linear_reg() not responsive to penalty with "glmnet" engine

[ftvalentini]

I'm trying to run a Lasso model with a specific value for lambda (penalty). I reproduce the simple example available at https://www.tidyverse.org/articles/2018/11/parsnip-0-0-1/.

library(parsnip)
penalized <- linear_reg(mixture = 1, penalty = 0.1)
penalized
glmn_fit <-
  penalized %>%
  set_engine("glmnet") %>%
  fit(mpg ~ ., data = mtcars)
glmn_fit$spec
head(glmn_fit$fit$lambda)
tail(glmn_fit$fit$lambda)

The output:

> glmn_fit$spec
Linear Regression Model Specification (regression)

Main Arguments:
  penalty = 0.1
  mixture = 0

Computational engine: glmnet 

Model fit template:
glmnet::glmnet(x = as.matrix(x), y = y, weights = missing_arg(), 
    alpha = 0, family = "gaussian")
> head(glmn_fit$fit$lambda)
[1] 5146.981 4689.737 4273.114 3893.502 3547.614 3232.454
> tail(glmn_fit$fit$lambda)
[1] 0.8195449 0.7467388 0.6804006 0.6199557 0.5648805 0.5146981

Alpha seems to be correctly set in the call of glmnet, but no lambda is set.

Session info:

> sessionInfo()
R version 3.6.1 (2019-07-05)
Platform: x86_64-w64-mingw32/x64 (64-bit)
Running under: Windows 10 x64 (build 17134)

Matrix products: default

locale:
[1] LC_COLLATE=Spanish_Argentina.1252 
[2] LC_CTYPE=Spanish_Argentina.1252   
[3] LC_MONETARY=Spanish_Argentina.1252
[4] LC_NUMERIC=C                      
[5] LC_TIME=Spanish_Argentina.1252    

attached base packages:
[1] stats     graphics  grDevices utils     datasets  methods  
[7] base     

other attached packages:
[1] parsnip_0.0.3.1

loaded via a namespace (and not attached):
 [1] Rcpp_1.0.2       rstudioapi_0.7   magrittr_1.5    
 [4] devtools_1.13.6  tidyselect_0.2.5 lattice_0.20-38 
 [7] R6_2.4.0         rlang_0.4.0      foreach_1.4.4   
[10] dplyr_0.8.3      tools_3.6.1      grid_3.6.1      
[13] packrat_0.4.9-3  glmnet_2.0-18    withr_2.1.2     
[16] iterators_1.0.10 digest_0.6.18    assertthat_0.2.0
[19] tibble_2.1.3     crayon_1.3.4     Matrix_1.2-17   
[22] purrr_0.2.5      tidyr_0.8.1      codetools_0.2-16
[25] vctrs_0.2.0      zeallot_0.1.0    memoise_1.1.0   
[28] glue_1.3.0       compiler_3.6.1   pillar_1.4.2    
[31] generics_0.0.2   backports_1.1.2  pkgconfig_2.0.1 

[smingerson]

From News for 0.0.3

• For glmnet models, the full regularization path is always fit regardless of the value given to penalty. Previously, the model was fit with passing penalty to glmnet's lambda argument and the model could only make predictions at those specific values. (glmnet penalty parameter  #195)

Reading in the linear_reg() details, this is also mentioned:

For glmnet models, the full regularization path is always fit regardless of the value given to penalty. Also, there is the option to pass multiple values (or no values) to the penalty argument. When using the predict() method in these cases, the return value depends on the value of penalty. When using predict(), only a single value of the penalty can be used. When predicting on multiple penalties, the multi_predict() function can be used. It returns a tibble with a list column called .pred that contains a tibble with all of the penalty results.

[topepo]

@ftvalentini You can still get your lambda values via glmn_fit$spec$args$penalty.

[ftvalentini]

I know glmnet fits the whole path, but one might want to change that path.
I think it might be misleading that glmn_fit$fit$lambda and glmn_fit$spec$args$penalty are different when the lambda sequence is chosen by the user. For instance:

library(parsnip)
lambda_seq = exp(seq(0,-10,length.out=50))
penalized = linear_reg(mixture = 1, penalty = lambda_seq)
glmn_fit =
  penalized %>%
  set_engine("glmnet") %>%
  fit(mpg ~ ., data = mtcars)

all.equal(glmn_fit$fit$lambda, lambda_seq)
all.equal(glmn_fit$spec$args$penalty, lambda_seq)
multi_predict(glmn_fit, mtcars)

gives

> all.equal(glmn_fit$fit$lambda, lambda_seq)
[1] "Numeric: lengths (79, 50) differ"
> all.equal(glmn_fit$spec$args$penalty, lambda_seq)
[1] TRUE
> multi_predict(glmn_fit, mtcars)
# A tibble: 32 x 1
   .pred            
   <list>           
 1 <tibble [50 × 2]>
 2 <tibble [50 × 2]>
 3 <tibble [50 × 2]>
 4 <tibble [50 × 2]>
 5 <tibble [50 × 2]>
 6 <tibble [50 × 2]>
 7 <tibble [50 × 2]>
 8 <tibble [50 × 2]>
 9 <tibble [50 × 2]>
10 <tibble [50 × 2]>
# … with 22 more rows

So in the end the multi_predict works as it uses the correct penalty sequence for each row to predict.

[topepo]

The path is unique per data set and alpha value; you can't really change it. You can tell glmnet what values of lambda that you want to compute statistics for (such as predictions, coefficients, etc).

If you have specific values of lambda of interest, you can still get statistics related to your lambda sequence regardless of the value of penalty in the current implementation.

Setting lambda the way we do makes this slightly harder to do but just passing the value of penalty to lambda prevents other things from happening.

[github-actions]

This issue has been automatically locked. If you believe you have found a related problem, please file a new issue (with a reprex: https://reprex.tidyverse.org) and link to this issue.