bulid_codebook.py

"""

3DGS Codebook Builder

=====================

使用 KMeans 对 3D Gaussian Splatting 模型的四类特征分别构建 codebook：

  - scale    (3维)   → 16384 个离散索引

  - rotation (4维)   → 16384 个离散索引

  - DC       (3维)   → 4096  个离散索引

  - SH rest  (45维)  → 4096  个离散索引



每个 codebook 单独保存为 .npz 文件，包含：

  - codebook  : (K, D) float32  —— 聚类中心

  - indices   : (N,)  int32     —— 每个高斯点对应的索引

"""

import os
import argparse
import numpy as np
from plyfile import PlyData
from sklearn.cluster import MiniBatchKMeans
import time


# ─────────────────────────────────────────────
# 1. PLY 读取
# ─────────────────────────────────────────────

def read_ply(ply_path: str) -> dict:
    """读取 3DGS .ply 文件，返回各属性 numpy 数组。"""
    plydata = PlyData.read(ply_path)
    vertex  = plydata['vertex']

    positions  = np.stack([vertex['x'], vertex['y'], vertex['z']], axis=1)          # (N, 3)
    opacities  = vertex['opacity'][:, np.newaxis]                                    # (N, 1)
    scales     = np.stack([vertex['scale_0'], vertex['scale_1'],
                           vertex['scale_2']], axis=1)                               # (N, 3)
    rotations  = np.stack([vertex['rot_0'], vertex['rot_1'],
                           vertex['rot_2'], vertex['rot_3']], axis=1)                # (N, 4)
    dc         = np.stack([vertex['f_dc_0'], vertex['f_dc_1'],
                           vertex['f_dc_2']], axis=1)                                # (N, 3)

    sh_keys  = sorted(
        [k for k in vertex.data.dtype.names if k.startswith('f_rest_')],
        key=lambda s: int(s.split('_')[-1])
    )
    sh_rest  = np.stack([vertex[k] for k in sh_keys], axis=1) \
               if sh_keys else None                                                   # (N, 45)

    # filter_3D 是可选字段（部分版本有，部分没有）
    filter_3d = None
    if 'filter_3D' in vertex.data.dtype.names:
        filter_3d = vertex['filter_3D'][:, np.newaxis]                              # (N, 1)

    print(f"[read_ply] 读取完成：{positions.shape[0]} 个高斯点")
    if sh_rest is not None:
        print(f"           SH rest 维度：{sh_rest.shape[1]}  "
              f"（期望 45 = 15 球谐系数 × 3 通道）")

    return {
        'positions':  positions,
        'opacities':  opacities,
        'scales':     scales,
        'rotations':  rotations,
        'dc':         dc,
        'sh_rest':    sh_rest,
        'filter_3d':  filter_3d,
        'plydata':    plydata,
    }


# ─────────────────────────────────────────────
# 2. KMeans 聚类（MiniBatchKMeans，速度快）
# ─────────────────────────────────────────────

def build_codebook(

    features: np.ndarray,

    n_clusters: int,

    name: str,

    random_state: int = 42,

    batch_size: int = 65536,

    max_iter: int = 300,

) -> tuple[np.ndarray, np.ndarray]:
    """

    对 features (N, D) 执行 MiniBatchKMeans，返回：

      codebook : (K, D) float32

      indices  : (N,)  int32

    """
    N, D = features.shape
    # 若点数少于 cluster 数，直接把每个点当一个 cluster
    K = min(n_clusters, N)
    if K < n_clusters:
        print(f"[{name}] 警告：高斯点数 ({N}) < 目标 cluster 数 ({n_clusters})，"
              f"自动调整为 K={K}")

    print(f"[{name}] 开始 KMeans：N={N}, D={D}, K={K} ...")
    t0 = time.time()

    kmeans = MiniBatchKMeans(
        n_clusters=K,
        batch_size=min(batch_size, N),
        max_iter=max_iter,
        random_state=random_state,
        n_init=3,
        verbose=0,
    )
    kmeans.fit(features.astype(np.float32))

    codebook = kmeans.cluster_centers_.astype(np.float32)   # (K, D)
    indices  = kmeans.labels_.astype(np.int32)               # (N,)

    elapsed = time.time() - t0
    inertia = kmeans.inertia_
    print(f"[{name}] 完成！耗时 {elapsed:.1f}s  |  inertia={inertia:.4f}")
    print(f"         codebook shape: {codebook.shape}  |  "
          f"索引范围: [{indices.min()}, {indices.max()}]")

    return codebook, indices


# ─────────────────────────────────────────────
# 3. 保存单个 codebook
# ─────────────────────────────────────────────

def save_codebook(

    save_dir: str,

    name: str,

    codebook: np.ndarray,

    indices: np.ndarray,

) -> None:
    """将 codebook 和 indices 存为 <name>_codebook.npz。"""
    os.makedirs(save_dir, exist_ok=True)
    out_path = os.path.join(save_dir, f"{name}_codebook.npz")
    np.savez_compressed(out_path, codebook=codebook, indices=indices)
    size_mb = os.path.getsize(out_path) / 1024 / 1024
    print(f"[{name}] 已保存 → {out_path}  ({size_mb:.2f} MB)")


# ─────────────────────────────────────────────
# 4. 主流程
# ─────────────────────────────────────────────

CODEBOOK_CONFIG = {
    #  name        n_clusters
    'scale':    16384,
    'rotation': 16384,
    'dc':        4096,
    'sh':        4096,
}


def build_all_codebooks(

    ply_path: str,

    save_dir: str,

    random_state: int = 42,

) -> dict:
    """

    读取 PLY → 对四类特征分别聚类 → 分开保存。



    返回字典：

      {

        'scale':    (codebook_array, indices_array),

        'rotation': ...,

        'dc':       ...,

        'sh':       ...,

      }

    """
    # ── 读取数据 ──────────────────────────────
    data = read_ply(ply_path)

    scales    = data['scales']     # (N, 3)
    rotations = data['rotations']  # (N, 4)
    dc        = data['dc']         # (N, 3)
    sh_rest   = data['sh_rest']    # (N, 45)  已去除 DC 的 SH

    if sh_rest is None:
        raise ValueError("PLY 文件中未找到 f_rest_* 字段，无法构建 SH codebook。")

    feature_map = {
        'scale':    scales,
        'rotation': rotations,
        'dc':       dc,
        'sh':       sh_rest,   # SH codebook 使用去掉 DC 的 45 维高频分量
    }

    # ── 逐一聚类并保存 ────────────────────────
    results = {}
    for name, n_clusters in CODEBOOK_CONFIG.items():
        features = feature_map[name]
        print(f"\n{'='*55}")
        print(f"  构建 [{name}] codebook  |  特征维度: {features.shape[1]}"
              f"  |  目标 K: {n_clusters}")
        print(f"{'='*55}")

        codebook, indices = build_codebook(
            features,
            n_clusters=n_clusters,
            name=name,
            random_state=random_state,
        )
        save_codebook(save_dir, name, codebook, indices)
        results[name] = (codebook, indices)

    print(f"\n{'='*55}")
    print("  所有 codebook 构建完毕！")
    print(f"  输出目录：{os.path.abspath(save_dir)}")
    print(f"{'='*55}")
    return results


# ─────────────────────────────────────────────
# 5. 验证：从 codebook 重建特征并计算误差
# ─────────────────────────────────────────────

def evaluate_codebooks(

    ply_path: str,

    save_dir: str,

) -> None:
    """

    加载已保存的四个 codebook，重建特征，

    计算每个维度的均方根误差（RMSE）。

    """
    data = read_ply(ply_path)
    feature_map = {
        'scale':    data['scales'],
        'rotation': data['rotations'],
        'dc':       data['dc'],
        'sh':       data['sh_rest'],
    }

    print("\n[评估] 重建误差（RMSE）：")
    for name in CODEBOOK_CONFIG:
        path = os.path.join(save_dir, f"{name}_codebook.npz")
        if not os.path.exists(path):
            print(f"  [{name}] 文件不存在，跳过")
            continue

        npz      = np.load(path)
        codebook = npz['codebook']   # (K, D)
        indices  = npz['indices']    # (N,)

        original    = feature_map[name].astype(np.float32)
        reconstructed = codebook[indices]                    # (N, D)

        rmse = np.sqrt(np.mean((original - reconstructed) ** 2))
        max_err = np.abs(original - reconstructed).max()
        print(f"  [{name:8s}]  K={codebook.shape[0]:6d}  D={codebook.shape[1]:3d}"
              f"  RMSE={rmse:.6f}  MaxErr={max_err:.6f}")


# ─────────────────────────────────────────────
# 6. CLI 入口
# ─────────────────────────────────────────────

def parse_args():
    parser = argparse.ArgumentParser(
        description="为 3DGS .ply 文件构建四个 KMeans codebook"
    )
    parser.add_argument('ply_path', type=str,default="./merge/original_3dgs.ply",
                        help='输入的 3DGS .ply 文件路径')
    parser.add_argument('--save_dir', type=str, default='./codebooks',
                        help='codebook 保存目录（默认：./codebooks）')
    parser.add_argument('--seed', type=int, default=42,
                        help='随机种子（默认：42）')
    parser.add_argument('--evaluate', action='store_true',
                        help='构建完成后计算 RMSE 重建误差')
    return parser.parse_args()


if __name__ == '__main__':
    args = parse_args()

    build_all_codebooks(
        ply_path=args.ply_path,
        save_dir=args.save_dir,
        random_state=args.seed,
    )

    if args.evaluate:
        evaluate_codebooks(
            ply_path=args.ply_path,
            save_dir=args.save_dir,
        )