New/REG/generate_2400000_100.sh

#!/usr/bin/env bash
# ODE 采样（euler_sampler，与 samplers.py 中 REG ODE 一致）：分别用 5 / 10 / 50 步，输出到不同子目录。
set -euo pipefail

SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
cd "$SCRIPT_DIR"

NUM_GPUS=1
NUM_SAMPLES=100
PER_PROC_BATCH_SIZE=10

CFG_SCALE=1
CLS_CFG_SCALE=1
GH=0.85

CKPT_PATH="/gemini/space/zhaozy/guzhenyu/UAVFlow/UAV_Flow_base/REG-XLarge-256/REG-XLarge-256/2400000.pt"
#CKPT_PATH="/your_path/reg_xlarge_dinov2_base_align_8_cls/linear-dinov2-b-enc8/checkpoints/0040000.pt"

# 根输出目录（可用环境变量覆盖）；其下会生成 ode_steps_5 / ode_steps_10 / ode_steps_50
OUTPUT_BASE="${OUTPUT_BASE:-./reg_xl_2400000_ode_multistep}"
NOISE_BANK_PATH="${NOISE_BANK_PATH:-${OUTPUT_BASE}/noise_bank.pt}"

export NCCL_P2P_DISABLE=1
export OUTPUT_BASE

GLOBAL_BATCH_SIZE=$((PER_PROC_BATCH_SIZE * NUM_GPUS))
TOTAL_SAMPLES=$((((NUM_SAMPLES + GLOBAL_BATCH_SIZE - 1) / GLOBAL_BATCH_SIZE) * GLOBAL_BATCH_SIZE))
export NOISE_BANK_PATH TOTAL_SAMPLES

REBUILD_NOISE_BANK=0
if [[ ! -f "${NOISE_BANK_PATH}" ]]; then
  REBUILD_NOISE_BANK=1
else
  EXISTING_SAMPLES=$(python - <<'PY'
import os
import torch
p = os.environ["NOISE_BANK_PATH"]
try:
    d = torch.load(p, map_location="cpu", weights_only=True)
except TypeError:
    d = torch.load(p, map_location="cpu")
print(int(d["z"].shape[0]))
PY
)
  if [[ "${EXISTING_SAMPLES}" != "${TOTAL_SAMPLES}" ]]; then
    echo "noise_bank size mismatch: existing=${EXISTING_SAMPLES}, required=${TOTAL_SAMPLES}. Rebuilding..."
    REBUILD_NOISE_BANK=1
  fi
fi

if [[ "${REBUILD_NOISE_BANK}" -eq 1 ]]; then
  mkdir -p "$(dirname "${NOISE_BANK_PATH}")"
  echo "Creating fixed noise bank: ${NOISE_BANK_PATH} (total_samples=${TOTAL_SAMPLES})"
  python - <<'PY'
import os
import torch

noise_path = os.environ["NOISE_BANK_PATH"]
total_samples = int(os.environ["TOTAL_SAMPLES"])
num_classes = 1000
in_channels = 4
latent_size = 32  # 256 / 8
cls_dim = 768

z = torch.randn(total_samples, in_channels, latent_size, latent_size)
y = torch.randint(0, num_classes, (total_samples,), dtype=torch.long)
cls_z = torch.randn(total_samples, cls_dim)

torch.save({"z": z, "y": y, "cls_z": cls_z}, noise_path)
print(f"Saved fixed noise bank to {noise_path}")
PY
fi

for NUM_STEP in 5 10 20 50 100; do
  RANDOM_PORT=$((RANDOM % 100 + 1200))
  OUTPUT_DIR="${OUTPUT_BASE}/ode_steps_${NUM_STEP}/checkpoints"
  echo "=== ODE num_steps=${NUM_STEP} -> ${OUTPUT_DIR} ==="

  python -m torch.distributed.launch \
    --master_port="${RANDOM_PORT}" \
    --nproc_per_node="${NUM_GPUS}" \
    generate.py \
    --model "SiT-XL/2" \
    --num-fid-samples "${NUM_SAMPLES}" \
    --ckpt "${CKPT_PATH}" \
    --path-type=linear \
    --encoder-depth=8 \
    --projector-embed-dims=768 \
    --per-proc-batch-size="${PER_PROC_BATCH_SIZE}" \
    --mode=ode \
    --num-steps="${NUM_STEP}" \
    --fixed-noise-file "${NOISE_BANK_PATH}" \
    --cfg-scale="${CFG_SCALE}" \
    --cls-cfg-scale="${CLS_CFG_SCALE}" \
    --guidance-high="${GH}" \
    --sample-dir "${OUTPUT_DIR}" \
    --cls=768
done

echo "All ODE runs finished. Building side-by-side pairs..."

python - <<'PY'
import os
from PIL import Image

output_base = os.environ.get("OUTPUT_BASE", "./reg_xl_2400000_ode_multistep")
steps = [5, 10, 50]

def find_single_subdir(step):
    root = os.path.join(output_base, f"ode_steps_{step}", "checkpoints")
    if not os.path.isdir(root):
        raise RuntimeError(f"Missing directory: {root}")
    subdirs = [d for d in os.listdir(root) if os.path.isdir(os.path.join(root, d))]
    if len(subdirs) != 1:
        raise RuntimeError(f"Expected exactly 1 run folder under {root}, got {len(subdirs)}: {subdirs}")
    return os.path.join(root, subdirs[0])

run_dirs = {s: find_single_subdir(s) for s in steps}
pair_dir = os.path.join(output_base, "pair")
os.makedirs(pair_dir, exist_ok=True)

files = sorted([f for f in os.listdir(run_dirs[steps[0]]) if f.endswith(".png")])
for name in files:
    imgs = [Image.open(os.path.join(run_dirs[s], name)).convert("RGB") for s in steps]
    w = sum(im.width for im in imgs)
    h = max(im.height for im in imgs)
    canvas = Image.new("RGB", (w, h))
    x = 0
    for im in imgs:
        canvas.paste(im, (x, 0))
        x += im.width
    canvas.save(os.path.join(pair_dir, name))

print(f"Saved paired comparisons to: {pair_dir}")
PY

echo "Done. Outputs under: ${OUTPUT_BASE}/ode_steps_{5,10,50}/checkpoints ; pairs under: ${OUTPUT_BASE}/pair"