Strategies for Minimizing Model Size in R While Maintaining Predictive Ability

Question

In training regression models with the text package in R, the model's size increases with the number of training datapoints, resulting in unnecessarily large model objects. The models are created using the parsnip package with the glmnet engine. R's memory handling system, which prevents data duplication, makes it difficult to distinguish what components/attributes of the model that take up space; for instance: object_size(model) shows 700 MB, but object_size(model$final_recipe) and object_size(model$final_model) are nearly the same at 698 MB respectively, and thus doesn’t show the actual size of the components.

• How can I efficiently identify and remove the memory-heavy components of the model to reduce its size, while maintaining its predictive ability?

Example:

object_size(model) # 700Mb
object_size(model$final_recipe) # 698Mb
object_size(model$final_model) # 698Mb

When removing the final_recipe attribute (just as an example, nothing I would do in practice):

model$final_recipe <- NULL 

The size of the model is still 700Mb:

object_size(model) # 700Mb

Answer 1

If you only care about the predictions, you can extract the coefficients as a matrix and predict with a matrix multiplication. This example shows the result is the same as predict(model, data).

library(glmnet)
#> Loading required package: Matrix
#> Loaded glmnet 4.1-7

x <- mtcars[, c("cyl", "wt", "gear")] |> as.matrix()
y <- mtcars$mpg

fit1 <- cv.glmnet(x, y)

# extract coefficients
b <- coef(fit1, "lambda.min")

# compare predictions
data.frame(
  y,
  x,
  pred1 = as.vector(cbind(1, x) %*% b), # predict with matrix multiplication
  pred2 = as.vector(predict(fit1, x, "lambda.min"))
) |> 
  head()
#>                      y cyl    wt gear    pred1    pred2
#> Mazda RX4         21.0   6 2.620    4 22.16138 22.16138
#> Mazda RX4 Wag     21.0   6 2.875    4 21.39066 21.39066
#> Datsun 710        22.8   4 2.320    4 25.90130 25.90130
#> Hornet 4 Drive    21.4   6 3.215    3 20.36304 20.36304
#> Hornet Sportabout 18.7   8 3.440    3 16.84980 16.84980
#> Valiant           18.1   6 3.460    3 19.62254 19.62254

# compare sizes
object.size(fit1)
#> 26496 bytes
object.size(b)
#> 1976 bytes

^{Created on 2024-01-19 with reprex v2.0.2}

Answer 2

The butcher package likely does what you want: You can inspect your fitted model to find out which parts are so big with butcher::weigh() and reduce the size of the object without losing the ability to predict with it with butcher::butcher():

library(butcher)
library(lobstr)

our_model <- function() {
  some_junk_in_the_environment <- runif(1e6) # we didn't know about
  lm(mpg ~ ., data = mtcars) 
}

obj_size(our_model())
#> 8.02 MB

small_lm <- lm(mpg ~ ., data = mtcars) 
obj_size(small_lm)
#> 22.22 kB

big_lm <- our_model()
weigh(big_lm)
#> # A tibble: 25 × 2
#>    object            size
#>    <chr>            <dbl>
#>  1 terms         8.01    
#>  2 qr.qr         0.00666 
#>  3 residuals     0.00286 
#>  4 fitted.values 0.00286 
#>  5 effects       0.0014  
#>  6 coefficients  0.00109 
#>  7 call          0.000728
#>  8 model.mpg     0.000304
#>  9 model.cyl     0.000304
#> 10 model.disp    0.000304
#> # ℹ 15 more rows

butchered_lm <- butcher(big_lm)
obj_size(butchered_lm)
#> 22.74 kB

predict(butchered_lm, mtcars[1:2,])
#>     Mazda RX4 Mazda RX4 Wag 
#>      22.59951      22.11189

^{Created on 2024-01-22 with reprex v2.0.2}