使用`caret`交叉验证`glmnet`

Question

我正在阅读ISLR书，我正在尝试使用10倍交叉验证为Ridge回归模型找到最佳lambda。我尝试过具有非常相似配置的cv.glmnet和caret train函数，但结果却截然不同：

初始设置

library(ISLR)
library(glmnet)
library(caret)
hit = na.omit(Hitters)

grid = 10 ^ seq(10, -2, length= 100) # from 10B to .01, the grid for lambda
x = model.matrix(Salary ~ ., hit)[,-1]
y = hit$Salary

train.x = x[inTrain,]
train.y = y[inTrain]
test.x = x[-inTrain,]
test.y = y[-inTrain]

使用cv.glmnet

set.seed(1)
cv.out = cv.glmnet(x, y, alpha = 0, nfolds = 10,
               type.measure = "mse", lambda = grid)

best.lam = cv.out$lambda.min
best.lam
#231.013

cv.out$cvm[which(cv.out$lambda == cv.out$lambda.min)]
# mean cross-validated MSE: 120385.5

使用caret

ridge.model.caret = train(x, y, method = "glmnet",
                      tuneGrid = expand.grid(alpha = 0, lambda = grid),
                      tuneLength = 100, metric = "RMSE",
                      trControl = trainControl(method = "cv", number = 10))
ridge.model.caret$bestTune
# 18.73

ridge.model.caret$results$RMSE[which.min(ridge.model.caret$results$RMSE)]^2
# mean cross-validated MSE: 108339

请你帮我弄清楚我错过了什么？这是否意味着这两种不同的lambda具有足够接近的结果？