Quantise the ONNX export, and record how the model performs
Aryan Sethi+1: Claude Opus 5 (1M context)
35206ab
METRICS.md collects what the fine-tuned model actually does: mAP per class, the
confidence operating point, distance accuracy by range and by scene, and speed
per export format. Every row names the script that produced it.
Completed the ONNX quantisation sketch into a working one -- it referenced an
undefined preprocess, an undefined image list, and a file that did not exist.
Two settings in it each changed the outcome completely, and both failed
silently, with the model loading and running either way.
Quantising the whole graph does not degrade this model, it destroys it: 213
detections to 0. The head decodes boxes into pixel coordinates spanning 0 to 640
and runs a softmax over the DFL bins, and one INT8 activation scale cannot carry
that range. Excluding the 85 nodes within twelve steps of the output restores it
to 214 detections at the same mean confidence.
Activations must be unsigned on x64. QInt8 activations run at 53.1 ms against
QUInt8's 23.9 -- slower than fp32, which defe…
