Training — train once, run everywhere (module 5)

This is where the course stops consuming models and starts making them. You collect a dataset on the board, train a model on your PC in Docker, and deploy the same model to four targets: the MCU (Cortex-M55 + Ethos-U55), the PC, the web browser, and a Cortex-A Linux board.

The one artifact, four targets

                    train.py  (Keras, in Docker)
                        │
                        ▼
              model_int8.tflite   ← the "train once" artifact
                        │
        ┌───────────────┼───────────────┬────────────────┐
        ▼               ▼               ▼                ▼
  quantize_vela.sh   eval_pc.py    convert_web.py   (same file)
  → *_vela.tflite    ai-edge-litert  → browser        ai-edge-litert
  MCU / Ethos-U55    PC / Docker      LiteRT.js/ORT    Cortex-A (RPi/Jetson)

Only the MCU needs the extra Vela compile; the browser and Cortex-A reuse the plain model_int8.tflite. int8 is the common denominator (the NPU requires it; the others accept it). Lessons 5.3 and 5.8 walk through the full matrix.

Files

File	What it does	Where it runs
Dockerfile	reproducible TensorFlow env (works on macOS/Win/Linux)	PC
dataset_tools.py	load / window / normalize / split an IMU CSV dataset	PC
train.py	train a small Conv1D gesture classifier, export int8 .tflite	PC (Docker)
eval_pc.py	run the .tflite on PC via ai-edge-litert, print accuracy + confusion	PC
quantize_vela.sh	compile model_int8.tflite → _vela.tflite for the Ethos-U55	PC → MCU
convert_web.py	make a browser-friendly variant + notes for LiteRT.js/ORT-Web	PC → Web

The dataset is IMU 6-axis windows labelled idle / circle / shaking — the same problem the board’s built-in Motion model solves, so you can compare your model against the shipped one.

Quick start (Docker)

# 1. build the training image (once)
docker build -t edgeai-train .

# 2. train (expects data/gestures.csv captured on the board in lesson 5.2)
docker run --rm -v "$PWD":/work edgeai-train python train.py \
    --data data/gestures.csv --out model_int8.tflite

# 3. check it on the PC
docker run --rm -v "$PWD":/work edgeai-train python eval_pc.py \
    --model model_int8.tflite --data data/gestures.csv

# 4a. MCU:   ./quantize_vela.sh model_int8.tflite       -> model_int8_vela.tflite
# 4b. Web:   docker run ... python convert_web.py model_int8.tflite
# 4c. Cortex-A: copy model_int8.tflite to the Pi/Jetson and run eval_pc.py there

data/gestures.csv is produced on the board by s04_daq_logger.py (lesson 2.2) or s11_dataset.py (lesson 5.2). A tiny synthetic generator is included in dataset_tools.py (--synthesize) so the pipeline runs before you have real data.