4836929bc7
Steps 5 and 6: pre-trained M2 passes the criterion for Obstruction, not for Visibility; episodes and sensors (decode.py)
main
maximeandClaude Opus 5.52026-11-17 11:59:49 +01:00
5337dd0dac
Step 5: pre-training on the rules' outputs (pretrain.py, m2.py --pretrain); regressions as detectors/bench/compare.py; the reader from the tools repository
maximeandClaude Opus 5.52026-11-17 09:51:56 +01:00
7312a1c3ca
M2 results: no better than M1 on stations seen, worse on a station left out
maximeandClaude Opus 5.52026-11-16 20:54:46 +01:00
fef30b5ef4
Step 4: M2, multiple-instance model in PyTorch (shared sensor MLP, attention per family, averaged seeds); the time-split report shared with M1
maximeandClaude Opus 5.52026-11-16 19:22:53 +01:00
975ebc7c1a
The criterion: stations seen, cut in time; M1's regressions station by station; no regression when nothing is flagged
maximeandClaude Opus 5.52026-11-16 17:59:06 +01:00
f93cdf3cd1
Step 3: M1, gradient boosting on station aggregates (leave one station out, and a time split)
maximeandClaude Opus 5.52026-11-16 17:48:11 +01:00
c04cbc90bb
Step 2: the evaluator (the measure of bench/score.py on predictions per instant; the rules reproduce it exactly)
maximeandClaude Opus 5.52026-11-16 16:25:03 +01:00
b6eecf3ab7
Step 1: the dataset (features per sensor and per station, families per instant as the benchmark)
maximeandClaude Opus 5.52026-11-16 15:48:51 +01:00