onnx-quantize-tflmlisted

# Model to microcontroller Two things go wrong on this path. The model does not fit, which is a budget problem you can answer in ten minutes. Or it fits and the accuracy quietly drops, which is a measurement problem — and usually a calibration-data problem wearing a disguise. ## Budget first Write down four numbers before converting anything: flash available, RAM available for the tensor arena, latency allowed per inference, and the accuracy you are willing to lose. A model that cannot meet these does not get better with quantisation tuning, and finding that out now costs ten minutes instead of a week. The arena has to hold the largest set of simultaneously live tensors, which is usually driven by the widest early activation rather than by the weights. A model that fits flash comfortably can still miss RAM by a factor of three. ## Quantising ONNX ```bash python3 scripts/quantize_onnx.py model.onnx -o model.int8.onnx \ --calib data/calibration --compare ``` The script preprocesses (shape inference and folding, skipping which is the usual cause of "quantisation ran and the model got slower"), calibrates on real samples, quantises to QDQ format with per-channel weights, and then reports max absolute difference, cosine similarity and argmax agreement against the float model. That last part is the point: a quantised model without a measured accuracy delta is a guess. For a model that loses too much at int8, `--activation-type uint16 --weight-type int8` recovers most of