Upload quantize.py

quantize.py

@@ -0,0 +1,314 @@
+"""
+3DGS Codebook Quantizer
+========================
+使用已训练好的 codebook 对新的 3DGS .ply 文件进行量化，
+将连续特征映射为离散索引，并可选地重建量化后的特征写回 .ply。
+
+输出：
+  <scene>_quantized.npz   —— 四类索引 + 重建误差统计
+  <scene>_quantized.ply   —— （可选）用 codebook 重建特征后写回的新 .ply
+"""
+
+import os
+import argparse
+import numpy as np
+from plyfile import PlyData, PlyElement
+import time
+
+
+# ─────────────────────────────────────────────
+# 1. PLY 读取（复用训练时的读法）
+# ─────────────────────────────────────────────
+
+def read_ply(ply_path: str) -> dict:
+    plydata = PlyData.read(ply_path)
+    vertex  = plydata['vertex']
+
+    positions = np.stack([vertex['x'], vertex['y'], vertex['z']], axis=1)
+    opacities = vertex['opacity'][:, np.newaxis]
+    scales    = np.stack([vertex['scale_0'], vertex['scale_1'],
+                          vertex['scale_2']], axis=1)
+    rotations = np.stack([vertex['rot_0'], vertex['rot_1'],
+                          vertex['rot_2'], vertex['rot_3']], axis=1)
+    dc        = np.stack([vertex['f_dc_0'], vertex['f_dc_1'],
+                          vertex['f_dc_2']], axis=1)
+
+    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
+
+    filter_3d = None
+    if 'filter_3D' in vertex.data.dtype.names:
+        filter_3d = vertex['filter_3D'][:, np.newaxis]
+
+    print(f"[read_ply] {os.path.basename(ply_path)}：{positions.shape[0]} 个高斯点")
+    return {
+        'positions':  positions,
+        'opacities':  opacities,
+        'scales':     scales,
+        'rotations':  rotations,
+        'dc':         dc,
+        'sh_rest':    sh_rest,
+        'filter_3d':  filter_3d,
+        'plydata':    plydata,
+        'sh_keys':    sh_keys,
+    }
+
+
+# ─────────────────────────────────────────────
+# 2. 加载 codebook
+# ─────────────────────────────────────────────
+
+def load_codebook(codebook_dir: str, name: str):
+    """
+    加载单个 codebook .npz，返回 codebook 矩阵 (K, D)。
+    训练时保存的 indices 属于训练集，量化时不使用。
+    """
+    path = os.path.join(codebook_dir, f"{name}_codebook.npz")
+    if not os.path.exists(path):
+        raise FileNotFoundError(f"找不到 codebook 文件：{path}")
+    npz = np.load(path)
+    codebook = npz['codebook'].astype(np.float32)   # (K, D)
+    print(f"[load] {name}_codebook：K={codebook.shape[0]}, D={codebook.shape[1]}")
+    return codebook
+
+
+# ─────────────────────────────────────────────
+# 3. 最近邻量化（核心）
+# ─────────────────────────────────────────────
+
+def quantize(features: np.ndarray, codebook: np.ndarray, name: str,
+             batch_size: int = 65536):
+    """
+    对 features (N, D) 在 codebook (K, D) 中做最近邻搜索。
+    采用分批计算避免一次性构造 (N, K) 的巨型矩阵撑爆内存。
+
+    返回：
+      indices      : (N,)  int32   每个点对应的 codebook 索引
+      reconstructed: (N, D) float32 量化后重建的特征
+    """
+    features  = features.astype(np.float32)
+    N, D      = features.shape
+    K         = codebook.shape[0]
+    indices   = np.empty(N, dtype=np.int32)
+
+    # ── 分批最近邻 ──────────────────────────
+    # 利用展开的 L2 距离公式：
+    #   ||x - c||^2 = ||x||^2 + ||c||^2 - 2 * x @ c^T
+    cb_norm2 = np.sum(codebook ** 2, axis=1)   # (K,)
+
+    t0 = time.time()
+    for start in range(0, N, batch_size):
+        end  = min(start + batch_size, N)
+        feat = features[start:end]               # (B, D)
+
+        feat_norm2 = np.sum(feat ** 2, axis=1, keepdims=True)   # (B, 1)
+        # (B, K) 距离矩阵
+        dist2 = feat_norm2 + cb_norm2[np.newaxis, :] \
+                - 2.0 * (feat @ codebook.T)
+        indices[start:end] = np.argmin(dist2, axis=1)
+
+    elapsed = time.time() - t0
+    reconstructed = codebook[indices]            # (N, D)
+
+    # ── 误差统计 ──────────────────────────
+    diff  = features - reconstructed
+    rmse  = float(np.sqrt(np.mean(diff ** 2)))
+    max_e = float(np.abs(diff).max())
+    usage = len(np.unique(indices))              # 实际用到多少个 cluster
+
+    print(f"[{name:8s}] 量化完成 {elapsed:.1f}s  |  "
+          f"RMSE={rmse:.6f}  MaxErr={max_e:.6f}  "
+          f"使用 {usage}/{K} 个 cluster "
+          f"({100*usage/K:.1f}%)")
+
+    return indices, reconstructed, {'rmse': rmse, 'max_err': max_e, 'cluster_usage': usage}
+
+
+# ─────────────────────────────────────────────
+# 4. 量化全部特征
+# ─────────────────────────────────────────────
+
+def quantize_all(data: dict, codebook_dir: str):
+    """
+    加载四个 codebook，对新场景的高斯点逐一量化。
+
+    返回：
+      results  dict  {name: {'indices', 'reconstructed', 'stats'}}
+      codebooks dict {name: np.ndarray}
+    """
+    feature_map = {
+        'scale':    data['scales'],
+        'rotation': data['rotations'],
+        'dc':       data['dc'],
+        'sh':       data['sh_rest'],
+    }
+
+    if data['sh_rest'] is None:
+        raise ValueError("PLY 中无 f_rest_* 字段，无法量化 SH。")
+
+    results   = {}
+    codebooks = {}
+    for name, features in feature_map.items():
+        print(f"\n{'='*55}")
+        print(f"  量化 [{name}]  特征维度: {features.shape[1]}")
+        print(f"{'='*55}")
+
+        cb = load_codebook(codebook_dir, name)
+        codebooks[name] = cb
+
+        idx, recon, stats = quantize(features, cb, name)
+        results[name] = {
+            'indices':       idx,
+            'reconstructed': recon,
+            'stats':         stats,
+        }
+
+    return results, codebooks
+
+
+# ─────────────────────────────────────────────
+# 5. 保存量化结果（索引 + 统计）
+# ─────────────────────────────────────────────
+
+def save_quantized(save_path: str, data: dict, results: dict) -> None:
+    """
+    保存量化后的四类索引和统计信息到单个 .npz。
+
+    文件内容：
+      scale_indices    (N,) int32
+      rotation_indices (N,) int32
+      dc_indices       (N,) int32
+      sh_indices       (N,) int32
+      positions        (N, 3) float32   原始坐标（方便后续对齐）
+      opacities        (N, 1) float32
+    """
+    save_dict = {
+        'positions':         data['positions'].astype(np.float32),
+        'opacities':         data['opacities'].astype(np.float32),
+        'scale_indices':     results['scale']['indices'],
+        'rotation_indices':  results['rotation']['indices'],
+        'dc_indices':        results['dc']['indices'],
+        'sh_indices':        results['sh']['indices'],
+    }
+    np.savez_compressed(save_path, **save_dict)
+    size_mb = os.path.getsize(save_path) / 1024 / 1024
+    print(f"\n[保存] 量化索引 → {save_path}  ({size_mb:.2f} MB)")
+
+
+# ─────────────────────────────────────────────
+# 6. （可选）写回量化重建的 .ply
+# ─────────────────────────────────────────────
+
+def save_reconstructed_ply(
+    save_path: str,
+    data: dict,
+    results: dict,
+) -> None:
+    """
+    用 codebook 重建的特征替换原始值，写出新的 .ply 文件。
+    positions 和 opacities 保持不变（未量化）。
+    """
+    plydata  = data['plydata']
+    vertex   = plydata['vertex']
+    sh_keys  = data['sh_keys']
+
+    # ── 取出重建值 ──────────────────────────
+    scales_r    = results['scale']['reconstructed']       # (N, 3)
+    rotations_r = results['rotation']['reconstructed']    # (N, 4)
+    dc_r        = results['dc']['reconstructed']          # (N, 3)
+    sh_r        = results['sh']['reconstructed']          # (N, 45)
+
+    # ── 修改 vertex 数组 ────────────────────
+    # 注意：vertex.data 是结构化 numpy 数组，直接按字段赋值
+    arr = vertex.data.copy()
+
+    arr['scale_0'] = scales_r[:, 0]
+    arr['scale_1'] = scales_r[:, 1]
+    arr['scale_2'] = scales_r[:, 2]
+
+    arr['rot_0'] = rotations_r[:, 0]
+    arr['rot_1'] = rotations_r[:, 1]
+    arr['rot_2'] = rotations_r[:, 2]
+    arr['rot_3'] = rotations_r[:, 3]
+
+    arr['f_dc_0'] = dc_r[:, 0]
+    arr['f_dc_1'] = dc_r[:, 1]
+    arr['f_dc_2'] = dc_r[:, 2]
+
+    for i, key in enumerate(sh_keys):
+        arr[key] = sh_r[:, i]
+
+    # ── 写出新 ply ──────────────────────────
+    new_vertex  = PlyElement.describe(arr, 'vertex')
+    new_plydata = PlyData([new_vertex], text=plydata.text)
+    new_plydata.write(save_path)
+
+    size_mb = os.path.getsize(save_path) / 1024 / 1024
+    print(f"[保存] 重建 .ply → {save_path}  ({size_mb:.2f} MB)")
+
+
+# ─────────────────────────────────────────────
+# 7. 打印汇总统计
+# ─────────────────────────────────────────────
+
+def print_summary(results: dict) -> None:
+    print(f"\n{'='*55}")
+    print(f"  量化汇总")
+    print(f"{'='*55}")
+    print(f"  {'特征':<10} {'RMSE':>10} {'MaxErr':>10} {'Cluster使用率':>14}")
+    print(f"  {'-'*46}")
+    for name, res in results.items():
+        s = res['stats']
+        print(f"  {name:<10} {s['rmse']:>10.6f} {s['max_err']:>10.6f} "
+              f"  {s['cluster_usage']:>5} / {len(np.unique(res['indices'])):>5}"
+              f"  ({100*s['cluster_usage']/s['cluster_usage']:.0f}%)")
+    print(f"{'='*55}")
+
+
+# ─────────────────────────────────────────────
+# 8. CLI 入口
+# ─────────────────────────────────────────────
+
+def parse_args():
+    parser = argparse.ArgumentParser(
+        description="用已训练的 codebook 量化新的 3DGS .ply 文件"
+    )
+    parser.add_argument('ply_path', type=str,
+                        help='待量化的 3DGS .ply 文件路径')
+    parser.add_argument('--codebook_dir', type=str, default='./codebooks',
+                        help='存放四个 *_codebook.npz 的目录（默认：./codebooks）')
+    parser.add_argument('--save_dir', type=str, default='./quantized',
+                        help='量化结果输出目录（默认：./quantized）')
+    parser.add_argument('--save_ply', action='store_true',
+                        help='同时输出用 codebook 重建特征后的 .ply 文件')
+    return parser.parse_args()
+
+
+if __name__ == '__main__':
+    args = parse_args()
+    os.makedirs(args.save_dir, exist_ok=True)
+
+    # ── 读取新场景 ──────────────────────────
+    data = read_ply(args.ply_path)
+
+    # ── 量化 ────────────────────────────────
+    results, codebooks = quantize_all(data, args.codebook_dir)
+
+    # ── 打印汇总 ────────────────────────────
+    print_summary(results)
+
+    # ── 保存索引 ────────────────────────────
+    scene_name = os.path.splitext(os.path.basename(args.ply_path))[0]
+    npz_path   = os.path.join(args.save_dir, f"{scene_name}_quantized.npz")
+    save_quantized(npz_path, data, results)
+
+    # ── （可选）写回重建 ply ─────────────────
+    if args.save_ply:
+        ply_out = os.path.join(args.save_dir, f"{scene_name}_reconstructed.ply")
+        save_reconstructed_ply(ply_out, data, results)
+
+    print("\n全部完成！")