Magnetron is a minimalistic, C99 and Python-based deep learning framework designed for learning and research, offering a PyTorch-like API with automatic differentiation and multithreaded CPU compute. It targets developers and researchers seeking a transparent and modifiable alternative to larger frameworks.

How It Works

Magnetron leverages a C99 core for performance-critical operations, including SIMD-optimized tensor computations (SSE4, AVX2, AVX512, ARM NEON), and provides a modern Python API for ease of use. Its design emphasizes a dynamic computation graph for eager evaluation and includes features like broadcasting, in-place operations, and high-level neural network building blocks.

Quick Start & Requirements

• Install: Clone the repo, cd magnetron/python, pip install -r requirements.txt (for examples), then cd magnetron_framework && bash install_wheel_local.sh.
• Prerequisites: Linux, macOS, or Windows; C99 compiler (gcc, clang, msvc); Python 3.6+.
• Dependencies: CFFI (for Python bindings). Matplotlib and NumPy are optional for examples.
• Demo: python examples/simple/xor.py
• Docs: Explore the docs »

Highlighted Details

• Multithreaded CPU backend with dynamic scaling and thread pooling.
• SIMD optimized operators (SSE4, AVX2, AVX512, ARM NEON).
• Modern Python API with broadcasting and in-place variants.
• Dynamic computation graph (eager evaluation).

Maintenance & Community

Developed by a single person in their free time; currently a Work in Progress (WIP). No community links (Discord/Slack) or roadmap details are provided in the README.

Licensing & Compatibility

Distributed under the Apache 2 License. Permissive for commercial use and closed-source linking.

Limitations & Caveats

The project is in its early stages (WIP) with many features missing and is not yet fully optimized. GPU compute (CUDA), low-precision datatypes, and distributed training are planned but not yet implemented.