On-Device Augmented Image Training and Inference utilizing the TinyEngine Library on a STM32F746G-DISCO Microcontroller
Published in Journal, 2023
This paper presents Tiny Engine, an optimization-driven inference library tai- lored for microcontrollers, which facilitates on-device training and inference in resource-constrained environments. We demonstrate the effectiveness of Tiny Engine through a series of benchmarks, highlighting significant improvements in computational efficiency and cost-effectiveness compared to traditional cloud- based and mobile AI platforms by training and fine-tuning an augmented image data set live on a microcontroller with a camera. Our approach leverages a combi- nation of in-place depth-wise convolution, patch-based inference, and advanced data augmentation techniques, which collectively enable the execution of sophis- ticated neural network models on edge devices. We validate our model using a dataset augmented through transformations, showcasing the potential for broader application in real-world scenarios.
Recommended citation: Zehao Zhao, Javi Ocampo.(2023). On-Device Augmented Image Training and Inference utilizing the TinyEngine Library on a STM32F746G-DISCO Microcontroller. DOI: 10.13140/RG.2.2.34074.21447. http://zehao-zhao.github.io/me/files/TinyML_Writeup.pdf