AffineFlow-ML

AffineFlow-ML

A high-performance Machine Learning engine for the AffineFlow ecosystem.

Designed for rapid experimentation without the Python Global Interpreter Lock (GIL) overhead, AffineFlow-ML executes compute-intensive tasks strictly in native C++ using Eigen and OpenBLAS. It features a dynamically compiled multi-architecture backend (Generic, AVX2, AVX-512) for maximum hardware-accelerated throughput safely dispatched at runtime.

Highlights

• Dynamic Hardware Dispatch: Automatically detects your CPU at runtime to dispatch highly optimized AVX2 or AVX-512 instructions.
• Zero-Copy Buffer Protocol: Uses Eigen::Ref and RowMajor structures to natively ingest NumPy arrays without memory duplication.
• Advanced KNN Classifier: OpenMP parallelized C++17 implementation with efficient top-k neighbor selection (std::nth_element).
• Distance Metrics: Native vectorized support for Euclidean, Manhattan, and Cosine distance algorithms.
• Weighted Voting: Choose between standard uniform voting or distance-weighted predictions.
• Adaptive PCA: Built-in dimensionality reduction with standard and thin-SVD modes for high-dimensional data.

Installation

Install the released wheel from PyPI:

pip install AffineFlow-ML

Quick Start

import numpy as np
import affineflow_ml as ml

# Example data
X_train = np.random.rand(1000, 50).astype(np.float64)
y_train = ["class_A" if x[0] > 0.5 else "class_B" for x in X_train]

# Configure the engine with new features
cfg = ml.KNNConfig()
cfg.k = 5
cfg.variance = 0.95
cfg.metric = ml.DistanceMetric.COSINE
cfg.weights = ml.Weights.DISTANCE

engine = ml.KNNEngine(cfg)
engine.train(X_train, y_train, scale=True)

X_test = np.random.rand(10, 50).astype(np.float64)
predictions = engine.predict_batch(X_test)
print(predictions)

Performance Benchmark

AffineFlow-ML achieves strict mathematical parity with Scikit-Learn while delivering up to 5.4x faster inference and transformation pipelines on high-dimensional datasets. Powered by a multithreaded C++17 Eigen backend with dynamic hardware dispatch (AVX2/AVX-512) and OpenMP parallelism, it eliminates Python loop and GIL bottlenecks.