The course
Machine learning on a microcontroller is an expert Coding course in seven lessons. The task is a wristband that tells a wave, a punch and a circle apart from an accelerometer. There is no numpy and no library: every model is written in plain Python, so you see every number. The circuit is the gallery's traffic light build, one lamp per gesture and a button to make one. Every step is checked against what the Pi prints or what its lamps do, and every number in the lessons is one the program really prints.
- Data and features: raw (ax, ay, az) traces, six features (mean and standard deviation per axis), class averages that show where the mistakes will be, a shuffled 66/24 train and test split, and z-scores from the training set only.
- k-nearest neighbors: distance and votes, 21 of 24 at k = 3, choosing k without peeking at the test set, a confusion matrix, why units wreck an unscaled model, and why storing the whole training set is costly on a chip.
- Decision trees: the best first question by Gini impurity, a depth-2 tree printed as rules (22 of 24), overfitting made visible with three mislabeled gestures, and the tree printed as C if statements.
- A neuron: a weighted sum and a step, weights set by hand, the perceptron rule learning them, linear separability and where one neuron gives up, and a lamp fired by its output.
- Training on the Pi: softmax and cross-entropy, gradient descent with the loss falling epoch by epoch, learning rates too small and too big, and a 6-4-3 neural network trained by backpropagation (43 parameters).
- Weights in C: the z-score constants and weights printed as C arrays, int8 quantization with one scale, a check that the int8 model agrees with the float model on every test gesture, and the C classify function.
- Boss: gesture recognition: the finale.
The finale
The boss trains on boot, then classifies a fresh gesture on every press of the button and lights green for a wave, red for a punch and amber for a circle. It shows a confident mistake (a circle called a wave at 92 percent), adds a confidence threshold that answers "not sure" to a jiggle it was never taught, and measures itself honestly on 30 gestures from seeds no part of training ever used: 28 of 30.
What is different here
- The sensor. The catalog kit is an ESP32 with an MPU6050. The Pi gallery
has no accelerometer circuit, so the course reads its gestures from
gestures.py, a short, clearly labeled model of a wrist that makes MPU6050-shaped traces with per-person size, speed, wrist tilt and sensor noise. The algorithms and the exported C are exactly what the real wristband would use; the course says so where it matters. - The engine. The Pi runs MicroPython 1.29: no numpy, so vectors are
lists;
randomhas one generator, sorandom.seedmakes every run the same; dicts do not keep order, so classes are list indexes. - The button bounces like a real one, so the boss uses gpiozero's
bounce_time. - No C compiler on the Pi: the C arrays and function are printed for you to paste into a sketch, and the course is honest that it stops there.