@echo off REM ============================================================ REM PoolCoach - BC v3 classification head, 20/07/2026 (double-click) REM REM Boi canh: canonical fix duoc phi (MSE 0.026) nhung V0/side/vert REM van dinh variance baseline o ca 5 config -> spin bi nghien ve 0. REM v3: label (V0,side,vert) nam tren grid 10x5x5 = 250 combo roi rac REM -> CLASSIFICATION (CE khong trung binh hoa mode), phi giu regression. REM REM Step 1: train classifier + offline oracle-check REM + EVAL TRUC TIEP 1000 cu trong env <-- KET QUA CHINH REM Step 2: distill policy classifier -> SB3 zip (can cho --init-from) REM Step 3: eval 1000 cu ban distill -> gate G1 chinh thuc REM REM Doc ket qua: REM - Tang 2 (step 1): Q|pot > 0.65? -> classification pha duoc REM multi-modality hay khong, khong dinh nhieu distill REM - Step 2: val MSE per-component cua distill co THOAT baseline? REM (fail o day = 64x64 khong bieu dien noi policy, van de RIENG) REM - Step 3: gate G1 tren ban SB3 (Q|pot > 0.65, pot >= 10%) REM ============================================================ if "%~1"=="_go" goto :inner cmd /k "%~f0" _go exit /b :inner set "VENV=D:\Khoa luan\poolcoach-env\Scripts\activate.bat" set "REPO=D:\Khoa luan\poolcoach-rl" if not exist "%VENV%" ( echo [loi] Khong tim thay venv: "%VENV%" goto :eof ) call "%VENV%" cd /d "%REPO%" echo. echo ===== Step 1/3: BC v3 classifier + eval truc tiep 1000 cu ===== python scripts\train_bc_v3.py --run-name bc_v3_20260720 echo. echo ===== Step 2/3: Distill -^> SB3 zip ===== python scripts\train_bc.py --dataset data\bc_dataset_10000_124_v3distill.npz --run-name bc_v3_20260720_distill echo. echo ===== Step 3/3: Gate G1 ban distill (eval 1000) ===== python scripts\eval_position.py models\bc_v3_20260720_distill\bc_model.zip --episodes 1000 --aim-mode any echo. echo ============================================================ echo XONG. Neu tang 2 PASS ma step 3 FAIL -^> van de la distill echo (thu --hidden lon hon o train_bc? / net_arch PPO), KHONG phai echo classification. Neu tang 2 cung FAIL -^> multi-modality chua echo giai quyet duoc bang class head -^> ban tinh tiep sau. echo ============================================================