poolcoach / scripts /experiments /run_bc_v3_20260720.bat
masterdanh's picture
deploy: snapshot for HF Space
78738de
Raw
History Blame Contribute Delete
2.26 kB
@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 ============================================================