#!/bin/bash
# Wait for Kermany SD-VAE stage2, then sample and run C-sdvae downstream.
set -euo pipefail

export https_proxy=http://10.140.15.68:3128 http_proxy=http://10.140.15.68:3128
export HF_DATASETS_CACHE=/data/temp/qinshengqian/c3/hf_cache
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True

BASE=/data/temp/qinshengqian/c3
CODE=$BASE/Code/RAEv2
PIPE=$BASE/kermany_pipeline
DOWN=$BASE/kermany_downstream
SYN=$DOWN/synth/sdvae
JSONS=$DOWN/jsons
LOGS=$DOWN/logs
CFGS=$DOWN/cfgs
mkdir -p "$SYN" "$JSONS" "$LOGS" "$CFGS"

FINAL_EPOCH=${FINAL_EPOCH:-50}
STOP_TRAINING_AFTER_FINAL=${STOP_TRAINING_AFTER_FINAL:-1}
FINAL=$(printf "%s/results/stage2/KERMANY_sdvae/checkpoints/ep-%07d.pt" "$CODE" "$FINAL_EPOCH")
RUNLOG=$DOWN/sdvae_completion.log
exec >>"$RUNLOG" 2>&1

ts() { date '+%Y-%m-%d %H:%M:%S'; }

echo "[$(ts)] waiting for $FINAL"
while [ ! -f "$FINAL" ]; do
  if ! pgrep -f "configs/stage2/training/KERMANY/sdvae" >/dev/null 2>&1; then
    echo "[$(ts)] sdvae training process not found before final checkpoint; abort"
    exit 1
  fi
  sleep 120
done
echo "[$(ts)] final checkpoint found"

if [ "$STOP_TRAINING_AFTER_FINAL" = "1" ]; then
  echo "[$(ts)] stopping SD-VAE training after checkpoint ep-$FINAL_EPOCH"
  pkill -TERM -f "configs/stage2/training/KERMANY/sdvae_8gpu_gb64.yaml" || true
  sleep 30
  pkill -KILL -f "configs/stage2/training/KERMANY/sdvae_8gpu_gb64.yaml" || true
fi

CSV=$SYN/synth.csv
if [ ! -f "$CSV" ] || [ "$(tail -n +2 "$CSV" 2>/dev/null | wc -l)" -lt 8000 ]; then
  SCFG=$CFGS/sample_sdvae.yaml
  cd "$CODE"
  /root/miniconda3/envs/rae_v2/bin/python "$PIPE/build_kermany_sample_cfg.py" sdvae "$SCFG"
  rm -f "$SYN"/synth_*.csv "$CSV"
  CLASSES=(CNV DME DRUSEN NORMAL)
  for i in "${!CLASSES[@]}"; do
    c=${CLASSES[$i]}
    CUDA_VISIBLE_DEVICES=$i PYTHONPATH=src /root/miniconda3/envs/rae_v2/bin/python sample_perclass_kermany.py \
      "$SCFG" "$SYN" 2000 80 "$c" "$c" >"$LOGS/sample_sdvae_${c}.log" 2>&1 &
    echo "[$(ts)] launched SD-VAE sampling class=$c gpu=$i pid=$!"
  done
  sample_fail=0
  for pid in $(jobs -rp); do
    wait "$pid" || sample_fail=$((sample_fail + 1))
  done
  [ "$sample_fail" -eq 0 ] || { echo "[$(ts)] sampling failed count=$sample_fail"; exit 1; }
  head -1 "$SYN/synth_CNV.csv" > "$CSV"
  for c in "${CLASSES[@]}"; do tail -n +2 "$SYN/synth_${c}.csv" >> "$CSV"; done
fi

rows=$(tail -n +2 "$CSV" | wc -l)
echo "[$(ts)] synth csv rows=$rows"
[ "$rows" -ge 8000 ] || { echo "[$(ts)] synth csv too small"; exit 1; }

DOSES=(0.05 0.25 1.0)
SEEDS=(0 1 2)
GPUS=(0 1 2 3 4 5 6 7)
MAXJOBS=8
wait_for_slot() {
  while [ "$(jobs -rp | wc -l)" -ge "$MAXJOBS" ]; do
    sleep 10
  done
}

for seed in "${SEEDS[@]}"; do
  out="$JSONS/A_full_s${seed}.json"
  if [ -f "$out" ]; then
    echo "[$(ts)] skip existing full-train baseline $out"
    continue
  fi
  wait_for_slot
  gpu=${GPUS[$((seed % ${#GPUS[@]}))]}
  log="$LOGS/A_full_s${seed}.log"
  /root/miniconda3/envs/rae_v2/bin/python "$PIPE/kermany_dose_rn50.py" \
    --dose 1.0 --regime A --seed "$seed" --device "cuda:$gpu" \
    --real-cap-per-class 0 --epochs 15 --out-json "$out" >"$log" 2>&1 &
  echo "[$(ts)] launched A_full seed=$seed gpu=$gpu pid=$! out=$out"
  sleep 3
done

i=0
for dose in "${DOSES[@]}"; do
  for seed in "${SEEDS[@]}"; do
    out="$JSONS/C-sdvae_d${dose}_s${seed}.json"
    if [ -f "$out" ]; then
      echo "[$(ts)] skip existing $out"
      continue
    fi
    wait_for_slot
    gpu=${GPUS[$((i % ${#GPUS[@]}))]}
    log="$LOGS/C-sdvae_d${dose}_s${seed}.log"
    /root/miniconda3/envs/rae_v2/bin/python "$PIPE/kermany_dose_rn50.py" \
      --dose "$dose" --regime C --seed "$seed" --device "cuda:$gpu" \
      --synth-csv "$CSV" --epochs 15 --out-json "$out" >"$log" 2>&1 &
    echo "[$(ts)] launched C-sdvae d=$dose seed=$seed gpu=$gpu pid=$! out=$out"
    i=$((i + 1))
    sleep 3
  done
done

for seed in "${SEEDS[@]}"; do
  out="$JSONS/C-sdvae_full_s${seed}.json"
  if [ -f "$out" ]; then
    echo "[$(ts)] skip existing full-train $out"
    continue
  fi
  wait_for_slot
  gpu=${GPUS[$((i % ${#GPUS[@]}))]}
  log="$LOGS/C-sdvae_full_s${seed}.log"
  /root/miniconda3/envs/rae_v2/bin/python "$PIPE/kermany_dose_rn50.py" \
    --dose 1.0 --regime C --seed "$seed" --device "cuda:$gpu" \
    --real-cap-per-class 0 --synth-csv "$CSV" --epochs 15 --out-json "$out" >"$log" 2>&1 &
  echo "[$(ts)] launched C-sdvae full seed=$seed gpu=$gpu pid=$! out=$out"
  i=$((i + 1))
  sleep 3
done

fail=0
for pid in $(jobs -rp); do
  wait "$pid" || fail=$((fail + 1))
done

/root/miniconda3/envs/rae_v2/bin/python "$PIPE/aggregate_kermany.py" --arms sdvae --metric macro >"$DOWN/SUMMARY_sdvae_macro.txt"
/root/miniconda3/envs/rae_v2/bin/python "$PIPE/aggregate_kermany.py" --arms sdvae --metric disease_mean >"$DOWN/SUMMARY_sdvae_disease_mean.txt"
echo "[$(ts)] KERMANY_SDVAe_COMPLETION_DONE fail=$fail"
echo "KERMANY_SDVAe_COMPLETION_DONE fail=$fail $(date)" > "$DOWN/SDVAE_COMPLETION_DONE"
exit "$fail"