On-Device Machine Learning

10 Lektionen in dieser Stufe

1. Training Versus Inference and What Fits on a Pi
   Explain model size, operations and memory limits on edge hardware. · Lab: Estimate from model cards whether three models can run at five frames per second.
2. Image Classification with a TensorFlow Lite Runtime
   Load a model, preprocess a frame and read the top results. · Lab: Classify live camera frames with a pre-trained MobileNet and print the top three labels.
3. Object Detection and Reading Bounding Boxes
   Interpret detector outputs and apply thresholds and non-maximum suppression. · Lab: Run a pre-trained detector and draw labelled boxes on the preview.
4. Quantisation: Trading Accuracy for Speed
   Compare float and 8-bit integer models. · Lab: Benchmark float32 and int8 versions of one model for speed and accuracy.
5. Collecting and Labelling Your Own Dataset
   Build a balanced, well-labelled image set. · Lab: Capture 300 images of your target in varied conditions and label them.
6. Transfer Learning on a PC and Exporting to the Pi
   Fine-tune a small detector and convert it for edge use. · Lab: Train on a laptop or free notebook service, export an int8 model and run it on the Pi.
7. Running ONNX and NCNN Models
   Use alternative runtimes and convert between formats. · Lab: Export a model to ONNX and NCNN and compare speed across three runtimes.
8. Accelerating Inference with an AI HAT
   Offload a compiled model to a neural processing unit over PCIe. · Lab: Install the AI HAT, run a compiled detection model and compare frame rate with the CPU.
9. In-Sensor Inference with an AI Camera
   Run a network inside the camera and receive results as metadata. · Lab: Deploy a detection model to an AI camera and read detections from frame metadata in picamera2.
10. Evaluating a Model: Accuracy, Latency and Heat
    Measure precision, recall, latency percentiles and thermal behaviour. · Lab: Score your model on a held-out set and run a thirty-minute soak test logging temperature.

Abschlussprojekt

A detector trained on your own images that recognises a chosen object class, publishes events over MQTT and comes with a measured accuracy, latency and temperature report.

Teile für diese Stufe

• 1 × Raspberry Pi 5 (8 GB recommended) set from Level 1
• 1 × Camera module with autofocus
• 1 × AI accelerator HAT for Raspberry Pi 5 (M.2 NPU type)
• 1 × AI camera module with on-sensor inference (IMX500 type)
• 1 × Active cooler
• 1 × Camera mount or tripod

In jeder Lektion legst du die passenden Teile direkt in den Warenkorb. Raspberry Pi →