This video is a live webinar providing a comprehensive walkthrough of the Edge Impulse platform for building and deploying on-device machine learning models. Led by Shrea, a Qualcomm software engineering intern, the session begins with a simple, hands-on computer vision project to detect a pen. This initial demonstration covers the entire user-friendly workflow, including effortless data collection using a smartphone, creating an "impulse" with a MobileNetV2 architecture, training the model, and deploying it in real-time on a Qualcomm Rubik Pi.
The presentation then transitions to a more advanced "expert mode," showing how developers can directly edit the Python training script to implement custom logic like early stopping. A more complex American Sign Language (ASL) classification project is also demonstrated, highlighting real-world challenges such as lighting conditions and the performance differences between quantized and unquantized models. The webinar concludes with a brief overview of running TF Lite models directly via the Hexagon delegate and using the Qualcomm AI Hub for model conversion, offering a complete guide for developers at all skill levels.
Getting Started with Edge Impulse: A Live Walkthrough for On-Device Machine Learning
0:02:56 What is Edge Impulse?: An introduction to Edge Impulse as a machine learning development platform designed to simplify building, training, optimizing, and deploying models directly onto edge devices.
0:05:06 Effortless Data Collection: The demonstration shows how easily a user can collect a dataset by scanning a QR code with their phone, which instantly connects the device and allows for direct image capture into the project.
0:07:27 Building a Model: Shrea walks through creating an "impulse," which combines a pre-processing block (for image data) and a learning block, selecting the MobileNetV2 architecture for a transfer learning task.
0:12:38 The Importance of Data Augmentation: The presenters explain that data augmentation (e.g., rotating, flipping images) is a crucial step to prevent the model from overtraining and help it generalize better to real-world variations not present in the original dataset.
0:15:36 Live Deployment on Rubik Pi: The trained model is deployed to a Qualcomm Rubik Pi using the edge-impulse-runner. The live demo shows the pen detection model running with very low latency, leveraging the device's built-in NPU for accelerated performance.
0:20:28 Expert Mode for Advanced Customization: The video showcases the "expert mode," which allows developers to directly edit the Python training code to add advanced features like custom callbacks (e.g., early stopping) for more efficient training.
0:26:43 Advanced ASL Classification Demo: A more complex model for classifying American Sign Language gestures is demonstrated, highlighting real-world challenges where accuracy is affected by factors like lighting and model quantization.
0:38:31 Key Takeaway - Data is Critical: The speakers emphasize that data curation is one of the most critical aspects of any ML project. The quality and representativeness of the dataset (e.g., matching lighting conditions) have a massive impact on the model's real-world performance.
0:40:11 Running Models Without the Runner: A brief look at Python code that runs a TF Lite model directly on the device using the Hexagon delegate, providing an alternative to the Edge Impulse runner for integration into custom applications.
0:42:06 Introduction to Qualcomm AI Hub: The webinar briefly introduces the Qualcomm AI Hub as a resource for developers to convert models from different frameworks (like ONNX) into a format compatible with Qualcomm's hardware.