Embedded software engineers who have struggled with Machine Learning (ML) model creation, testing and deployment will now be able to seamlessly execute all these steps, and much more, on our SAMA7G54 Arm® Cortex®-A7 based 32-bit microprocessor (MPU), using the Edge Impulse Platform.
Accelerate AI Adoption with Edge Impulse
To accelerate the adoption of ML and artificial intelligence (AI) at the edge, Microchip’s 32-bit MPU is fully integrated into the Edge Impulse platform designed to provide Software engineers with an easy-to-use device to develop AI/ML models.
This blog post provides an overview of this innovative solution that leverages the SAMA7G54 and Edge Impulse platform.
SAMA7G54, Arm Cortex-A7 32-bit MPU
The SAMA7G54 is a single-core Arm® Cortex®-A7 based microprocessor that can run up to 1GHz. It features various user interfaces like a Parallel 14-bit Interface Image sensor controller, a MIPI CSI-2® Camera interface and an audio subsystem.
This microprocessor also integrates advanced security features like hardware cryptography elements (AES/TEDS/SHA/RSA/ECC).
To learn more about our Cortex-A7 based microprocessors, please visit our SAMA7 MPU web page.
Overview of Edge Impulse Platform
Edge Impulse is a company that provides an easy-to-use toolset that allows developers to train, evaluate and deploy ML models on embedded targets; the main steps are illustrated below:
Figure 1: Illustration of all the steps enabled by Edge Impulse
To be more precise, most of these steps are done thanks to the Edge Impulse Studio, an online platform that is very visual, and therefore well suited for both beginners and experts in AI/ML:
Figure 2: Snapshot of the Edge Impulse Studio
The first key feature is the data auto-labeller for object detection projects. This tool can automatically locate objects with similar features (form, colours, etc.) through the different images of your training dataset, and therefore you can label it quicker.
To learn more about this feature, feel free to visit Edge Impulse Documentation.
After successfully creating and preprocessing your data, you are then ready to construct your model architecture, which encompasses the input, output, processing modules and learning blocks.
After this stage, you may proceed to train various models, using a range of parameters such as:
Number of Epochs
Learning rate
Batch Size
Data Augmentation
Number of neurons for the final layer
Dropout rate
There are advanced features the further assist developers including the EON Tuner.
This tool is designed to evaluate a variety of architectures and train multiple models, considering specific settings and hardware prerequisites. It systematically examines the input, assessing various potential signal processing blocks and neural network frameworks. Ultimately, you will be able to select the model that best aligns with your requirements.
In the image below, you can see that the Eon Tuner has been used for the SAMA7G54, with a targeted inference time of 100ms. Upon completion of the process, many different models are proposed, and the view is adaptable depending on the most important criteria—for example, accuracy or inference time.
Figure 3: Brief overview of the EON Tuner
Upon successfully optimizing your model and ensuring it meets your satisfaction, you may proceed with deployment. This can be accomplished either by directly linking the target to your Edge Impulse project or by downloading the model onto your host computer and subsequently transferring it to your target:
Figure 4: Edge Impulse Studio - Deployment
With the Edge Impulse platform, developers can easily build applications based on computer-vision at the edge, signal and audio processing and so many other use cases.
As demonstrated, this tool is well suited for both experts and beginners in AI/ML.
Getting Started
To learn more about how to use the SAMA7G54 with Edge Impulse, review the Edge Impulse Documentation.
To test the tool, please visit the Edge Impulse site.
Otherwise, if you need more information, or if you would like to discuss your AI/ML project, feel free to reach out to me at hakim.cherif@microchip.com.
Hakim Cherif, Sep 12, 2024
Tags/Keywords: AI-ML
Comentários