@echo off REM ============================================================ REM BG29 buoc 2 (12/08/2026) - train ShotNet c4a/c4b tren bb9_synth REM 150k voi augmentation "detections thua bi muc tieu". REM VI SAO: BG28 chan doan khe sim->real la PHAN PHOI mat do REM detection bi muc tieu (cu 11 = 1.01 det khong-cue/frame, cu 12 REM = 3.09, synth train 95-96%% visibility ~ 5.3 det/frame), KHONG REM phai kien truc hay input path (slot-builder da duoc minh oan REM bang thi nghiem cach ly 30 cu, lech 3.83 deg / 3.3%%). REM c4a = c3 + dropout slot bi muc tieu (keep ~U[0.10,0.95] tren REM 75%% cu, cue KHONG dropout) REM c4b = c4a + reweight loss ve lat target (cham dau >=0.3s, REM 13.62%% train, w=3.0 - weight loss chu KHONG oversample REM de giu nguyen so buoc moi epoch nhu c3) REM Kien truc/optimizer/ngan sach buoc = y het c3 (194.8 phut moc). REM NGAN SACH CHOT TRUOC: dung 2 config. Muon config thu 3 phai hoi. REM Checkpoint chon theo VAL (augmentation TAT khi val - so val phai REM so duoc voi c3). Ket qua -> poolcoach-rl\models\shotnet_20260812_c4?\ REM NGHIEM THU o run_shotnet_eval_p2prime.bat: gate P2' tren lat REM >=0.3s held-out (n=669) - bar 2 deg / 10%% / 80%% / 80%% VA hon moc REM c3 tren chinh lat do (3.03 deg / 9.73%% / 69.7%% / 65.3%%). REM ============================================================ set PYTHONUTF8=1 call "D:\Khoa luan\poolcoach-cv-env\Scripts\activate.bat" cd /d "D:\Khoa luan\poolcoach-rl" python -u scripts\broadcast\train_shotnet.py --config c4a python -u scripts\broadcast\train_shotnet.py --config c4b pause