Error when using the predict function

[SantiagoD999]

Hello to all,

The problem

I have been using the excellent parsnip package, I have been running some simulations to assess the effects of changing the regularization parameter. Now, there are cases in which I get the following message:

Error in .check_glmnet_penalty_predict():
! The glmnet model was fit with a single penalty value of 10. Predicting with a value of 10 will give incorrect results from glmnet().
Run rlang::last_trace() to see where the error occurred.

This even though In all the simulations I am using the same penalty value of 10. Here post I reproducible example, all the simulations run smoothly but for some reason, the 120th simulation (that uses the same data generating process and parametrization) gives me this error.

Reproducible example

## copy your code to the clipboard and run:
reprex::reprex(si = TRUE)

library(tidymodels)
asses_cv<-1
n_folds<-1
pen_val<-seq(from=10,to=10,length.out=1)
tuning_g<-expand_grid(penalty=pen_val,mixture=c(1))
for (i in 1:120){
  set.seed(i)
  x1<-rnorm(100)
  x2<-rnorm(100)
  y<-3+rnorm(100)
  
  train<-tibble(y,x1,x2)
  lm_elne<-linear_reg(penalty = tune(),mixture = tune()) %>%
    set_engine("glmnet",path_values = pen_val,thresh=1E-25)
  
  rec<-recipe(y~.,train)
  
  folds<-rolling_origin(train,initial = (NROW(train)-n_folds-asses_cv+1),assess = asses_cv)
  
  wf_elne<-workflow() %>%
    add_model(lm_elne) %>%
    add_recipe(rec)
  
  lm_elne_res<-tune_grid(wf_elne,resamples=folds,grid=tuning_g,metrics = yardstick::metric_set(rmse))
  
  wf_elne %>%
    finalize_workflow(select_best(lm_elne_res,metric = "rmse")) %>%
    fit(train) %>%
    predict(train[1,])
  print(i)
}

I would appreciate any comment.

Thank you very much,

[topepo]

The notion for this comes from glmnet itself:

WARNING: use with care. Avoid supplying a single value for lambda (for predictions after CV use predict() instead). Supply instead a decreasing sequence of lambda values. glmnet relies on its warms starts for speed, and its often faster to fit a whole path than compute a single fit.

I'll discuss with the group; we could convert this to a warning but there is probably a good reason we made it an error.