#!/bin/bash
# Held-out test (fixed before the pilot): the 200 reserved fresh scenes (seed 9100003), anti-aliasing off,
# eval_policy.py --fast (checked identical to the normal evaluation).
#   primary:   RL checkpoint with the best validation success vs the baseline, both with the action
#              expert in float32 (the precision RL trains and runs it in)
#   secondary: the baseline as always evaluated (expert rounded to bfloat16 on load)
cd /root/play_machanize
export MUJOCO_GL=egl OPENBLAS_NUM_THREADS=1 PHASE2_PROFILE=baseline SIM_OFFSAMPLES=0
R=sim/reports/rl/heldout; mkdir -p $R
RL=${RL_CKPT:-outputs/rl_pilot/best}
E=".venv/bin/python sim/eval_policy.py --fast --episodes 200 --workers 4 --seed 9100003 --video-episodes 2"
$E --expert-fp32 --policy models/baseline --out $R/baseline_fp32 > logs/heldout_baseline_fp32.log 2>&1 &
$E --expert-fp32 --policy $RL --out $R/rl > logs/heldout_rl.log 2>&1
wait
$E --policy models/baseline --out $R/baseline_native > logs/heldout_baseline_native.log 2>&1
.venv/bin/python sim/paired_compare.py baseline_fp32=$R/baseline_fp32 rl=$R/rl --out $R/paired.md
.venv/bin/python sim/paired_compare.py baseline_native=$R/baseline_native rl=$R/rl --out $R/paired_vs_native.md
echo HELDOUT_DONE