rust_metrics

rust_metrics is an ML evaluation toolkit that brings Torchmetrics-style metrics to Rust. Each metric implements the same incremental Metric trait, so you can feed batched predictions over time and ask for the final score when ready. Every metric shares the same Torch-inspired test cases and examples so that the crate docs mirror the upstream behavior where the functionality matches.

Some benefits of rust-metrics:

• A standardized interface to increase reproducibility
• Reduces Boilerplate
• Rigorously tested
• Automatic accumulation over batches

Getting started

Add the crate to your project:

cargo add rust_metrics
# or enable the BERT-based similarity metric
cargo add rust_metrics --features text-bert

TorchMetrics-aligned examples

All snippets below (and method examples) reuse the exact inputs from the public TorchMetrics docs so you can cross-check the expected values.

Classification

use rust_metrics::{BinaryAccuracy, BinaryAuroc, Metric};

let target = [0_usize, 1, 0, 1, 0, 1];
let preds = [0.11, 0.22, 0.84, 0.73, 0.33, 0.92];

let mut acc = BinaryAccuracy::default();
acc.update((&preds[..], &target[..])).unwrap();
assert!((acc.compute().unwrap() - 2.0 / 3.0).abs() < f64::EPSILON);

let mut auroc = BinaryAuroc::new(0); 
let auroc_scores = [0.0, 0.5, 0.7, 0.8];
let auroc_target = [0_usize, 1, 1, 0];
auroc.update((&auroc_scores, &auroc_target)).unwrap();
assert!((auroc.compute().unwrap() - 0.5).abs() < f64::EPSILON);

Regression

use rust_metrics::{MeanSquaredError, MeanAbsoluteError, Metric};

let mut mse = MeanSquaredError::default();
mse.update((&[3.0, 5.0, 2.5, 7.0], &[2.5, 5.0, 4.0, 8.0])).unwrap();
assert!((mse.compute().unwrap() - 0.875).abs() < f64::EPSILON);

let mut mae = MeanAbsoluteError::default();
mae.update((&[2.5, 0.0, 2.0, 8.0], &[3.0, -0.5, 2.0, 7.0])).unwrap();
assert!((mae.compute().unwrap() - 0.5).abs() < f64::EPSILON);

Text

use rust_metrics::{Bleu, EditDistance, Metric};

let preds = ["the cat is on the mat"];
let targets = ["a cat is on the mat"];

let mut bleu = Bleu::default();
bleu.update((&preds, &targets)).unwrap();
assert!(bleu.compute().unwrap() > 0.5); 

let mut edit = EditDistance::default();
edit.update((&["rain"], &["shine"])).unwrap();
assert_eq!(edit.compute(), Some(3.0));

For SentenceEmbeddingSimilarity enable the text-bert feature; it mirrors the BERTScore example sentences and reports cosine similarities for each pair instead of precision/recall triples.

Implemented metrics

Classification

• BinaryAccuracy, MulticlassAccuracy
• BinaryPrecision, BinaryRecall, MulticlassPrecision
• BinaryF1Score, MulticlassF1Score
• BinaryHingeLoss, MulticlassHingeLoss
• BinaryJaccardIndex, MulticlassJaccardIndex
• BinaryConfusionMatrix
• BinaryAuroc

Regression

• MeanSquaredError
• NormalizedRootMeanSquaredError
• MeanAbsoluteError
• MeanAbsolutePercentageError
• R2Score

Text

• Bleu with optional smoothing and arbitrary n-gram depth
• EditDistance with sum or mean reduction
• RougeScore
• SentenceEmbeddingSimilarity (requires the text-bert feature) backed by fastembed. This metric embeds each sentence pair with lightweight BERT embeddings and reports cosine similarity scores.

Feature flags