TinyEngine
TinyEngine is a memory-efficient, high-performance neural network library designed for microcontrollers. It serves as the inference and training engine for MCUNet, a system-algorithm co-design framework co-developed with TinyNAS for tiny deep learning on IoT devices. The library enables deep learning workloads under tight memory constraints, supporting both inference and on-device training with as little as 256KB of memory. Key features include a memory-optimized operator library for microcontroller inference, patch-based inference for handling larger models, a tiny training engine for on-device learning, and code generation tools that produce standalone C code deployable on platforms like OpenMV Cam H7 and STM32 boards. It supports applications such as visual wake words detection, person detection, and face mask detection. Developed at MIT Han Lab, TinyEngine is backed by publications including MCUNet at NeurIPS 2020, MCUNetV2 at NeurIPS 2021, and MCUNetV3 at NeurIPS 2022. It targets edge AI, embedded system