Changes to allow sparsity predictor representations

[topepo]

Related to discussion in tidymodels/tidymodels#42

• Adds a new encoding fields called allow_sparse_x that is TRUE for glmnet, xgboost, and ranger models.

• ranger now uses their x/y interface. Tests indicates that there is no difference.

• Some refactoring of xgboost code.

Only changes the use of fit_xy(). For example:

library(tidymodels)
#> ── Attaching packages ─────────────────────────────────────────────────────────── tidymodels 0.1.1 ──
#> ✓ broom     0.7.0          ✓ recipes   0.1.13    
#> ✓ dials     0.0.9          ✓ rsample   0.0.8     
#> ✓ dplyr     1.0.2          ✓ tibble    3.0.3     
#> ✓ ggplot2   3.3.2          ✓ tidyr     1.1.2     
#> ✓ infer     0.5.2          ✓ tune      0.1.1.9000
#> ✓ modeldata 0.0.2          ✓ workflows 0.2.0     
#> ✓ parsnip   0.1.3.9000     ✓ yardstick 0.0.7     
#> ✓ purrr     0.3.4
#> ── Conflicts ────────────────────────────────────────────────────────────── tidymodels_conflicts() ──
#> x purrr::discard() masks scales::discard()
#> x dplyr::filter()  masks stats::filter()
#> x dplyr::lag()     masks stats::lag()
#> x recipes::step()  masks stats::step()
library(Matrix)
#> 
#> Attaching package: 'Matrix'
#> The following objects are masked from 'package:tidyr':
#> 
#>     expand, pack, unpack

xgb_spec <-
  boost_tree(trees = 10) %>% 
  set_engine("xgboost") %>% 
  set_mode("regression")

mtcar_x <- mtcars[, -1]
mtcar_mat <- as.matrix(mtcar_x)
mtcar_smat <- Matrix(mtcar_mat, sparse = TRUE)

set.seed(1)
from_df <- xgb_spec %>% fit_xy(mtcar_x, mtcars$mpg)
set.seed(1)
from_mat <- xgb_spec %>% fit_xy(mtcar_mat, mtcars$mpg)
set.seed(1)
from_sparse <- xgb_spec %>% fit_xy(mtcar_smat, mtcars$mpg)

all.equal(from_df$fit, from_mat$fit)
#> [1] TRUE
all.equal(from_df$fit, from_sparse$fit)
#> [1] TRUE

^{Created on 2020-09-29 by the reprex package (v0.3.0)}

[juliasilge]

Here is some benchmarking FWIW:

library(glmnet)
#> Loading required package: Matrix
#> Loaded glmnet 4.0-2
library(parsnip)

n <- 1e5 
y <- sample(0:1, n, replace = TRUE) 
x1 <- rnorm(n)
x2 <- sample(1:400, n, replace = TRUE)
X1 <- sparse.model.matrix(~ x1 + factor(x2))
X2 <- model.matrix(~ x1 + factor(x2))

lasso_spec <- linear_reg(mixture = 1) %>%
  set_mode("regression") %>%
  set_engine("glmnet")

bench::mark(
  check = FALSE, iterations = 10,
  lasso_spec %>% fit_xy(x = X2, y = y),
  lasso_spec %>% fit_xy(x = X1, y = y),
  glmnet(x = X2, y = y),
  glmnet(x = X1, y = y)
)
#> Warning: Some expressions had a GC in every iteration; so filtering is disabled.
#> # A tibble: 4 x 6
#>   expression                                min   median `itr/sec` mem_alloc
#>   <bch:expr>                           <bch:tm> <bch:tm>     <dbl> <bch:byt>
#> 1 lasso_spec %>% fit_xy(x = X2, y = y)    6.76s    6.81s     0.147   931.9MB
#> 2 lasso_spec %>% fit_xy(x = X1, y = y) 276.37ms  282.1ms     3.49     16.7MB
#> 3 glmnet(x = X2, y = y)                   6.57s    6.68s     0.149     623MB
#> 4 glmnet(x = X1, y = y)                164.27ms 166.48ms     5.75     16.4MB
#> # … with 1 more variable: `gc/sec` <dbl>

^{Created on 2020-10-02 by the reprex package (v0.3.0.9001)}

Looking so great! We can see the small parsnip overhead here, and the huge improvement to being able to use a sparse data structure.

[juliasilge]

Super happy about these changes. 😄

In the benchmarking I have been working on, this makes a big difference for glmnet especially in terms of fitting time.

We don't have this change/option documented anywhere in user-facing docs, and I think it should be surfaced at least mildly somewhere (not be entirely undocumented). Some options:

• In fit.R, change the x parameter to something like

A matrix, sparse matrix, or data frame of predictors. Only some models have support for sparse matrix input. 
See `parsnip::get_encoding()` for details.

• Add something to the Details of the models that have the support.

I think I like option 1 better.

[DavisVaughan]

Quick thought, only version 0.12.0 of ranger supports the x/y interface (the newest one), so you may want to add a version requirement to ranger in Suggests https://cran.r-project.org/web/packages/ranger/NEWS

[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.