AareLC_train/experiments/looptype

Loop-type classifier, first test (2026-09-29)

Question: can a small classifier tell the loop type, and above all a broken loop, from the frame? Data: the 364 labelled production frames whose reviewer picked a loop type in the annotator (nylon 123, broken 97, litholoop 88, mitegen microloop 30, micromesh 26). Method: train.py (a copy of AareLC_studio/src/pipeline/looptype/train.py): ResNet-18 from ImageNet weights, 256 px crop around every labelled object widened by 40 %, five cross-validation folds that never split the frames of one sample, class-weighted loss. Result: looptype_head_2026-09-29.json.

accuracy, five classes	0.849
broken: precision / recall	0.88 / 0.95
broken vs rest AUROC	0.985 overall, 0.993 within PXII, 0.81 within PXIII (4 broken frames only)
broken recall at 90 % precision	0.95
weakest classes	micromesh and mitegen microloop, confused with each other (26 and 30 frames)

Caveat: 93 of the 97 broken tags come from PXII, so the PXIII number rests on four frames; the next tags to collect are broken loops from PXIII.