{SLmetrics}: Machine learning performance evaluation on steroids

{SLmetrics} is a lightweight R package written in C++ and {Rcpp} for memory-efficient and lightning-fast machine learning performance evaluation; it’s like using a supercharged {yardstick} but without the risk of soft to super-hard deprecations. {SLmetrics} covers both regression and classification metrics and provides (almost) the same array of metrics as {scikit-learn} and {PyTorch} all without {reticulate} and the Python compile-run-(crash)-debug cycle.

Depending on the mood and alignment of planets {SLmetrics} stands for Supervised Learning metrics, or Statistical Learning metrics. If {SLmetrics} catches on, the latter will be the core philosophy and include unsupervised learning metrics. If not, then it will remain a {pkg} for Supervised Learning metrics, and a sandbox for me to develop my C++ skills.

:rocket: Gettting Started

Below you’ll find instructions to install {SLmetrics} and get started with your first metric, the Root Mean Squared Error (RMSE).

:package: CRAN version

## install latest CRAN build
install.packages("SLmetrics")

:books: Basic Usage

Below is a minimal example demonstrating how to compute both unweighted and weighted RMSE.

library(SLmetrics)

actual    <- c(10.2, 12.5, 14.1)
predicted <- c(9.8, 11.5, 14.2)
weights   <- c(0.2, 0.5, 0.3)

cat(
  "Root Mean Squared Error", rmse(
    actual    = actual,
    predicted = predicted,
  ),
  "Root Mean Squared Error (weighted)", weighted.rmse(
    actual    = actual,
    predicted = predicted,
    w         = weights
  ),
  sep = "\n"
)
#> Root Mean Squared Error
#> 0.6244998
#> Root Mean Squared Error (weighted)
#> 0.7314369

That’s all! Now you can explore the rest of this README for in-depth usage, performance comparisons, and more details about {SLmetrics}.

:information_source: Why?

Machine learning can be a complicated task; the steps from feature engineering to model deployment require carefully measured actions and decisions. One low-hanging fruit to simplify this process is performance evaluation.

At its core, performance evaluation is essentially just comparing two vectors - a programmatically and, at times, mathematically trivial step in the machine learning pipeline, but one that can become complicated due to:

1. Dependencies and potential deprecations
2. Needlessly complex or repetitive arguments
3. Performance and memory bottlenecks at scale

{SLmetrics} solves these issues by being:

1. Fast: Powered by C++ and {Rcpp}
2. Memory-efficient: Everything is structured around pointers and references
3. Lightweight: Only depends on {Rcpp} and {lattice}
4. Simple: S3-based, minimal overhead, and flexible inputs

Performance evaluation should be plug-and-play and “just work” out of the box - there’s no need to worry about quasiquations, dependencies, deprecations, or variations of the same functions relative to their arguments when using {SLmetrics}.

:zap: Performance Comparison

One, obviously, can’t build an R-package on C++ and {Rcpp} without a proper pissing contest at the urinals - below is a comparison in execution time and memory efficiency of two simple cases that any {pkg} should be able to handle gracefully; computing a 2 x 2 confusion matrix and computing the RMSE[^1].

:fast_forward: Speed comparison

As shown in the chart, {SLmetrics} maintains consistently low(er) execution times across different sample sizes.

:floppy_disk: Memory-efficiency

Below are the results for garbage collections and total memory allocations when computing a 2×2 confusion matrix (N = 1e7) and RMSE (N = 1e7) [^2]. Notice that {SLmetrics} requires no GC calls for these operations.

2 x 2 Confusion Matrix (N = 1e7)

RMSE (N = 1e7)

In both tasks, {SLmetrics} remains extremely memory-efficient, even at large sample sizes.

[!IMPORTANT]

From {bench} documentation: Total amount of memory allocated by R while running the expression. Memory allocated outside the R heap, e.g. by malloc() or new directly is not tracked, take care to avoid misinterpreting the results if running code that may do this.

:information_source: Basic usage

In its simplest form, {SLmetrics}-functions work directly with pairs of <numeric> vectors (for regression) or <factor> vectors (for classification). Below we demonstrate this on two well-known datasets, mtcars (regression) and iris (classification).

:books: Regression

We first fit a linear model to predict mpg in the mtcars dataset, then compute the in-sample RMSE:

## Evaluate a linear model on mpg (mtcars)
model <- lm(mpg ~ ., data = mtcars)
rmse(mtcars$mpg, fitted(model))
#> [1] 2.146905

:books: Classification

Now we recode the iris dataset into a binary problem (“virginica” vs. “others”) and fit a logistic regression. Then we generate predicted classes, compute the confusion matrix and summarize it.

## 1) recode iris
## to binary problem
iris$species_num <- as.numeric(
  iris$Species == "virginica"
)

## 2) fit the logistic
## regression
model <- glm(
  formula = species_num ~ Sepal.Length + Sepal.Width,
  data    = iris,
  family  = binomial(
    link = "logit"
  )
)

## 3) generate predicted
## classes
predicted <- factor(
  as.numeric(
    predict(model, type = "response") > 0.5
  ),
  levels = c(1,0),
  labels = c("Virginica", "Others")
)

## 4) generate actual
## values as factor
actual <- factor(
  x = iris$species_num,
  levels = c(1,0),
  labels = c("Virginica", "Others")
)

## 4) generate
## confusion matrix
summary(
  confusion_matrix <- cmatrix(
    actual    = actual,
    predicted = predicted
  )
)
#> Confusion Matrix (2 x 2) 
#> ================================================================================
#>           Virginica Others
#> Virginica        35     15
#> Others           14     86
#> ================================================================================
#> Overall Statistics (micro average)
#>  - Accuracy:          0.81
#>  - Balanced Accuracy: 0.78
#>  - Sensitivity:       0.81
#>  - Specificity:       0.81
#>  - Precision:         0.81

:information_source: Enable OpenMP

[!IMPORTANT]

OpenMP support in {SLmetrics} is experimental. Use it with caution, as performance gains and stability may vary based on your system configuration and workload.

You can control OpenMP usage within {SLmetrics} using openmp.on() and openmp.off() . Below are examples demonstrating how to enable and disable OpenMP:

## enable OpenMP
SLmetrics::openmp.on()
#> OpenMP enabled!

## disable OpenMP
SLmetrics::openmp.off()
#> OpenMP disabled!

To illustrate the impact of OpenMP on performance, consider the following benchmarks for calculating entropy on a 1,000,000 x 200 matrix over 100 iterations[^3].

:books: Entropy without OpenMP

1e6 x 200 matrix without OpenMP

:books: Entropy with OpenMP

1e6 x 200 matrix with OpenMP

:package: Install from source

Github release

## install github release
pak::pak(
    pkg = "serkor1/SLmetrics@*release",
    ask = FALSE
)

Nightly build

Clone repository with submodules

git clone --recurse-submodules https://github.com/serkor1/SLmetrics.git

Installing with build tools

make build

Installing with {pak}

## install nightly build
pak::pak(
    pkg = ".",
    ask = FALSE
)

:information_source: Code of Conduct

Please note that the {SLmetrics} project is released with a Contributor Code of Conduct. By contributing to this project, you agree to abide by its terms.

[^1]: The source code is available here and here.

[^2]: The source code is available here.

[^3]: The source code is available here.