truong_stft_cnn: a single-fold CUDA trace, and a real 6-fold run

2026-08-22 · baseline comparison, performance

Two distinct truong_stft_cnn prediction-mode runs happened in the same session, at different scales and for different purposes — worth keeping separate since both are easy to conflate under the same pipeline name.

A fast, fully-instrumented single-fold trace

A debug harness forces exactly one fold and one epoch to get a fast, verbose look at a single training pass rather than a real result. On real CUDA hardware: dataset construction 3.62s, model fit 6.12s (including one-time model build), test prediction 0.79s — training throughput 101.8 windows/sec, test throughput 188.8 windows/sec. CUDA memory after fit: 17.0MB allocated, 214.0MB reserved. That's a genuinely tiny working set, and it's architectural, not incidental: this pipeline runs a single 2D/3D CNN over a compact STFT tensor, not a 253-edge dense coherence graph — which is exactly why it's so much cheaper than dense_edge_gru to begin with, independent of anything about caching or GPU utilization.

A real 6-fold leave-one-seizure-out run

Separately, a full run across the subject's six seizures, real epoch schedule, completed successfully.

seizure	preictal	hit (smoothed)	precision	recall	f1	roc_auc	avg_precision	FAR/hr
03	30	✓	0.181	0.700	0.288	0.875	0.136	12.3
04	30	✓	0.201	0.933	0.331	0.919	0.227	13.9
15	30	✗	0.000	0.000	0.000	0.971	0.400	1.6
16	23	✓	0.259	0.957	0.407	0.980	0.616	3.2
18	30	✗	0.857	0.200	0.324	0.995	0.852	0.0
26	30	✓	0.558	0.967	0.707	0.983	0.619	2.8

Event-level hit rate: raw 5/6, smoothed 4/6 — the same hit/miss pattern (misses on seizures 15 and 18) as the earlier truong_stft_cnn run on the same subject and folds. Per-fold numbers shifted somewhat from that run (seizure 03's precision, for instance, moved from 0.206 to 0.181) but not investigated further here, since the pipeline itself changed materially in between — the torch-native CWT swap and the disk cache removal both landed in that window. The qualitative shape held steady across both runs regardless: seizures 16 and 26 strong, 15 and 18 missing at threshold despite ranking well, 03 and 04 borderline.

Open items

The real 6-fold run has no accompanying verbose log, only result CSVs — if it needs to be reproduced or debugged later, it will need a fresh run. No label-permutation null control has been run yet for any truong_stft_cnn prediction run to date. The precision/recall drift against the earlier run on the same folds hasn't been attributed to a specific one of the intervening changes. Result CSVs from this run are untracked in git.