, y, size = 3, decay = .1, softmax = TRUE)
pred <- predict(fit, iris[,1:4])
rowSums(head(pred))

来源：https://stackoverflow.com/questions/37621320/understanding-multinomial-nnet

标签

nnet