| """ |
| Rewrite generated children from infer_upsample.py debug_npz with a fixed opacity. |
| |
| This is a controlled diagnostic: positions, scales, rotations, DC, and SH stay |
| identical; only opacity changes. |
| |
| Example: |
| python rewrite_generated_opacity.py --debug_npz debug_sample.npz \ |
| --codebook_dir codebooks --alpha 0.2 --save_path same_points_alpha02.ply |
| |
| python rewrite_generated_opacity.py --debug_npz debug_sample.npz \ |
| --codebook_dir codebooks --alpha 0.2 --force_rgb 1 1 1 \ |
| --save_path same_points_alpha02_white.ply |
| """ |
|
|
| import argparse |
| import os |
| from typing import Optional |
|
|
| import numpy as np |
| from plyfile import PlyData, PlyElement |
|
|
|
|
| SH_C0 = 0.28209479177387814 |
|
|
|
|
| def alpha_to_logit(alpha: float) -> float: |
| alpha = float(np.clip(alpha, 1e-6, 1.0 - 1e-6)) |
| return float(np.log(alpha / (1.0 - alpha))) |
|
|
|
|
| def normalize_quaternions(rotations: np.ndarray) -> np.ndarray: |
| rotations = rotations.astype(np.float32, copy=True) |
| norms = np.linalg.norm(rotations, axis=1, keepdims=True) |
| valid = np.isfinite(norms) & (norms > 1e-8) |
| rotations = np.where(valid, rotations / np.maximum(norms, 1e-8), rotations) |
| bad = ~valid.squeeze(1) |
| if bad.any(): |
| rotations[bad] = np.array([1.0, 0.0, 0.0, 0.0], dtype=np.float32) |
| return rotations |
|
|
|
|
| def load_codebook(codebook_dir: str, name: str) -> np.ndarray: |
| path = os.path.join(codebook_dir, f"{name}_codebook.npz") |
| if not os.path.exists(path): |
| raise FileNotFoundError(path) |
| return np.load(path)["codebook"].astype(np.float32) |
|
|
|
|
| def print_percentiles(name: str, values: np.ndarray) -> None: |
| pct = np.percentile(values, [0, 1, 50, 90, 95, 99, 99.9, 100], axis=0) |
| print(f"\n[{name}]") |
| for p, row in zip([0, 1, 50, 90, 95, 99, 99.9, 100], pct): |
| row = np.ravel(row) |
| print(f" p{p:>5}: " + " ".join(f"{float(x):.6g}" for x in row)) |
|
|
|
|
| def write_ply(debug_npz: str, |
| codebook_dir: str, |
| save_path: str, |
| alpha: Optional[float], |
| opacity_logit: Optional[float], |
| force_rgb: Optional[list], |
| scale_offset: float) -> None: |
| data = np.load(debug_npz) |
| keys = set(data.files) |
| positions = data["positions"].astype(np.float32) |
|
|
| |
| |
| |
| |
| |
| if {"scale_idx", "rot_idx", "dc_idx", "sh_idx"}.issubset(keys): |
| scale_idx = data["scale_idx"].astype(np.int64) |
| rot_idx = data["rot_idx"].astype(np.int64) |
| dc_idx = data["dc_idx"].astype(np.int64) |
| sh_idx = data["sh_idx"].astype(np.int64) |
| source_kind = "generated-debug" |
| elif {"scale_indices", "rotation_indices", "dc_indices", "sh_indices"}.issubset(keys): |
| scale_idx = data["scale_indices"].astype(np.int64) |
| rot_idx = data["rotation_indices"].astype(np.int64) |
| dc_idx = data["dc_indices"].astype(np.int64) |
| sh_idx = data["sh_indices"].astype(np.int64) |
| source_kind = "quantized" |
| else: |
| raise KeyError( |
| "Unsupported npz format. Expected either generated debug keys " |
| "(scale_idx/rot_idx/dc_idx/sh_idx) or quantized keys " |
| "(scale_indices/rotation_indices/dc_indices/sh_indices). " |
| f"Found keys: {sorted(keys)}" |
| ) |
|
|
| if alpha is not None: |
| raw_opacity = np.full(positions.shape[0], alpha_to_logit(alpha), dtype=np.float32) |
| elif opacity_logit is not None: |
| raw_opacity = np.full(positions.shape[0], float(opacity_logit), dtype=np.float32) |
| elif "opacities" in keys: |
| raw_opacity = data["opacities"].astype(np.float32) |
| if raw_opacity.ndim > 1: |
| raw_opacity = raw_opacity.reshape(-1) |
| else: |
| raise KeyError("No opacity source found. Pass --alpha or --opacity_logit.") |
|
|
| cb_scale = load_codebook(codebook_dir, "scale") |
| cb_rot = normalize_quaternions(load_codebook(codebook_dir, "rotation")) |
| cb_dc = load_codebook(codebook_dir, "dc") |
| cb_sh = load_codebook(codebook_dir, "sh") |
|
|
| scales = cb_scale[scale_idx].copy() |
| if scale_offset != 0.0: |
| scales = scales + float(scale_offset) |
| rotations = cb_rot[rot_idx] |
| dc = cb_dc[dc_idx] |
| sh = cb_sh[sh_idx] |
|
|
| if force_rgb is not None: |
| rgb = np.asarray(force_rgb, dtype=np.float32) |
| if rgb.shape != (3,): |
| raise ValueError("--force_rgb expects exactly three values: R G B") |
| rgb = np.clip(rgb, 0.0, 1.0) |
| dc_value = (rgb - 0.5) / SH_C0 |
| dc = np.broadcast_to(dc_value.reshape(1, 3), dc.shape).astype(np.float32) |
| sh = np.zeros_like(sh, dtype=np.float32) |
|
|
| fields = ( |
| [("x", "f4"), ("y", "f4"), ("z", "f4"), |
| ("opacity", "f4"), |
| ("scale_0", "f4"), ("scale_1", "f4"), ("scale_2", "f4"), |
| ("rot_0", "f4"), ("rot_1", "f4"), ("rot_2", "f4"), ("rot_3", "f4"), |
| ("f_dc_0", "f4"), ("f_dc_1", "f4"), ("f_dc_2", "f4"), |
| ("filter_3D", "f4")] |
| + [(f"f_rest_{i}", "f4") for i in range(sh.shape[1])] |
| ) |
| vd = np.zeros(positions.shape[0], dtype=np.dtype(fields)) |
| vd["x"] = positions[:, 0] |
| vd["y"] = positions[:, 1] |
| vd["z"] = positions[:, 2] |
| vd["opacity"] = raw_opacity |
| vd["scale_0"] = scales[:, 0] |
| vd["scale_1"] = scales[:, 1] |
| vd["scale_2"] = scales[:, 2] |
| vd["rot_0"] = rotations[:, 0] |
| vd["rot_1"] = rotations[:, 1] |
| vd["rot_2"] = rotations[:, 2] |
| vd["rot_3"] = rotations[:, 3] |
| vd["f_dc_0"] = dc[:, 0] |
| vd["f_dc_1"] = dc[:, 1] |
| vd["f_dc_2"] = dc[:, 2] |
| vd["filter_3D"] = 0.0 |
| for i in range(sh.shape[1]): |
| vd[f"f_rest_{i}"] = sh[:, i] |
|
|
| os.makedirs(os.path.dirname(os.path.abspath(save_path)), exist_ok=True) |
| PlyData([PlyElement.describe(vd, "vertex")]).write(save_path) |
|
|
| alpha_values = 1.0 / (1.0 + np.exp(-np.clip(raw_opacity, -80.0, 80.0))) |
| volumes = np.exp(np.clip(scales.sum(axis=1), -80.0, 80.0)) |
| print(f"[write] source={source_kind}") |
| print(f"[write] {positions.shape[0]:,} gaussians -> {save_path}") |
| print_percentiles("written raw opacity", raw_opacity) |
| print_percentiles("written alpha", alpha_values) |
| print_percentiles("written f_dc", dc) |
| approx_rgb = np.clip(dc * SH_C0 + 0.5, 0.0, 1.0) |
| print_percentiles("approx RGB from DC only", approx_rgb) |
| print_percentiles("written raw scale", scales) |
| print_percentiles("exp(written scale)", np.exp(np.clip(scales, -80.0, 80.0))) |
| print_percentiles("decoded physical volume", volumes) |
|
|
|
|
| def parse_args() -> argparse.Namespace: |
| parser = argparse.ArgumentParser(description="Rewrite generated debug_npz as PLY with fixed opacity.") |
| parser.add_argument("--debug_npz", required=True) |
| parser.add_argument("--codebook_dir", required=True) |
| parser.add_argument("--save_path", required=True) |
| group = parser.add_mutually_exclusive_group() |
| group.add_argument("--alpha", type=float, help="Fixed alpha value to write, converted to raw logit.") |
| group.add_argument("--opacity_logit", type=float, help="Fixed raw opacity logit to write.") |
| parser.add_argument("--force_rgb", nargs=3, type=float, |
| help="Force DC color to this RGB in [0,1] and zero SH rest.") |
| parser.add_argument("--scale_offset", type=float, default=0.0, |
| help="Add this value to all raw log-scales. 0.693 roughly doubles physical scale.") |
| return parser.parse_args() |
|
|
|
|
| if __name__ == "__main__": |
| args = parse_args() |
| write_ply( |
| debug_npz=args.debug_npz, |
| codebook_dir=args.codebook_dir, |
| save_path=args.save_path, |
| alpha=args.alpha, |
| opacity_logit=args.opacity_logit, |
| force_rgb=args.force_rgb, |
| scale_offset=args.scale_offset, |
| ) |
|
|