Embedded Artificial Intelligence (AI) is when AI inferencing is done at the endpoint, rather than on a server in the cloud.
The Renesas RZ/G FOSS (Free and Open Source Software) AI Board Support Package (BSP) is a collection of Yocto meta-layers that enable a number of popular Open Source AI frameworks to run natively on the RZ/G2 series of reference platforms. This allows users to test AI models directly on an embedded platform using Arm Cortex CPU and NEON cores.
Currently, supported frameworks are ArmNN, ONNX Runtime and TensorFlow Lite.
The AI BSP also includes sample benchmarking applications for each framework that will test the performance of running several popular pre-trained image classification models. The inference timings below are a sub-selection of those provided by the meta-benchmark Yocto layer. The RZ/G2H SoC has the fastest inference timing, followed closely by RZ/G2L.
The source code for the meta-layers is published on GitHub: https://github.com/renesas-rz/meta-renesas-ai
As with any open source project, code contributions and pull requests are welcome.
Shopping Basket Demo Application
Using the RZ/G AI BSP as a base, Renesas has developed a mock-up shopping basket demo application that uses Machine Learning (ML) to identify items in a shopping basket.
This demo uses the TensorFlow Lite AI framework to process a custom MobileNet v2 SSD model trained to identify 10 common shopping items. The application then adds up the cost of each item and provides a total, mimicking a “smart checkout” that is efficient and reduces waiting time in the retail market.
Download all source code from GitHub: https://github.com/renesas-rz/meta-renesas-ai-demos