Training and deploying to several targets

Training and deploying to several targets · Course page

Prepare a balanced dataset, train our own model in Docker, squeeze it into int8, then take one single file and run it on the web, Cortex-A and an MCU through Vela, while measuring parity.

Module goal

Build a model of our own, from data all the way to the chip, and be able to measure exactly what trade-off each target makes.

Lessons

Lesson	Topic	Time (min)	Slides
5.1	Dataset engineering: class balance, windows and the train/val/test split	65	slides.md
5.2	Hands-on: recording a balanced dataset on the board, then splitting it on the PC	75	slides.md
5.3	Training a model in Docker: one artifact, four targets	65	slides.md
5.4	Inside training: Keras, Conv1D, gradient descent, int8 and the confusion matrix	70	slides.md
5.5	Hands-on: fill in a training script and run it in Docker	75	slides.md
5.6	Running a model on the web: LiteRT.js, int8 I/O and parity	70	slides.md
5.7	Hands-on: matching the web’s verdict to the PC’s, and the Cortex-A story	75	slides.md
5.8	Quantizing and Vela: getting our model onto the Ethos-U55	70	slides.md
5.9	Hands-on: comparing three targets — MCU, web and PC	75	slides.md

Lessons come in pairs: a concept lesson followed by a hands-on lesson with a practice file, a solution, and a lab.

Module checkpoint

You pass this module once you can do all of the following (details are in the Lab section of each hands-on lesson):

• A clean, balanced, split dataset from the board — every set (train/val/test) has all three classes in close-to-equal proportion (lesson 5.2).
• Successfully train a Keras model in Docker, getting a report of float32 accuracy, int8 accuracy and a confusion matrix on a test set the model has never seen, along with the model_int8.tflite and .norm.npz files (lesson 5.5).
• The web file’s verdict matches the PC side within tolerance (max|score_pc − score_web| ≤ TOL, and the winning class matches), and can explain why they don’t need to match bit for bit (lesson 5.7).
• A comparison table of three targets (MCU, web, PC) with real numbers, explaining why accuracy matches but latency differs, and why the MCU needs Vela (lesson 5.9).