TorchLogix Documentation

TorchLogix is a PyTorch library for differentiable logic gate neural networks. It extends the original DiffLogic work with enhanced features, improved usability, and comprehensive documentation.

Note

TorchLogix builds upon the foundational work of DiffLogic by Felix Petersen et al. This library provides extensions and improvements while maintaining compatibility with the original concepts.

Key Features

  • Differentiable Logic Gates: Implement neural networks using 16 different binary logical operations

  • Convolutional Support: 2D and 3D logic convolutional layers with flexible receptive fields

  • Model Compilation: Convert trained logic networks to optimized implementations

  • CUDA Acceleration: Optional CUDA extensions for high-performance computing

  • Easy Integration: New layers mostly follow PyTorch conventions

Quick Start

import torch
from torchlogix.layers import LogicDense, LogicConv2d
from torchlogix.models import CNN

# Create a simple logic dense layer
layer = LogicDense(in_dim=784, out_dim=128, tree_depth=3)

# Create a logic convolutional layer
conv = LogicConv2d(
    in_dim=(28, 28),
    num_kernels=16,
    tree_depth=3,
    receptive_field_size=5
)

# Use pre-built models
model = CNN(class_count=10, tau=1.0)

Documentation Contents

User Guide

API Reference

Development

Indices and Tables