Zhaolan Huang

Freie Universität Berlin - "U-TOE: Universal TinyML On-board Evaluation Toolkit for Low-Power AIoT"


U-TOE: Universal TinyML On-board Evaluation Toolkit for Low-Power AIoT



We developed and released a generic toolkit enabling low-power neural network inference on various boards supported by RIOT. This toolkit, named U-TOE, is fully open source and its architecture is detailed in our preprint, which we recently published. U-TOE is based on a seamless combination of the model compiler TVM with RIOT, and enables a workflow taking as input models from diverse frameworks such as PyTorch, TensorflowLite etc. and producing as output a compressed model binary integrated in a RIOT image which can be flashed on a large variety of boards based on all the popular 32-bit microcontroller architecters Cortex-M, RISC-V or ESP-32. As such, U-TOE allows the evaluation of resource consumption of inference with arbitrary neural networks, locally or remotely on a testbed, on all the popular low-power IoT hardware including Raspberry Pi Pico, nrf52840dk, Arduino Nano, ESP32-wroom-3, Hifive1b, stm32f746g-disco and many more.