#!/bin/bash
#SBATCH --job-name=EHD_v2_cnn
#SBATCH --nodes=5
#SBATCH --exclusive
#SBATCH --time=06:00:00
#SBATCH --output=/home/tanpv/Progs/claude/QG_1/logs/cnn_ann_%j.out
#SBATCH --error=/home/tanpv/Progs/claude/QG_1/logs/cnn_ann_%j.err
set -euo pipefail
ROOT=/home/tanpv/Progs/claude/QG_1
PY=/home/tanpv/.conda/envs/ehd/bin/python
# torch/sklearn need the conda libstdc++ ahead of the system one (no CXXABI_1.3.9 in /lib64)
export LD_LIBRARY_PATH=/home/tanpv/.conda/envs/ehd/lib${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}
cd "$ROOT"
mkdir -p logs outputs/loyo_pred_v2_cnn

YEARLIST=$(seq 2003 2022)                       # 20 LOYO folds
NFOLD=$(echo $YEARLIST | wc -w)

# Bound each task's math/torch libs to its 5 cores so 4 tasks/node don't oversubscribe.
export OMP_NUM_THREADS=5 MKL_NUM_THREADS=5 OPENBLAS_NUM_THREADS=5

echo "=== v2 CNN+ANN LOYO: $NFOLD folds | 5 nodes x 4 tasks x 5 cpus ==="
srun --nodes=5 --ntasks=$NFOLD --ntasks-per-node=4 --cpus-per-task=5 \
     --kill-on-bad-exit=1 bash -c \
     "YEARS=($YEARLIST); YR=\${YEARS[\$SLURM_PROCID]}; \
      export OMP_NUM_THREADS=5; \
      echo \"[proc \$SLURM_PROCID @ \$(hostname)] fold \$YR start \$(date +%T)\"; \
      $PY scripts/12b_run_cnn_ann.py --folds \$YR; \
      echo \"[proc \$SLURM_PROCID] fold \$YR done \$(date +%T)\""

echo "=== CNN+ANN folds returned; pred files: $(ls outputs/loyo_pred_v2_cnn/pred_v2cnn_*.npz 2>/dev/null | wc -l)/$NFOLD ==="
echo "=== regenerating ensemble + maps + report (now including cnn, ann_torch) ==="
$PY scripts/14_ensemble_v2.py
$PY scripts/15_skill_maps_v2.py
$PY scripts/16_report_v2_html.py
echo "=== done $(date +%T) ==="