Skip to content

README.md

Latest commit

 

History

History
97 lines (73 loc) · 3.65 KB

README.md

File metadata and controls

97 lines (73 loc) · 3.65 KB
Raw

HypercubeCNN

Build wheels

Python bindings for HypercubeCNN — a dependency-free convolutional neural network whose feature map is a Boolean hypercube. Choose a dimension DIM; each channel then lives on exactly N = 2^DIM vertices (for example DIM 6 → 64 sites, DIM 10 → 1024). A local filter at a vertex reaches only that site and its nearest neighbors, and every neighbor index is a single XOR on the binary address — no spatial grid, no adjacency list, no stencil table to store. The activations stay ordinary real-valued units (ReLU, tanh, …); only the topology is binary, so capacity is power-of-two by construction and packing non-cube data is host work.

You stack local layers, train for classification or regression, and save weights with a small architecture sidecar. Optional helpers map images onto length N; data that already lives at 2^D (reservoir or ESN state, fingerprints, product features) can drive the network with no packing step.

Compared with a standard vision CNN, the shared-weight stack and end-to-end training stay familiar — only the domain changes. A usual network slides a window on a rectangle and pads the borders; here every site has the same neighbors under the cube’s symmetry, with no image edge and no stencil table. Pixels are not native: pack them onto the N sites first, then train on those length-N inputs.

Installation

pip install hypercube-cnn

Pre-built wheels for Python 3.10–3.13 on Windows (x64), Linux (x86_64, aarch64), and macOS (x86_64, arm64). No compiler required.

From source

git clone https://github.com/dliptak001/HypercubeCNN.git
cd HypercubeCNN
pip install .

Requires Python 3.10+, a C++23 compiler, and CMake ≥ 3.21. On Windows with CLion’s bundled MinGW, put that toolchain’s bin (and Ninja) on PATH, set CMAKE_GENERATOR=Ninja, and point CC/CXX at the MinGW gcc/g++ — exact install paths change with the CLion version. Then:

pip install . --no-build-isolation --force-reinstall --no-deps

Quick start

import numpy as np
import hypercube_cnn as hc

net = hc.HCNNConfig(
    dim=6,
    num_outputs=3,
    layers=[
        hc.LayerSpec.conv(8, bn=True),
        hc.LayerSpec.pool("max"),
        hc.LayerSpec.conv(8),
    ],
    weight_seed=1,
).build()

x = np.random.randn(net.N).astype(np.float32)  # full capacity N = 2**dim
logits = net.predict(x)
cls = net.predict_class(x)
net.train_step(x, target=0, params=hc.TrainParams(learning_rate=1e-3))
net.save("model")  # model.hcnw + model.arch.json

Features

  • Core train/infer — classification (CE) and regression (MSE); NumPy float32
  • Architecture product — LayerSpec / HCNNConfig, export/import arch JSON
  • Model I/O — HCNW weights + arch sidecar (C++ interop); pickle as secondary
  • Spatial pack — SpatialEmbedder / SpatialAugmenter for H×W → length N
  • Train helpers — evaluate_classification / evaluate_regression, cosine_lr
  • Contracts — capacity input_channels * 2**dim; after packing, pass length-N inputs

Documentation

Full API reference: docs/Python_SDK.md

C++ contracts: docs/CPP_SDK.md

In-repo recipes: examples/python/

Project repository: github.com/dliptak001/HypercubeCNN

License

Apache-2.0