tidymodel tuning fails for lasso regression using family = Gamma

[andrewwilson]

I'm trying to fit a lasso regression to some simulated data. The outcome variable is called 'loss' and is drawn from a gamma distribution. I can successfully fit a model when I hard code the penalty parameter but when I try to tune it as per the reproducible example below I get the following error:

"Error in class(): 0 arguments passed to 'class' which requires 1"

library(tidymodels)
library(dplyr)

# make up data
df <- tibble(loss = rgamma(10000, 1, scale = 14),
             predictor1 = runif(10000),
             predictor2 = runif(10000))

# create bootstraps
df_boots <- bootstraps(df)

# set up workflow
wf <- workflow() %>% 
  add_formula(loss ~ .) %>% 
  add_model(linear_reg(penalty = tune(), mixture = 1) %>% 
              set_engine("glmnet", family = Gamma))

# tune
tune_model <- tune_grid(
  wf,
  resamples = df_boots,
  grid = grid_regular(penalty(), levels = 50)
)
#> Warning: package 'rlang' was built under R version 3.6.3
#> 
#> Attaching package: 'rlang'
#> The following objects are masked from 'package:purrr':
#> 
#>     %@%, as_function, flatten, flatten_chr, flatten_dbl, flatten_int,
#>     flatten_lgl, flatten_raw, invoke, list_along, modify, prepend,
#>     splice
#> Warning: package 'vctrs' was built under R version 3.6.3
#> 
#> Attaching package: 'vctrs'
#> The following object is masked from 'package:tibble':
#> 
#>     data_frame
#> The following object is masked from 'package:dplyr':
#> 
#>     data_frame
#> Warning: package 'glmnet' was built under R version 3.6.3
#> Loading required package: Matrix
#> 
#> Attaching package: 'Matrix'
#> The following objects are masked from 'package:tidyr':
#> 
#>     expand, pack, unpack
#> Loaded glmnet 4.0-2
#> x Bootstrap01: preprocessor 1/1, model 1/1 (predictions): Error in class(): 0 arguments...
#> x Bootstrap02: preprocessor 1/1, model 1/1 (predictions): Error in class(): 0 arguments...
#> x Bootstrap03: preprocessor 1/1, model 1/1 (predictions): Error in class(): 0 arguments...
#> x Bootstrap04: preprocessor 1/1, model 1/1 (predictions): Error in class(): 0 arguments...
#> x Bootstrap05: preprocessor 1/1, model 1/1 (predictions): Error in class(): 0 arguments...
#> x Bootstrap06: preprocessor 1/1, model 1/1 (predictions): Error in class(): 0 arguments...
#> x Bootstrap07: preprocessor 1/1, model 1/1 (predictions): Error in class(): 0 arguments...
#> x Bootstrap08: preprocessor 1/1, model 1/1 (predictions): Error in class(): 0 arguments...
#> x Bootstrap09: preprocessor 1/1, model 1/1 (predictions): Error in class(): 0 arguments...
#> x Bootstrap10: preprocessor 1/1, model 1/1 (predictions): Error in class(): 0 arguments...
#> x Bootstrap11: preprocessor 1/1, model 1/1 (predictions): Error in class(): 0 arguments...
#> x Bootstrap12: preprocessor 1/1, model 1/1 (predictions): Error in class(): 0 arguments...
#> x Bootstrap13: preprocessor 1/1, model 1/1 (predictions): Error in class(): 0 arguments...
#> x Bootstrap14: preprocessor 1/1, model 1/1 (predictions): Error in class(): 0 arguments...
#> x Bootstrap15: preprocessor 1/1, model 1/1 (predictions): Error in class(): 0 arguments...
#> x Bootstrap16: preprocessor 1/1, model 1/1 (predictions): Error in class(): 0 arguments...
#> x Bootstrap17: preprocessor 1/1, model 1/1 (predictions): Error in class(): 0 arguments...
#> x Bootstrap18: preprocessor 1/1, model 1/1 (predictions): Error in class(): 0 arguments...
#> x Bootstrap19: preprocessor 1/1, model 1/1 (predictions): Error in class(): 0 arguments...
#> x Bootstrap20: preprocessor 1/1, model 1/1 (predictions): Error in class(): 0 arguments...
#> x Bootstrap21: preprocessor 1/1, model 1/1 (predictions): Error in class(): 0 arguments...
#> x Bootstrap22: preprocessor 1/1, model 1/1 (predictions): Error in class(): 0 arguments...
#> x Bootstrap23: preprocessor 1/1, model 1/1 (predictions): Error in class(): 0 arguments...
#> x Bootstrap24: preprocessor 1/1, model 1/1 (predictions): Error in class(): 0 arguments...
#> x Bootstrap25: preprocessor 1/1, model 1/1 (predictions): Error in class(): 0 arguments...
#> Warning: All models failed. See the `.notes` column.

^{Created on 2021-02-15 by the reprex package (v1.0.0)}

[Max]

Can you please provide a minimal reprex (reproducible example)? The goal of a reprex is to make it as easy as possible for me to recreate your problem so that I can fix it: please help me help you!

If you've never heard of a reprex before, start by reading "What is a reprex", and follow the advice further down that page.

[andrewwilson]

Hi Max,
Sorry - I tried to follow the reprex guidelines when creating my original post. I've amended it now to try and make the reproducible example more minimal!
Thanks

[Max]

That looks great.

It seems that glmnet doesn't offer all of the exponential families that glm does. See their docs for the list of acceptable values of family

[andrewwilson]

Thanks Max,

The documentation that you link to above suggests that glmnet should support all exponential families from glm . See the following section:

From version 4.0 onwards, glmnet supports both the original built-in families, as well as any family object as used by stats:glm() . This opens the door to a wide variety of additional models. For example family=binomial(link=cloglog) or family=negative.binomial(theta=1.5) (from the MASS library). Note that the code runs faster for the built-in families.

The reproducible example below shows that it is possible to fit a model with a Gamma family and then make predictions so it seems that the issue is with the tuning.

library(tidymodels)
library(dplyr)

# make up data
df <- tibble(loss = rgamma(10000, 1, scale = 14),
             predictor1 = runif(10000),
             predictor2 = runif(10000))


# fit model
fitted_model <- workflow() %>% 
  add_formula(loss ~ .) %>% 
  add_model(linear_reg(penalty = 0.1, mixture = 1) %>% 
              set_engine("glmnet", family = Gamma)) %>% 
  fit(df)


# make predictions
predict(fitted_model, df)
#> # A tibble: 10,000 x 1
#>    .pred
#>    <dbl>
#>  1  14.1
#>  2  13.8
#>  3  14.0
#>  4  14.0
#>  5  14.0
#>  6  14.0
#>  7  14.0
#>  8  13.8
#>  9  13.8
#> 10  14.0
#> # ... with 9,990 more rows

^{Created on 2021-02-15 by the reprex package (v1.0.0)}

[andrewwilson]

Hi Max,
Since it does appear that glmnet supports all built in families from glm I just wanted to check whether support for this will be built in to future versions of tune?
Thanks,