experiments/learnability/run_bestfirst_eval.sh

#!/bin/bash
# =============================================================================
# Best-first eval driver (value-free, no MCTS). Sibling of run_policy_only_eval.sh.
#
# Evaluates the SAME checkpoints with bestfirst_search.py (priority = cumulative
# log policy prob; budget = --max-pops, mirrors MCTS max_mcts_nodes=4000).
#
#   eval set:  base minimo_0pt (0.pt), base minimo_1pt (1.pt), and every
#              policy-only finetuned epoch_2.pt (final) + epoch_1.pt (mid).
#   eval:      bestfirst_search.py, extrinsic-95, --max-pops 4000, --timeout 600.
#              Action cap OFF (faithful to MCTS).
#
# Uses all 4 GPUs, one eval per GPU at a time.
# =============================================================================
set -u

LEARN=/datadrive/ayush/home/minimoX/learning
cd "$LEARN" || exit 1

source "$(conda info --base)/etc/profile.d/conda.sh"
conda activate minimo_new
export WANDB_MODE=offline

CKPT0=/datadrive/ayush/home/minimoX/learning/checkpoints/dnn3_bootstrap_bs_800_para/2026-03-12_23-41-21/0.pt
CKPT1=/datadrive/ayush/home/minimoX/learning/checkpoints/dnn3_bootstrap_bs_800_para/2026-03-12_23-41-21/1.pt
TRAIN_BASE=$LEARN/experiments/learnability/train_policy_only
BF_EVAL=$LEARN/experiments/learnability/bestfirst_eval
mkdir -p "$BF_EVAL"

PROBLEMSET=extrinsic-95
BF_MAXPOPS=4000
BF_WORKERS=6
BF_TIMEOUT=600          # per-problem wall-clock budget (s)
GPUS=(0 1 2 3)
NGPU=${#GPUS[@]}

LOG=$BF_EVAL/eval_driver.log
ts() { date +"%Y-%m-%d %H:%M:%S"; }
log() { echo "[$(ts)] $*" | tee -a "$LOG"; }

# ---- Build eval job table (parallel arrays) --------------------------------
E_CKPT=(); E_JSON=(); E_LOG=(); E_LABEL=()
add_eval() { E_CKPT+=("$1"); E_JSON+=("$2"); E_LOG+=("$3"); E_LABEL+=("$4"); }

add_eval "$CKPT0" "$BF_EVAL/minimo_0pt_base_bf_${PROBLEMSET}.json" \
         "$BF_EVAL/minimo_0pt_base_bf_${PROBLEMSET}.log" "base/minimo_0pt"
add_eval "$CKPT1" "$BF_EVAL/minimo_1pt_base_bf_${PROBLEMSET}.json" \
         "$BF_EVAL/minimo_1pt_base_bf_${PROBLEMSET}.log" "base/minimo_1pt"

for CKPT_NAME in epoch_2.pt epoch_1.pt; do
  TAG=$([[ $CKPT_NAME == epoch_2.pt ]] && echo final || echo mid)
  while IFS= read -r ckpt; do
    [[ -z "$ckpt" ]] && continue
    ckpt=$(realpath "$ckpt")
    label=$(echo "$ckpt" | sed "s#.*train_policy_only/##; s#/updates.*##")
    add_eval "$ckpt" \
             "$(dirname "$ckpt")/bf_${PROBLEMSET}_${TAG}.json" \
             "$(dirname "$ckpt")/bf_${PROBLEMSET}_${TAG}.log" \
             "$TAG/$label"
  done < <(find "$TRAIN_BASE" -name "$CKPT_NAME" -path "*updates_*" 2>/dev/null | sort)
done

NJOBS=${#E_CKPT[@]}
log "===== START: $NJOBS best-first evals (max_pops=${BF_MAXPOPS} timeout=${BF_TIMEOUT}s) on GPUs ${GPUS[*]} ====="

k=0
for ((j=0; j<NJOBS; j++)); do
  gpu=${GPUS[$((k % NGPU))]}
  log "EVAL [$k/$NJOBS] ${E_LABEL[$j]} -> gpu $gpu"
  (
    CUDA_VISIBLE_DEVICES=$gpu python -u bestfirst_search.py \
      --agent "${E_CKPT[$j]}" \
      --problemset "$PROBLEMSET" \
      --max-pops "$BF_MAXPOPS" \
      --workers "$BF_WORKERS" \
      --timeout "$BF_TIMEOUT" \
      --output "${E_JSON[$j]}" \
      > "${E_LOG[$j]}" 2>&1
    echo "[$(ts)] EVAL DONE ${E_LABEL[$j]}" >> "$LOG"
  ) &
  k=$((k + 1))
  if (( k % NGPU == 0 )); then wait; fi
done
wait
log "--- All evals complete ---"

for f in $(find "$BF_EVAL" "$TRAIN_BASE" -name "bf_${PROBLEMSET}_*.json" -o -name "minimo_*_base_bf_${PROBLEMSET}.json" 2>/dev/null | sort); do
  python -c "import json;d=json.load(open('$f'));print(d['num_solved'],'/',d['num_problems'],'$f')" 2>/dev/null | tee -a "$LOG"
done