hotfix Kermany FM pipeline env and prepare skip

code/kermany_pipeline/run_kermany_fm_pipeline.sh

@@ -0,0 +1,239 @@
+#!/bin/bash
+# Run the full Kermany2018 OCT FM pipeline in OCT_8/Baseline/RAE-main.
+set -euo pipefail
+
+export https_proxy=http://10.140.15.68:3128 http_proxy=http://10.140.15.68:3128
+export HF_HOME=/data/temp/qinshengqian/c3/hf_cache
+export TRANSFORMERS_CACHE=/data/temp/qinshengqian/c3/hf_cache
+export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
+
+ROOT=/data/temp/qinshengqian/c3
+PIPE=$ROOT/kermany_pipeline
+DOWN=$ROOT/kermany_downstream
+BASE=/mnt/tidal-alsh-share2/dataset/qinshengqian/research/c3/Data/Classification/OCT_8/Baseline/RAE-main
+DATA=$ROOT/kermany_imagefolder/train
+REAL=$ROOT/kermany_imagefolder/train
+RESULTS=$BASE/results_kermany
+LOGS=$RESULTS/logs
+JSONS=$DOWN/jsons
+SYNROOT=$DOWN/synth
+EVAL=$DOWN/eval
+RUNLOG=$DOWN/kermany_fm_pipeline.log
+
+CORE_ARMS=(RETFound VisionFM DINOv2L MAEL)
+EXTRA_ARMS=(EyeCLIP FMUE UrFound SigLIP2L)
+RUN_EXTRA_ARMS=${RUN_EXTRA_ARMS:-0}
+if [ "$RUN_EXTRA_ARMS" = "1" ]; then
+  ARMS=("${CORE_ARMS[@]}" "${EXTRA_ARMS[@]}")
+else
+  ARMS=("${CORE_ARMS[@]}")
+fi
+GPUS=(0,1 2,3 4,5 6,7)
+PORTS=(29600 29601 29602 29603)
+DOSES=(0.05 0.25 1.0)
+SEEDS=(0 1 2)
+
+mkdir -p "$LOGS" "$JSONS" "$SYNROOT" "$EVAL"
+exec >>"$RUNLOG" 2>&1
+
+ts() { date '+%Y-%m-%d %H:%M:%S'; }
+
+wait_for_sdvae() {
+  local marker=$DOWN/SDVAE_COMPLETION_DONE
+  echo "[$(ts)] waiting for SD-VAE completion marker: $marker"
+  while [ ! -f "$marker" ]; do
+    sleep 600
+  done
+  echo "[$(ts)] SD-VAE marker found"
+}
+
+env_on() {
+  source ~/miniconda3/etc/profile.d/conda.sh
+  conda activate rae_v2
+  cd "$BASE"
+}
+
+prepare() {
+  echo "[$(ts)] preparing Kermany ImageFolder and configs"
+  if [ -d "$ROOT/kermany_imagefolder/train/CNV" ] && [ -d "$ROOT/kermany_imagefolder/test/NORMAL" ]; then
+    echo "[$(ts)] Kermany ImageFolder already exists; skip rebuild"
+  else
+    /root/miniconda3/envs/rae_v2/bin/python "$PIPE/kermany_make_imagefolder.py"
+  fi
+  env_on
+  local missing=0
+  for arm in "${ARMS[@]}"; do
+    [ -f "$BASE/configs/stage1/training/KERMANY/${arm}_decXL.yaml" ] || missing=1
+    [ -f "$BASE/configs/stage2/training/KERMANY/DiTDH-XL_${arm}.yaml" ] || missing=1
+  done
+  if [ "$missing" = "0" ]; then
+    echo "[$(ts)] Kermany FM configs already exist; skip regenerate"
+  else
+    python "$PIPE/make_kermany_fm_configs.py" --base "$BASE" --stage1-epochs 30 --stage2-epochs 50 --arms "${ARMS[@]}"
+  fi
+}
+
+stage1_one() {
+  local arm=$1 gpu=$2 port=$3
+  local ngpu=$(echo "$gpu" | tr ',' '\n' | wc -l)
+  env_on
+  if [ -f "$RESULTS/stage1/stage1_${arm}/checkpoints/ep-last.pt" ]; then
+    echo "[$(ts)] [stage1] skip existing $arm"
+    return
+  fi
+  echo "[$(ts)] [stage1] $arm gpus=$gpu"
+  CUDA_VISIBLE_DEVICES=$gpu EXPERIMENT_NAME=stage1_$arm torchrun --nproc_per_node="$ngpu" --master_port="$port" \
+    src/train_stage1.py \
+    --config "$BASE/configs/stage1/training/KERMANY/${arm}_decXL.yaml" \
+    --data-path "$DATA" --results-dir "$RESULTS/stage1" --image-size 256 --precision bf16 \
+    >"$LOGS/stage1_${arm}.log" 2>&1
+}
+
+stat_one() {
+  local arm=$1 gpu=$2 port=$3
+  local ngpu=$(echo "$gpu" | tr ',' '\n' | wc -l)
+  env_on
+  if [ -f "$RESULTS/stats/$arm/normalization_stats.pt" ]; then
+    echo "[$(ts)] [stat] skip existing $arm"
+    return
+  fi
+  echo "[$(ts)] [stat] $arm gpus=$gpu"
+  local stat_cfg="$DOWN/cfgs/stat_${arm}.yaml"
+  python - <<PY
+import yaml
+from pathlib import Path
+src = Path("$BASE/configs/stage2/training/KERMANY/DiTDH-XL_${arm}.yaml")
+cfg = yaml.safe_load(src.read_text())
+cfg["stage_1"]["params"]["normalization_stat_path"] = None
+Path("$stat_cfg").parent.mkdir(parents=True, exist_ok=True)
+Path("$stat_cfg").write_text(yaml.safe_dump(cfg, sort_keys=False))
+PY
+  SAVE_FOLDER=$arm CUDA_VISIBLE_DEVICES=$gpu torchrun --nproc_per_node="$ngpu" --master_port="$port" \
+    src/calculate_stat.py \
+    --config "$stat_cfg" \
+    --data-path "$DATA" --sample-dir "$RESULTS/stats" --image-size 256 \
+    --per-proc-batch-size 64 --num-workers 8 --precision bf16 \
+    >"$LOGS/stat_${arm}.log" 2>&1
+}
+
+stage2_one() {
+  local arm=$1 gpu=$2 port=$3
+  local ngpu=$(echo "$gpu" | tr ',' '\n' | wc -l)
+  env_on
+  if [ -f "$RESULTS/stage2/stage2_${arm}/checkpoints/ep-last.pt" ]; then
+    echo "[$(ts)] [stage2] skip existing $arm"
+    return
+  fi
+  echo "[$(ts)] [stage2] $arm gpus=$gpu"
+  CUDA_VISIBLE_DEVICES=$gpu EXPERIMENT_NAME=stage2_$arm torchrun --nproc_per_node="$ngpu" --master_port="$port" \
+    src/train.py \
+    --config "$BASE/configs/stage2/training/KERMANY/DiTDH-XL_${arm}.yaml" \
+    --data-path "$DATA" --results-dir "$RESULTS/stage2" --image-size 256 --precision bf16 --compile \
+    >"$LOGS/stage2_${arm}.log" 2>&1
+}
+
+sample_eval_downstream_one() {
+  local arm=$1 gpu=$2
+  env_on
+  local scfg=$DOWN/cfgs/sample_${arm}.yaml
+  local sdir=$SYNROOT/$arm
+  local csv=$sdir/synth.csv
+  if [ ! -f "$csv" ] || [ "$(tail -n +2 "$csv" 2>/dev/null | wc -l)" -lt 8000 ]; then
+    python - <<PY
+import yaml
+from pathlib import Path
+src = Path("$BASE/configs/stage2/training/KERMANY/DiTDH-XL_${arm}.yaml")
+cfg = yaml.safe_load(src.read_text())
+cfg["stage_2"]["ckpt"] = "$RESULTS/stage2/stage2_${arm}/checkpoints/ep-last.pt"
+Path("$scfg").parent.mkdir(parents=True, exist_ok=True)
+Path("$scfg").write_text(yaml.safe_dump(cfg, sort_keys=False))
+PY
+    echo "[$(ts)] [sample] $arm gpu=$gpu"
+    CUDA_VISIBLE_DEVICES=$gpu PYTHONPATH="$BASE/src" python "$PIPE/kermany_fm_sample_conditional.py" \
+      --config "$scfg" --output-dir "$sdir" --num-per-class 2000 --cfg-scales 1.0 \
+      --batch-size 50 --precision bf16 >"$LOGS/sample_${arm}.log" 2>&1
+    /root/miniconda3/envs/rae_v2/bin/python "$PIPE/build_kermany_synth_csv.py" "$sdir" 1.0 "$csv" >>"$LOGS/sample_${arm}.log" 2>&1
+  else
+    echo "[$(ts)] [sample] skip existing $arm csv=$csv"
+  fi
+
+  if [ ! -f "$EVAL/${arm}.json" ]; then
+    echo "[$(ts)] [quality] $arm gpu=$gpu"
+    CUDA_VISIBLE_DEVICES=$gpu python src/evaluate_quality_v2.py \
+      --gen-dir "$sdir" --real-dir "$REAL" --output "$EVAL/${arm}.json" --batch-size 64 \
+      >"$LOGS/eval_${arm}.log" 2>&1 || echo "[$(ts)] [quality] $arm failed; continuing to downstream"
+  fi
+
+  local i=0
+  for dose in "${DOSES[@]}"; do
+    for seed in "${SEEDS[@]}"; do
+      local out="$JSONS/C-${arm}_d${dose}_s${seed}.json"
+      if [ -f "$out" ]; then
+        echo "[$(ts)] [downstream] skip existing $out"
+        continue
+      fi
+      local dgpu=$((i % 8))
+      local log="$DOWN/logs/C-${arm}_d${dose}_s${seed}.log"
+      /root/miniconda3/envs/rae_v2/bin/python "$PIPE/kermany_dose_rn50.py" \
+        --dose "$dose" --regime C --seed "$seed" --device "cuda:$dgpu" \
+        --synth-csv "$csv" --epochs 15 --out-json "$out" >"$log" 2>&1 &
+      echo "[$(ts)] [downstream] launched $arm d=$dose seed=$seed gpu=$dgpu pid=$!"
+      i=$((i + 1))
+      while [ "$(jobs -rp | wc -l)" -ge 8 ]; do sleep 10; done
+      sleep 2
+    done
+  done
+  for seed in "${SEEDS[@]}"; do
+    local out="$JSONS/C-${arm}_full_s${seed}.json"
+    if [ -f "$out" ]; then
+      echo "[$(ts)] [downstream] skip existing $out"
+      continue
+    fi
+    local dgpu=$((i % 8))
+    local log="$DOWN/logs/C-${arm}_full_s${seed}.log"
+    /root/miniconda3/envs/rae_v2/bin/python "$PIPE/kermany_dose_rn50.py" \
+      --dose 1.0 --regime C --seed "$seed" --device "cuda:$dgpu" \
+      --real-cap-per-class 0 --synth-csv "$csv" --epochs 15 --out-json "$out" >"$log" 2>&1 &
+    echo "[$(ts)] [downstream] launched $arm full seed=$seed gpu=$dgpu pid=$!"
+    i=$((i + 1))
+    while [ "$(jobs -rp | wc -l)" -ge 8 ]; do sleep 10; done
+    sleep 2
+  done
+  local fail=0
+  for pid in $(jobs -rp); do wait "$pid" || fail=$((fail + 1)); done
+  echo "[$(ts)] [downstream] $arm fail=$fail"
+}
+
+run_wave() {
+  local -a wave=("$@")
+  echo "[$(ts)] running wave: ${wave[*]}"
+  for idx in "${!wave[@]}"; do
+    stage1_one "${wave[$idx]}" "${GPUS[$idx]}" "$((PORTS[$idx] + 0))" &
+  done
+  wait
+  for idx in "${!wave[@]}"; do
+    stat_one "${wave[$idx]}" "${GPUS[$idx]}" "$((PORTS[$idx] + 100))" &
+  done
+  wait
+  for idx in "${!wave[@]}"; do
+    stage2_one "${wave[$idx]}" "${GPUS[$idx]}" "$((PORTS[$idx] + 200))" &
+  done
+  wait
+  for idx in "${!wave[@]}"; do
+    sample_eval_downstream_one "${wave[$idx]}" "$idx"
+  done
+}
+
+echo "[$(ts)] KERMANY_FM_PIPELINE_START arms=${ARMS[*]}"
+wait_for_sdvae
+prepare
+run_wave "${CORE_ARMS[@]}"
+if [ "$RUN_EXTRA_ARMS" = "1" ]; then
+  run_wave "${EXTRA_ARMS[@]}"
+else
+  echo "[$(ts)] skip extra arms: ${EXTRA_ARMS[*]}"
+fi
+/root/miniconda3/envs/rae_v2/bin/python "$PIPE/aggregate_kermany.py" --arms "${ARMS[@]}" sdvae --metric macro >"$DOWN/SUMMARY_all_macro.txt"
+/root/miniconda3/envs/rae_v2/bin/python "$PIPE/aggregate_kermany.py" --arms "${ARMS[@]}" sdvae --metric disease_mean >"$DOWN/SUMMARY_all_disease_mean.txt"
+echo "KERMANY_FM_PIPELINE_DONE $(date)" > "$DOWN/FM_PIPELINE_DONE"
+echo "[$(ts)] KERMANY_FM_PIPELINE_DONE"