change_lineage.py

import json
import pickle
import numpy as np
import os


def convert_lineage_to_split_genealogy(
        lineage_json_path: str,
        save_pkl_path: str,
        max_children: int = 8,  # merge后每个粗点平均对应4个细点，留余量
):
    """
    正确的转换方向：

    lineage["L1"][i] = [a, b, c, d]
      含义：L1第i个粗点 由 L0中的a,b,c,d细点 merge而来
      反转：L1第i个粗点 → 子节点是 L0中的 a,b,c,d

    genealogy[1]['children_ids'] shape (N_L1粗, max_children)
      含义：L1第i个粗点，在L0中对应的细节点索引

    使用时：
      父节点 = L3/L2/L1 的粗粒度高斯点
      子节点 = 更细粒度的高斯点（向原始方向展开）
      序列：L3(粗)→L2(细)→L1(更细)→L0(原始)
    """
    print(f"[convert] 加载 {lineage_json_path}")
    with open(lineage_json_path, 'r') as f:
        lineage = json.load(f)

    genealogy = {}

    # 注意方向：粗→细，所以 key 含义变了
    # genealogy[1]: L1粗节点 → L0细节点
    # genealogy[2]: L2粗节点 → L1细节点
    # genealogy[3]: L3粗节点 → L2细节点
    level_map = {
        "L1": 1,
        "L2": 2,
        "L3": 3,
    }

    for level_name, gen_key in level_map.items():
        if level_name not in lineage:
            continue

        # merge_list[粗点索引] = [细点索引, ...]
        merge_list = lineage[level_name]
        n_coarse = len(merge_list)

        # 统计每个粗点有多少个细点
        child_counts = [len(parents) for parents in merge_list]
        max_actual = max(child_counts)
        avg_actual = np.mean(child_counts)
        print(f"[convert] {level_name}（粗→细方向）：")
        print(f"           粗节点数={n_coarse}")
        print(f"           每个粗节点平均对应细节点数={avg_actual:.2f}")
        print(f"           最多子节点数={max_actual}")

        # 直接用 merge_list 构建 children_ids
        # merge_list[i] 就是 L(k-1)_粗节点i 对应的所有 L(k-2)_细节点索引
        actual_max = min(max_actual, max_children)
        children_ids = np.full((n_coarse, actual_max), fill_value=-1, dtype=np.int32)

        truncated = 0
        for coarse_idx, fine_indices in enumerate(merge_list):
            n = min(len(fine_indices), actual_max)
            children_ids[coarse_idx, :n] = fine_indices[:n]
            if len(fine_indices) > actual_max:
                truncated += 1

        if truncated:
            print(f"           ⚠️  {truncated} 个粗节点子节点数超过{actual_max}被截断")

        genealogy[gen_key] = {
            'children_ids': children_ids,  # (N_coarse, max_children)
        }

    with open(save_pkl_path, 'wb') as f:
        pickle.dump(genealogy, f, protocol=4)

    size_mb = os.path.getsize(save_pkl_path) / 1024 / 1024
    print(f"\n[convert] 已保存 → {save_pkl_path}  ({size_mb:.2f} MB)")
    return genealogy


def verify(lineage_json_path, genealogy_pkl_path, n_sample=5):
    with open(lineage_json_path, 'r') as f:
        lineage = json.load(f)
    with open(genealogy_pkl_path, 'rb') as f:
        genealogy = pickle.load(f)

    print("\n[verify] 抽查：粗节点的子节点应与 lineage 一致")
    for level_name, gen_key in [("L1", 1), ("L2", 2), ("L3", 3)]:
        if level_name not in lineage or gen_key not in genealogy:
            continue
        merge_list = lineage[level_name]
        children_ids = genealogy[gen_key]['children_ids']
        sample_idx = np.random.choice(len(merge_list), min(n_sample, len(merge_list)), replace=False)
        errors = 0
        for i in sample_idx:
            expected = set(merge_list[i])
            actual = set(int(x) for x in children_ids[i] if x >= 0)
            if not expected.issubset(actual):
                print(f"  ❌ {level_name}[{i}]: expected={expected}, actual={actual}")
                errors += 1
        if errors == 0:
            print(f"  ✅ {level_name}：{len(sample_idx)} 个粗节点验证通过")


if __name__ == '__main__':
    genealogy = convert_lineage_to_split_genealogy(
        lineage_json_path="outputs/lineage.json",
        save_pkl_path="outputs/genealogy.pkl",
        max_children=4,
    )
    verify("outputs/lineage.json", "outputs/genealogy.pkl", n_sample=10)