Add sparse matrix support for LiblineaR engine

[juliasilge]

This PR adds sparse matrix support for the LiblineaR engine models. Closes #434.

library(tidymodels)
#> Registered S3 method overwritten by 'tune':
#>   method                   from   
#>   required_pkgs.model_spec parsnip

data(two_class_dat, package = "modeldata")
example_split <- initial_split(two_class_dat, prop = 0.99)
example_train <- training(example_split)
example_test  <-  testing(example_split)

rec <- recipe(Class ~ ., data = example_train) %>%
  step_normalize(all_numeric_predictors())

spec1 <- svm_linear() %>%
  set_engine("LiblineaR") %>%
  set_mode("classification")

spec2 <- logistic_reg(penalty = 0.1, mixture = 1) %>%
  set_engine("LiblineaR") %>%
  set_mode("classification")

sparse_bp <- hardhat::default_recipe_blueprint(composition = "dgCMatrix")

wf <- workflow() %>%
  add_recipe(rec, blueprint = sparse_bp)

wf_1 <- .fit_pre(wf %>% add_model(spec1), data = example_train)
class(wf_1$pre$mold$predictors)
#> [1] "dgCMatrix"
#> attr(,"package")
#> [1] "Matrix"
wf %>% add_model(spec1) %>% fit(example_train)
#> ══ Workflow [trained] ══════════════════════════════════════════════════════════
#> Preprocessor: Recipe
#> Model: svm_linear()
#> 
#> ── Preprocessor ────────────────────────────────────────────────────────────────
#> 1 Recipe Step
#> 
#> ● step_normalize()
#> 
#> ── Model ───────────────────────────────────────────────────────────────────────
#> $TypeDetail
#> [1] "L2-regularized L2-loss support vector classification dual (L2R_L2LOSS_SVC_DUAL)"
#> 
#> $Type
#> [1] 1
#> 
#> $W
#>              A          B      Bias
#> [1,] 0.3586417 -0.9565394 0.1148524
#> 
#> $Bias
#> [1] 1
#> 
#> $ClassNames
#> [1] Class1 Class2
#> Levels: Class1 Class2
#> 
#> $NbClass
#> [1] 2
#> 
#> attr(,"class")
#> [1] "LiblineaR"

wf_2 <- .fit_pre(wf %>% add_model(spec2), data = example_train)
class(wf_2$pre$mold$predictors)
#> [1] "dgCMatrix"
#> attr(,"package")
#> [1] "Matrix"
wf %>% add_model(spec2) %>% fit(example_train)
#> ══ Workflow [trained] ══════════════════════════════════════════════════════════
#> Preprocessor: Recipe
#> Model: logistic_reg()
#> 
#> ── Preprocessor ────────────────────────────────────────────────────────────────
#> 1 Recipe Step
#> 
#> ● step_normalize()
#> 
#> ── Model ───────────────────────────────────────────────────────────────────────
#> $TypeDetail
#> [1] "L1-regularized logistic regression (L1R_LR)"
#> 
#> $Type
#> [1] 6
#> 
#> $W
#>             A         B      Bias
#> [1,] 1.056906 -2.709859 0.3345231
#> 
#> $Bias
#> [1] 1
#> 
#> $ClassNames
#> [1] Class1 Class2
#> Levels: Class1 Class2
#> 
#> $NbClass
#> [1] 2
#> 
#> attr(,"class")
#> [1] "LiblineaR"

^{Created on 2021-03-18 by the reprex package (v1.0.0)}

[juliasilge]

The way that the LiblineaR algorithm works means that we don't necessarily get performance gains from using sparse data:

library(tidymodels)
#> Registered S3 method overwritten by 'tune':
#>   method                   from   
#>   required_pkgs.model_spec parsnip
library(textrecipes)

data("small_fine_foods")
sparse_bp <- hardhat::default_recipe_blueprint(composition = "dgCMatrix")

text_rec <-
  recipe(score ~ review, data = training_data) %>%
  step_tokenize(review)  %>%
  step_stopwords(review) %>%
  step_tokenfilter(review, max_tokens = 1e3) %>%
  step_tfidf(review)

liblinear_spec <- logistic_reg(penalty = 0.02, mixture = 1) %>%
  set_engine("LiblineaR") %>%
  set_mode("classification")

glmnet_spec <- logistic_reg(penalty = 0.02, mixture = 1) %>%
  set_engine("glmnet") %>%
  set_mode("classification")

liblinear_sparse <- 
  workflow() %>%
  add_recipe(text_rec, blueprint = sparse_bp) %>%
  add_model(liblinear_spec)

liblinear_default <- 
  workflow() %>%
  add_recipe(text_rec) %>%
  add_model(liblinear_spec)

glmnet_sparse <- 
  workflow() %>%
  add_recipe(text_rec, blueprint = sparse_bp) %>%
  add_model(glmnet_spec)

glmnet_default <- 
  workflow() %>%
  add_recipe(text_rec) %>%
  add_model(glmnet_spec)


set.seed(123)
food_folds <- vfold_cv(training_data, v = 3)

library(bench)

results <- mark(
  iterations = 10, check = FALSE,
  liblinear_sparse = fit_resamples(liblinear_sparse, food_folds),  
  liblinear_default = fit_resamples(liblinear_default, food_folds), 
  glmnet_sparse = fit_resamples(glmnet_sparse, food_folds),  
  glmnet_default = fit_resamples(glmnet_default, food_folds), 
)
#> Warning: Some expressions had a GC in every iteration; so filtering is disabled.

results
#> # A tibble: 4 x 6
#>   expression             min   median `itr/sec` mem_alloc `gc/sec`
#>   <bch:expr>        <bch:tm> <bch:tm>     <dbl> <bch:byt>    <dbl>
#> 1 liblinear_sparse      2.7s    2.73s    0.362      867MB   1.20  
#> 2 liblinear_default    2.71s    2.73s    0.366      804MB   1.10  
#> 3 glmnet_sparse        7.67s    7.71s    0.130      794MB   0.389 
#> 4 glmnet_default       1.18m    1.21m    0.0138     933MB   0.0828

autoplot(results, type = "ridge")
#> Picking joint bandwidth of 0.00186

^{Created on 2021-03-18 by the reprex package (v1.0.0)}

But we still want to set up the possibility, in case we add the ability to have sparse representations inside of recipes later or someone brings a sparse matrix straight to parsnip, etc.

[hfrick]

Is there a place in the parsnip documentation that tells you if an engine accepts sparse matrices?

[juliasilge]

This is what we landed on for now:

In fit.model_spec:

@param x A matrix, sparse matrix, or data frame of predictors. Only some models have support for sparse matrix input. See parsnip::get_encoding() for details. x should have column names.

And then you can call parsnip::get_encoding() on a model to see details:

library(parsnip)
get_encoding("linear_reg")
#> # A tibble: 5 x 7
#>   model    engine mode     predictor_indicato… compute_interce… remove_intercept allow_sparse_x
#>   <chr>    <chr>  <chr>    <chr>               <lgl>            <lgl>            <lgl>         
#> 1 linear_… lm     regress… traditional         TRUE             TRUE             FALSE         
#> 2 linear_… glmnet regress… traditional         TRUE             TRUE             TRUE          
#> 3 linear_… stan   regress… traditional         TRUE             TRUE             FALSE         
#> 4 linear_… spark  regress… traditional         TRUE             TRUE             FALSE         
#> 5 linear_… keras  regress… traditional         TRUE             TRUE             FALSE         
get_encoding("svm_linear")
#> # A tibble: 4 x 7
#>   model    engine   mode     predictor_indica… compute_interce… remove_intercept allow_sparse_x
#>   <chr>    <chr>    <chr>    <chr>             <lgl>            <lgl>            <lgl>         
#> 1 svm_lin… Libline… regress… none              FALSE            FALSE            TRUE          
#> 2 svm_lin… Libline… classif… none              FALSE            FALSE            TRUE          
#> 3 svm_lin… kernlab  regress… none              FALSE            FALSE            FALSE         
#> 4 svm_lin… kernlab  classif… none              FALSE            FALSE            FALSE         

[github-actions]

This pull request 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.