addernet 1.6.0

Multiplication-free lookup neural models with scalar, multi-input, HDC, and additive APIs

AdderNet 1.6 development fork

This directory contains an audited and extended version of the addernet 1.5.0 package. It keeps the original scalar, HDC, attention, boosting, clustering, and additive APIs while adding safer native bindings and genuine multi-input models.

Main additions

• AdderNetLayer.fit(...): direct O(n + range) scalar LUT fitting.
• AdderNetLayer.partial_fit(...): blended streaming updates.
• Explicit close() and context-manager lifecycle.
• Atomic native saves and portable versioned .npz saves.
• Hardened native file loading and full validation of C arguments.
• AdderNetMultiInputLayer: N inputs, one or more outputs, additive and pairwise LUTs.
• AdderNetRegressor: alias for the multi-input regression model.
• AdderNetClassifier: classification wrapper with arbitrary class labels.
• Improved UniformQuantizer with feature-shape validation and inverse transform.
• OpenMP changed to opt-in because it slowed the memory-bound scalar lookup on the test host.

Two-input example

import numpy as np
from addernet import AdderNetMultiInputLayer

x = np.arange(32)
y = np.arange(32)
xx, yy = np.meshgrid(x, y, indexing="ij")
X = np.column_stack([xx.ravel(), yy.ravel()])
target = (xx * yy).ravel()

model = AdderNetMultiInputLayer(bins=32, interactions="auto")
model.fit(X, target)

print(model.predict(7, 9))  # approximately 63
print(model.predict_batch([[2, 3], [4, 5]]))

Build and test

python -m pip install -e .
addernet-selftest
python -m pytest

Set ADDERNET_OPENMP=1 before building only after benchmarking on the target hardware. The default single-threaded native loop is often faster for this LUT.