ScaleDF_rebuttal/launch_fixed500k_resampled_eval_host_queue.sh

#!/usr/bin/env bash
set -euo pipefail

if [[ $# -ne 4 ]]; then
  echo "Usage: $0 <train-launch-tag> <eval-launch-tag> <real-level> <fake-level>" >&2
  exit 2
fi

TRAIN_TAG="$1"
EVAL_TAG="$2"
REAL_LEVEL="$3"
FAKE_LEVEL="$4"

PERSIST_ROOT="${PERSIST_ROOT:-/mnt/bn/search-aipf-haoyu/wenhaowang/ScaleDF_rebuttal}"
EVAL_LAUNCHER="${EVAL_LAUNCHER:-$PERSIST_ROOT/launch_rebuttal_eval_flex.sh}"
SEEDS_RAW="${SEEDS:-1 2 3 4}"
SEED_LABEL="${SEED_LABEL:-1_4}"
EVAL_ROOT="${EVAL_ROOT:-$PERSIST_ROOT/eval/fixed500k_resampled_seed${SEED_LABEL}_benchmark_eval_${EVAL_TAG}}"
HOST_NAME="$(hostname)"

export PERSIST_ROOT
export PACKAGE_ROOT="${PACKAGE_ROOT:-/tmp/ScaleDF_repro_20260724}"
export SCALEDF_DATA="${SCALEDF_DATA:-/dev/shm/scaledf_bind_storage/ScaleDF}"
export BENCHMARK_ROOT="${BENCHMARK_ROOT:-/dev/shm/scaledf_bind_storage/Established_benchmarks_eval_names}"
export SKIP_MISSING_BENCHMARKS="${SKIP_MISSING_BENCHMARKS:-1}"
export BATCH_PER_GPU="${BATCH_PER_GPU:-256}"
export NPROC_PER_NODE="${NPROC_PER_NODE:-8}"
export PYTHONUNBUFFERED="${PYTHONUNBUFFERED:-1}"

find_checkpoint() {
  local family="$1"
  local level="$2"
  local seed="$3"
  local pattern="control_fixed500000_${family}_n${level}_classseed${seed}_resampled_*_${TRAIN_TAG}"
  local run_dir
  run_dir="$(find "$PERSIST_ROOT/runs" -maxdepth 1 -type d -name "$pattern" | sort | tail -n 1)"
  if [[ -z "$run_dir" ]]; then
    echo "missing run dir for pattern=$pattern" >&2
    exit 3
  fi
  if [[ ! -f "$run_dir/best_checkpoint.pth" ]]; then
    echo "missing checkpoint: $run_dir/best_checkpoint.pth" >&2
    exit 4
  fi
  printf "%s/best_checkpoint.pth" "$run_dir"
}

eval_complete() {
  local eval_log="$1"
  [[ -f "$eval_log" ]] && [[ "$(grep -c 'Accuracy of the network on' "$eval_log" || true)" -ge 7 ]]
}

run_eval() {
  local axis_short="$1"
  local family="$2"
  local level="$3"
  local seed="$4"
  local checkpoint
  checkpoint="$(find_checkpoint "$family" "$level" "$seed")"
  local tag="fixed500k_${axis_short}_n${level}_seed${seed}_resampled"
  local output_root="$EVAL_ROOT/$tag"
  local eval_log="$output_root/eval.log"
  local lock_dir="$output_root/.eval.lock"

  if eval_complete "$eval_log"; then
    echo "[$(date -u +%Y-%m-%dT%H:%M:%SZ)] eval_skip tag=$tag reason=complete output_root=$output_root"
    return
  fi

  mkdir -p "$output_root"
  while ! mkdir "$lock_dir" 2>/dev/null; do
    if eval_complete "$eval_log"; then
      echo "[$(date -u +%Y-%m-%dT%H:%M:%SZ)] eval_skip tag=$tag reason=complete_after_wait output_root=$output_root"
      return
    fi
    echo "[$(date -u +%Y-%m-%dT%H:%M:%SZ)] eval_wait tag=$tag reason=locked output_root=$output_root"
    sleep 60
  done

  if eval_complete "$eval_log"; then
    rmdir "$lock_dir" 2>/dev/null || true
    echo "[$(date -u +%Y-%m-%dT%H:%M:%SZ)] eval_skip tag=$tag reason=complete_after_lock output_root=$output_root"
    return
  fi

  echo "[$(date -u +%Y-%m-%dT%H:%M:%SZ)] eval_start tag=$tag checkpoint=$checkpoint"
  if OUTPUT_ROOT="$output_root" RUN_NAME="$tag" "$EVAL_LAUNCHER" "$checkpoint" deit_base_patch16_LS benchmarks; then
    rmdir "$lock_dir" 2>/dev/null || true
  else
    rmdir "$lock_dir" 2>/dev/null || true
    return 1
  fi
  echo "[$(date -u +%Y-%m-%dT%H:%M:%SZ)] eval_done tag=$tag"
}

echo "eval_queue_started_utc=$(date -u +%Y-%m-%dT%H:%M:%SZ)"
echo "hostname=$HOST_NAME"
echo "train_tag=$TRAIN_TAG"
echo "eval_tag=$EVAL_TAG"
echo "real_level=$REAL_LEVEL"
echo "fake_level=$FAKE_LEVEL"
echo "eval_root=$EVAL_ROOT"
echo "batch_per_gpu=$BATCH_PER_GPU"
echo "seeds=$SEEDS_RAW"

for seed in $SEEDS_RAW; do
  run_eval real real_domains "$REAL_LEVEL" "$seed"
  run_eval fake fake_methods "$FAKE_LEVEL" "$seed"
done

echo "eval_queue_done_utc=$(date -u +%Y-%m-%dT%H:%M:%SZ)"