sklearn_freezer 0.1.0

Utility package for compiling scikit-learn classifiers for faster single-sample inference

Sklearn-Freezer

High-performance scikit-learn classifier compilation for ultra-fast single-sample inference.

Overview

Sklearn-Freezer compiles scikit-learn classifiers' predict or predict_proba methods into optimized implementations, dramatically improving single-sample prediction performance for real-time applications.

Disclaimer: This is a simple optimization created to address performance bottlenecks from iterative scikit-learn model calls. The implementation is naive with limited model support. Currently only predict_proba for binary classification of RandomForestClassifier and DecisionTreeClassifier are supported.

Compilation Backends

• Python: Pure Python implementation (baseline)
• Cython: Cython-compiled implementation (significant speedup)
• C: Native C extension (maximum speed)

Installation

# Basic installation
pip install sklearn-freezer

# With optional backends
pip install sklearn-freezer[cython]  # Cython backend
pip install sklearn-freezer[c]       # C backend
pip install sklearn-freezer[all]     # All backends

Quick Start

from sklearn.ensemble import RandomForestClassifier
from sklearn.datasets import make_classification
from sklearn_freezer import compile_predict_proba

# Train and compile model
X, y = make_classification(n_samples=1000, n_features=4, random_state=42)
clf = RandomForestClassifier(random_state=42).fit(X, y)
compiled_func = compile_predict_proba(clf, backend="c")

# Fast single-sample prediction
probability = compiled_func(*X[0])

Performance

The benchmark compares five prediction approaches:

1. Iterative clf.predict_proba: Individual clf.predict_proba([sample]) calls
2. Compiled Python/Cython/C: Using compiled backends
3. Batch clf.predict_proba: Single clf.predict_proba(all_samples) call

Performance ordering: iterative single-sample calls < compiled functions < batch processing.

Run benchmark: python example/benchmark.py

Requirements

• Python 3.10+
• scikit-learn

Optional: cython, setuptools