STMicro Brings Its Free STM32 MCU Edge AI Toolset to the Cloud
The new cloud-based version includes a GitHub-based model resource, and remote access to a “board farm” to remotely benchmark edge-AI models on STM32 MCU boards.
With a goal of opening its STM32 MCU microcontroller (MCU) AI development resources to a broader community including AI developers and data scientists, yesterday STMicroelectronics (ST) unveiled its STM32Cube.AI Developer Cloud offering.
Engineers familiar to the company’s desktop version, STM32Cube.AI, should be able to embrace the new STM32Cube.AI Developer Cloud online version easily, according to the company. But the real power of the online toolset is to enable people such as AI developers and data scientists—experts who may not have traditionally worked with MCUs and embedded systems—the capability to easily run and benchmark their models in a virtual way.
STM32Cube.AI Developer Cloud shares the core features with STM32Cube.AI, except now users can leverage a web-based interface, and access a board farm of STM32 MCU-based boards.
In this article, we discuss the need that the cloud-based tool satisfies, examine the details of the service’s features, and share insights with our interview with Vincent Richard, ST’s AI product marketing manager.
An Online Web-based Interface
ST’s established desktop STM32Cube.Al was created to enable developers to validate and generate optimized STM32 AI libraries from trained neural networks. Indeed, Richard says that tool continues to see increased download numbers, and those numbers have increased steadily since the tool was launched in 2017. The new cloud-based version is intended to complement the desktop version.
Over STM32Cube.AI Developer Cloud’s online interface, users can generate optimized C-code for STM32 MCUs. No prior software installation is required. According to Richard, the optimization aspect is key. “This is an optimization tool and service that can convert trained AI models into C-code basically that are optimized to run on STM32 microcontrollers,” he says.
Richard says that this online way of working fits well with how AI developers and data scientists traditionally work. “Our aim with this new tool is to address a different category of people that most likely have used AI products and tools that are mainly online,” he says. “The advantages are no installation, no download, direct access to the service using their MyST login, and being able to run this tool to optimize their AI neural networks.”
Model Zoo Provides a Rich Set of Resources
Beyond the online capability, another key feature of STM32Cube.AI Developer Cloud is access to ST’s STM32 Model Zoo. Model Zoo is basically a repository of resources to aid AI development. This includes trainable deep-learning models and demos. Available use cases at launch include human motion sensing for activity recognition and tracking.
Other use cases available are computer vision for image classification or object detection, audio event detection for audio classification, and more. The Model Zoo is hosted on GitHub and it allows the automatic generation of “getting started” packages optimized for STM32.
Hosted on GitHub, STM32 Model Zoo is a repository of optimized models.
Again, for the Model Zoo, Richard stressed the optimization aspects. “Model Zoo is a collection of models that has been selected to run easily in terms of memory footprints, latency and performances on STM32 MCUs,” he says. “On the Model Zoo, users can retrain from a model with their own data and create an application automatically from the Model Zoo.”
The fact that Model Zoom is hosted on GitHub is significant, according to Richard. It also fits in with how today’s engineers and developers work. “On GitHub we have a collection of models, but we also have the training scripts in order to retrain the models,” he says.
Stepping through how it works, Richard uses an example. “For instance, let’s say I would like to have an object detector,” he says. “We provide an application nutshell, where we use one of our ‘identify’ models trained with a public data set. The user can then go into the tool, clone the GitHub repository, create their own object detector based on its own model train on its data set, and generate the application that is corresponding to this particular application.”
“A user has no need to invent from scratch the whole application. He only needs to have its data, use the models, and generate the AI application.”
A Board Farm Accessible Online
An important feature of ST’s STM32Cube.AI Developer Cloud—and one the company is calling an industry first—is a capability to access ST’s online “board farm.” This means remote access to STM32 MCU-based development boards, so that users can evaluate and test the performance of AI models that have been created through the online application.
Using the cloud-accessible board farm, data scientists and developers can remotely measure the actual performance of the optimized models. (Click image to enlarge)
According to Richard, ST is starting with ten boards—ten part numbers of boards. “Users can just ask the service to give them the results on the performance and latency based on the real hardware remotely,” he says. “We will also give users some scripts to simplify the machine learning flow of the creation of models that are implemented by data scientists.
Richards says they will be able to just copy and paste those scripts to retrain and deploy the models onto the designated hardware. Generally speaking, the STM32Cube.AI Developer Cloud is designed to support all versions of the STM32 MCU, from the C0 to H7. But, in the initial board farm set, only a limited set of MCU boards will be available, including H7 and L4 STM32 MCU models, according to Richard.
The company says the online tool has already been undergoing testing and evaluation by select embedded development customers. STM32Cube.AI Developer Cloud is available now for free to registered MyST users.
It’s an Online World for Engineers and AI
It’s become clear that AI has a powerful role to play in edge applications. But the edge has been more the comfort zone of embedded systems engineers used to working with MCUs and constrained embedded software situations.
New online tools like ST’s STM32Cube.AI Developer Cloud perhaps open the door for AI developers and data scientists that lack that embedded systems’ knowledge to more easily collaborate and bring AI to the edge.
All image used courtesy of STMicroelectronics
