No Hardware? No Problem: Simulating TensorFlow Lite Targets with Renode

Continuous integration and delivery tools have experienced widespread adoption by traditional software developers thanks to tools like Travis CI, CircleCI, and Concourse. While solutions are being developed to bring this productivity-enhancing workflow to the hardware world, the ability to create a complete pipeline can be complicated by the need to incorporate actual hardware targets to execute and test code on. Renode, an open-source project that spun out of and is commercially supported by Antmicro, allows simulation of Arm and RISC-V hardware, bringing the rapid development and iteration that most software developers now take for granted to the world of IoT. When the TensorFlow Lite team themselves needed to accelerate development by automating testing of Arm and RISC-V platforms, Antmicro and Renode made this possible, without the need for physical hardware. And now the Antmicro/Renode and TF Lite teams are sharing their productivity breakthroughs with the community.

In a guest blog post on TensorFlow, Michael Gielda of Antmicro shares an example using the popular Magic Wand demo. Renode simulates hardware by translating machine code from the target to the host's own native machine language. I/O and sensors are simulated accurately by emulating their register interfaces and mimicking their real-world behavior. This also means that sensors and I/O can be scripted, for example using pre-recorded data to input gestures into the virtual devices. And of course, when you're ready, Renode supports deployment to actual hardware, including a wide array of FPGAs, STM32s, and more — even the diminutive Fomu board!

Hello World and Magic Wand examples are available in the accompanying GitHub repo, and Antmicro promise follow-up blog posts in coming months to create a complete example of leveraging Continuous Integration with TensorFlow Lite on the Zephyr RTOS.