FerroML

High-performance ML in Rust with a scikit-learn-compatible Python API.

FerroML is a machine learning library written in Rust that provides 55+ algorithms with statistical rigor built in: confidence intervals on predictions, hypothesis testing for model comparison, and assumption checks on every model. It's 2-40x faster than scikit-learn on predict and up to 9x faster on fit for tree/ensemble models.

Installation

pip install ferroml

Requires Python 3.10+. Pre-built wheels available for Linux (x86_64, aarch64), macOS (x86_64, arm64), and Windows (x86_64).

Quick Start

from ferroml.linear import LinearRegression
import numpy as np

# Linear regression with full statistical diagnostics
X = np.random.randn(100, 5)
y = X @ np.array([1, 2, 3, 4, 5]) + np.random.randn(100) * 0.1

model = LinearRegression()
model.fit(X, y)
print(model.summary())  # R-style output: coefficients, std errors, p-values, R²

from ferroml.trees import RandomForestClassifier
from ferroml.preprocessing import StandardScaler
from ferroml.pipeline import Pipeline

# scikit-learn-compatible pipeline
pipe = Pipeline([
    ("scaler", StandardScaler()),
    ("clf", RandomForestClassifier(n_estimators=100)),
])
pipe.fit(X_train, y_train)
score = pipe.score(X_test, y_test)

from ferroml.preprocessing import CountVectorizer, TfidfTransformer
from ferroml.naive_bayes import MultinomialNB

# Text classification
cv = CountVectorizer()
X_counts = cv.fit_transform(documents)
tfidf = TfidfTransformer()
X_tfidf = tfidf.fit_transform(X_counts)
clf = MultinomialNB()
clf.fit(X_tfidf, y)

Performance vs scikit-learn

All benchmarks produce matching predictions. Speedup >1x = FerroML is faster.

FerroML is faster on predict universally (zero Python overhead) and on fit for tree/ensemble models (Rayon parallel construction). scikit-learn wins on fit for LAPACK/MKL-backed linear algebra.

Available Models

sklearn API Compatibility

FerroML supports the scikit-learn API conventions:

• fit() / predict() / transform() on all models
• score() on 56 models (R² for regressors, accuracy for classifiers)
• partial_fit() on 10 models for incremental learning
• decision_function() on 13 classifiers
• predict_proba() on probabilistic classifiers
• Pipeline and ColumnTransformer composition
• NumPy array input/output

Testing

5,650+ tests passing (3,550+ Rust + 2,100+ Python), validated against scikit-learn, scipy, xgboost, lightgbm, and statsmodels with 200+ cross-library correctness tests.

License

MIT OR Apache-2.0