methods/wrapper_c3dgs.py

import os
import sys
import random
import torch
import torch.nn.functional as F
from argparse import ArgumentParser

from core.registry import register_method
from core.base_method import BaseMethod

_UPSTREAM = '/root/autodl-tmp/C3DGS_offy'
sys.path.insert(0, _UPSTREAM)
# monorepo-flat: insert each first-level subdir individually
import os as _os
for _sub in ('utils', 'scene', 'gaussian_renderer', 'arguments', 'lpipsPyTorch', 'experiments'):
    _p = _os.path.join(_UPSTREAM, _sub)
    if _os.path.isdir(_p):
        sys.path.insert(0, _p)
del _os, _p, _sub, _UPSTREAM
from utils.loss_utils import l1_loss, ssim
from gaussian_renderer import render as native_render
from scene import Scene, GaussianModel
from arguments import ModelParams, PipelineParams, OptimizationParams, CompressionParams
from compression.vq import CompressionSettings, compress_gaussians

@register_method("c3dgs")
class C3DGSWrapper(BaseMethod):
    def __init__(self, dataset_config, hyperparams):
        self.parser = ArgumentParser()
        self.lp = ModelParams(self.parser)
        self.op = OptimizationParams(self.parser)
        self.pp = PipelineParams(self.parser)
        self.cp = CompressionParams(self.parser)
        self.args = self.parser.parse_args([])

        self.args.source_path = dataset_config["source_path"]
        self.args.model_path = dataset_config["model_path"]
        self.args.eval = True
        self.args.resolution = dataset_config.get("resolution", 1)
        self.track_decoupling = hyperparams.get("track_decoupling", False)

        self.dataset = self.lp.extract(self.args)
        self.opt = self.op.extract(self.args)
        self.pipe = self.pp.extract(self.args)
        self.comp = self.cp.extract(self.args)

        self.gaussians = GaussianModel(self.dataset.sh_degree, quantization=not self.opt.not_quantization_aware)
        # INJECTED_RES_FIX begin
        import sys as _sys
        _scene, _explicit_res = None, None
        for _i, _a in enumerate(_sys.argv[:-1]):
            _v = _sys.argv[_i + 1]
            if _a == "--scene":          _scene = _v
            elif _a == "--source_path":  _scene = _v.rstrip("/").split("/")[-1]
            elif _a == "--resolution":
                try: _explicit_res = int(_v)
                except: pass
        _OUTDOOR_360 = {"bicycle", "flowers", "garden", "stump", "treehill"}
        if _explicit_res is not None and _explicit_res > 0:
            _res = _explicit_res
        elif _scene is not None:
            _res = 4 if _scene in _OUTDOOR_360 else 2
        else:
            _res = None
        try:
            if _res is not None:
                self.dataset.resolution = _res
                print("[res-fix] scene=%s explicit=%s -> res=%s  (%s)" % (_scene, _explicit_res, _res, __file__))
        except Exception as _e:
            print("[res-fix] FAILED:", _e)
        # INJECTED_RES_FIX end
        self.scene = Scene(self.dataset, self.gaussians, load_iteration=-1, shuffle=True)
        self.gaussians.training_setup(self.opt)

        bg_color = [1, 1, 1] if self.dataset.white_background else [0, 0, 0]
        self.background = torch.tensor(bg_color, dtype=torch.float32, device="cuda")
        self.viewpoint_stack = self.scene.getTrainCameras().copy()
        
        self.last_n_gaussians = len(self.gaussians.get_xyz)
        self.compressed = False
        self.step_offset = 30000

    def calc_importance_internal(self):
        scaling = self.gaussians.scaling_qa(self.gaussians.scaling_activation(self.gaussians._scaling.detach()))
        cov3d = self.gaussians.covariance_activation(scaling, 1.0, self.gaussians.get_rotation.detach(), True).requires_grad_(True)
        scaling_factor = self.gaussians.scaling_factor_activation(self.gaussians.scaling_factor_qa(self.gaussians._scaling_factor.detach()))

        h1 = self.gaussians._features_dc.register_hook(lambda grad: grad.abs())
        h2 = self.gaussians._features_rest.register_hook(lambda grad: grad.abs())
        h3 = cov3d.register_hook(lambda grad: grad.abs())

        bg = torch.tensor([0.0, 0.0, 0.0], dtype=torch.float32, device="cuda")
        self.gaussians._features_dc.grad = None
        self.gaussians._features_rest.grad = None
        
        num_pixels = 0
        for camera in self.scene.getTrainCameras():
            cov3d_scaled = cov3d * scaling_factor.square()
            rendering = native_render(camera, self.gaussians, self.pipe, bg, clamp_color=False, cov3d=cov3d_scaled)["render"]
            loss = rendering.sum()
            loss.backward()
            num_pixels += rendering.shape[1] * rendering.shape[2]

        importance = torch.cat([self.gaussians._features_dc.grad, self.gaussians._features_rest.grad], 1).flatten(-2) / num_pixels
        cov_grad = cov3d.grad / num_pixels

        h1.remove()
        h2.remove()
        h3.remove()
        return importance.detach(), cov_grad.detach()

    def train_iteration(self, step):
        if not self.compressed:
            color_importance, gaussian_sensitivity = self.calc_importance_internal()
            color_importance_n = color_importance.amax(-1)
            gaussian_importance_n = gaussian_sensitivity.amax(-1)
            
            c_set = CompressionSettings(
                codebook_size=self.comp.color_codebook_size,
                importance_prune=self.comp.color_importance_prune,
                importance_include=self.comp.color_importance_include,
                steps=int(self.comp.color_cluster_iterations),
                decay=self.comp.color_decay,
                batch_size=self.comp.color_batch_size,
            )
            g_set = CompressionSettings(
                codebook_size=self.comp.gaussian_codebook_size,
                importance_prune=None,
                importance_include=self.comp.gaussian_importance_include,
                steps=int(self.comp.gaussian_cluster_iterations),
                decay=self.comp.gaussian_decay,
                batch_size=self.comp.gaussian_batch_size,
            )
            compress_gaussians(self.gaussians, color_importance_n, gaussian_importance_n, c_set, g_set, self.comp.color_compress_non_dir, prune_threshold=self.comp.prune_threshold)
            self.compressed = True
            self.gaussians.training_setup(self.opt)

        actual_step = step + self.step_offset
        self.gaussians.update_learning_rate(actual_step)

        if not self.viewpoint_stack:
            self.viewpoint_stack = self.scene.getTrainCameras().copy()
        viewpoint_cam = self.viewpoint_stack.pop(random.randint(0, len(self.viewpoint_stack) - 1))

        render_pkg = native_render(viewpoint_cam, self.gaussians, self.pipe, self.background)
        image = render_pkg["render"]
        gt_image = viewpoint_cam.original_image.cuda()

        Ll1 = l1_loss(image, gt_image)
        ssim_value = ssim(image, gt_image)
        loss_target = (1.0 - self.opt.lambda_dssim) * Ll1
        loss_parasitic = self.opt.lambda_dssim * (1.0 - ssim_value)
        loss = loss_target + loss_parasitic

        grad_cos_sim = 0.0
        parasitic_ratio = 0.0
        stats = {}

        if self.track_decoupling and step % 100 == 0:
            self.gaussians.optimizer.zero_grad(set_to_none=True)
            loss_target.backward(retain_graph=True)
            grad_target = self.gaussians._xyz.grad.clone() if self.gaussians._xyz.grad is not None else torch.zeros_like(self.gaussians._xyz)
            
            self.gaussians.optimizer.zero_grad(set_to_none=True)
            loss_parasitic.backward(retain_graph=True)
            grad_parasitic = self.gaussians._xyz.grad.clone() if self.gaussians._xyz.grad is not None else torch.zeros_like(self.gaussians._xyz)
            
            valid_mask = (torch.norm(grad_target, dim=1) > 0) & (torch.norm(grad_parasitic, dim=1) > 0)
            if valid_mask.any():
                grad_cos_sim = float(F.cosine_similarity(grad_target[valid_mask], grad_parasitic[valid_mask], dim=1).mean())
            parasitic_ratio = float(torch.norm(grad_parasitic, dim=1).mean() / (torch.norm(grad_target, dim=1).mean() + 1e-7))

            param_groups_map = {
                "spatial": [self.gaussians._xyz],
                "geometry": [self.gaussians._scaling, self.gaussians._rotation, self.gaussians._scaling_factor],
                "opacity": [self.gaussians._opacity],
                "appearance": [self.gaussians._features_dc, self.gaussians._features_rest]
            }

            self.gaussians.optimizer.zero_grad(set_to_none=True)
            loss_target.backward(retain_graph=True)
            grads_target = {}
            for group_name, params in param_groups_map.items():
                grads_target[group_name] = torch.cat([p.grad.clone().reshape(-1) for p in params if p.grad is not None])

            self.gaussians.optimizer.zero_grad(set_to_none=True)
            loss_parasitic.backward(retain_graph=True)
            grads_parasitic = {}
            for group_name, params in param_groups_map.items():
                grads_parasitic[group_name] = torch.cat([p.grad.clone().reshape(-1) for p in params if p.grad is not None])

            for group_name in param_groups_map:
                gt, gp = grads_target.get(group_name), grads_parasitic.get(group_name)
                if gt is not None and gp is not None and gt.numel() > 0 and gp.numel() > 0 and gt.norm() > 0 and gp.norm() > 0:
                    cos = float(F.cosine_similarity(gt.unsqueeze(0), gp.unsqueeze(0)))
                    r = float(gp.norm() / (gt.norm() + gp.norm() + 1e-7))
                    ti = r * max(0.0, -cos)
                else:
                    ti = 0.0
                stats[f"sti_{group_name}"] = ti

            self.gaussians.optimizer.zero_grad(set_to_none=True)
            loss.backward()
        else:
            loss.backward()

        with torch.no_grad():
            self.gaussians.optimizer.step()
            self.gaussians.optimizer.zero_grad(set_to_none=True)

        num_gaussians = self.gaussians.get_xyz.shape[0]
        metrics = {
            "loss": float(loss), "loss_l1": float(loss_target), "loss_ssim": float(loss_parasitic),
            "num_gaussians": int(num_gaussians), "delta_N": int(num_gaussians - self.last_n_gaussians),
            "peak_vram_GB": float(torch.cuda.max_memory_allocated() / (1024 ** 3)),
            "grad_cos_sim": float(grad_cos_sim), "parasitic_ratio": float(parasitic_ratio)
        }
        metrics.update(stats)
        self.last_n_gaussians = num_gaussians

        histograms = {}
        if step % 1000 == 0:
            histograms["opacity"] = torch.sigmoid(self.gaussians._opacity).clone().detach()
            scales = torch.exp(self.gaussians._scaling).clone().detach()
            histograms["scaling"] = scales
            scales_2d = scales[:, :2] if scales.shape[1] >= 2 else scales
            gamma = scales_2d.max(dim=-1)[0] / (scales_2d.min(dim=-1)[0] + 1e-7)
            histograms["anisotropy"] = gamma
            histograms["sh_dc_mag"] = self.gaussians._features_dc.detach().norm(dim=-1)

        return metrics, histograms

    def render(self, camera):
        with torch.no_grad():
            render_pkg = native_render(camera, self.gaussians, self.pipe, self.background)
        return {"image": render_pkg["render"], "depth": render_pkg.get("depth", None)}

    def save(self, save_dir, step):
        os.makedirs(os.path.join(save_dir, "point_cloud", f"iteration_{step}"), exist_ok=True)
        self.gaussians.save_npz(os.path.join(save_dir, "point_cloud", f"iteration_{step}", "point_cloud.npz"))

    def load(self, model_path, iteration):
        self.gaussians.load_npz(os.path.join(model_path, "point_cloud", f"iteration_{iteration}", "point_cloud.npz"))

    def get_spatial_centers(self):
        return self.gaussians._xyz

    def compute_physical_metrics(self, cameras=None):
        metrics = {}
        with torch.no_grad():
            raw_scales = self.gaussians._scaling
            scales = torch.exp(raw_scales)
            
            scales_2d = scales[:, :2] if scales.dim() > 1 and scales.shape[1] >= 2 else scales.unsqueeze(-1).expand(-1, 2)
            
            max_S, _ = torch.max(scales_2d, dim=1)
            min_S, _ = torch.min(scales_2d, dim=1)
            gamma = max_S / (min_S + 1e-7)
            
            metrics["gamma_median"] = float(torch.median(gamma))
            metrics["gamma_90th_percentile"] = float(torch.quantile(gamma, 0.90))
            metrics["scale_mean"] = float(torch.mean(scales_2d))
            metrics["alpha_mean"] = float(torch.mean(torch.sigmoid(self.gaussians._opacity)))
            
            dc, rest = self.gaussians._features_dc, self.gaussians._features_rest
            if rest is not None and rest.shape[1] > 0:
                metrics["sh_energy_ratio"] = float(rest.norm(dim=-1).mean() / (dc.norm(dim=-1).mean() + 1e-7))
                
            if cameras is not None and len(cameras) > 0:
                view_dirs = []
                for c in cameras:
                    view_dirs.append(c.world_view_transform[:3, 2].tolist())
                view_dirs = F.normalize(torch.tensor(view_dirs, dtype=torch.float32, device="cuda"), dim=1)

                rots = F.normalize(self.gaussians._rotation.clone(), dim=1)
                w, x, y, z = rots.unbind(dim=-1)
                normals = F.normalize(torch.stack([2*(x*z + w*y), 2*(y*z - w*x), 1-2*(x*x + y*y)], dim=-1), dim=1)
                
                max_cos, _ = torch.max(torch.abs(torch.matmul(normals, view_dirs.T)), dim=1)
                metrics["billboard_bias_ratio"] = float((max_cos > 0.90).float().mean())

        return metrics

    def evaluate_spatial_field(self, query_points: torch.Tensor, cameras=None) -> torch.Tensor:
        with torch.no_grad():
            V = query_points.shape[0]
            densities = torch.zeros(V, device="cuda")
            xyz, opacities = self.gaussians._xyz, torch.sigmoid(self.gaussians._opacity).squeeze()
            scales = torch.exp(self.gaussians._scaling)
            sigma_sq = (scales[:, :2].max(dim=1)[0].pow(2)) if scales.shape[1] >= 2 else scales.squeeze().pow(2)
            
            N_gaussians = xyz.shape[0]
            chunk_size = max(1, 30_000_000 // (N_gaussians + 1))
            for i in range(0, V, chunk_size):
                end = min(i + chunk_size, V)
                dist_sq = torch.cdist(query_points[i:end], xyz, p=2).pow(2)
                weights = torch.exp(-0.5 * dist_sq / (sigma_sq.unsqueeze(0) + 1e-7))
                densities[i:end] = torch.sum(weights * opacities.unsqueeze(0), dim=1)
            return densities