import os import sys import torch import random import numpy as np import torch.nn.functional as F from argparse import ArgumentParser from core.registry import register_method from core.base_method import BaseMethod SURFEL_ROOT = "/root/autodl-tmp/gaussian_surfels" if SURFEL_ROOT not in sys.path: sys.path.insert(0, SURFEL_ROOT) @register_method("gaussian_surfel") class SurfelWrapper(BaseMethod): def __init__(self, dataset_config, hyperparams): from scene import Scene, GaussianModel from arguments import ModelParams, PipelineParams, OptimizationParams self.parser = ArgumentParser() self.lp = ModelParams(self.parser) self.op = OptimizationParams(self.parser) self.pp = PipelineParams(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.resolution = dataset_config.get("resolution", 1) self.args.eval = True self.args.surface = True if any(x in self.args.source_path.lower() for x in ["tnt", "360", "tanks"]): self.args.images = "images" else: self.args.images = "image" self.dataset = self.lp.extract(self.args) self.opt = self.op.extract(self.args) self.pipe = self.pp.extract(self.args) _prev_cwd = os.getcwd(); os.chdir("/root/autodl-tmp/gaussian_surfels") self.gaussians = GaussianModel(self.dataset) os.chdir(_prev_cwd) # 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, self.opt.camera_lr, shuffle=False, resolution_scales=[1, 2, 4]) 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 = None self.last_n_gaussians = len(self.gaussians.get_xyz) self.track_decoupling = hyperparams.get("track_decoupling", False) self.cap_gaussians = hyperparams.get("cap_gaussians", None) def train_iteration(self, step): from gaussian_renderer import render as native_render from utils.loss_utils import l1_loss, ssim, cos_loss from utils.image_utils import depth2normal self.gaussians.update_learning_rate(step) if step % 1000 == 0: self.gaussians.oneupSHdegree() scale = 1 if step < 2000: scale = 4 elif step < 5000: scale = 2 if not self.viewpoint_stack: self.viewpoint_stack = self.scene.getTrainCameras(scale).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, [float('inf'), float('inf')]) image, normal, depth, opac = render_pkg["render"], render_pkg["normal"], render_pkg["depth"], render_pkg["opac"] gt_image = viewpoint_cam.get_gtImage(self.background, self.dataset.use_mask) mask_vis = (opac.detach() > 1e-5) loss_target = (1.0 - self.opt.lambda_dssim) * l1_loss(image, gt_image) loss_parasitic = self.opt.lambda_dssim * (1.0 - ssim(image, gt_image)) d2n = depth2normal(depth, mask_vis, viewpoint_cam) normal_norm = F.normalize(normal, dim=0) * mask_vis loss_surface = cos_loss(normal_norm, d2n) loss = loss_target + loss_parasitic loss += (0.01 + 0.1 * min(2 * step / self.opt.iterations, 1)) * loss_surface grad_cos_sim = 0.0 p_ratio = 0.0 if self.track_decoupling and step % 100 == 0: self.gaussians.optimizer.zero_grad() loss_target.backward(retain_graph=True) g_t = self.gaussians._xyz.grad.clone() if self.gaussians._xyz.grad is not None else None self.gaussians.optimizer.zero_grad() loss_parasitic.backward(retain_graph=True) g_p = self.gaussians._xyz.grad.clone() if self.gaussians._xyz.grad is not None else None if g_t is not None and g_p is not None: mask = (g_t.norm(dim=1) > 1e-8) & (g_p.norm(dim=1) > 1e-8) if mask.any(): grad_cos_sim = float(F.cosine_similarity(g_t[mask], g_p[mask], dim=1).mean()) p_ratio = float(g_p[mask].norm(dim=1).mean() / (g_t[mask].norm(dim=1).mean() + 1e-8)) self.gaussians.optimizer.zero_grad() loss.backward() else: loss.backward() # 修复点：包裹在 no_grad() 中，且在 optimizer.step() 之前执行 with torch.no_grad(): if step > self.opt.densify_from_iter: self.gaussians.max_radii2D[render_pkg["visibility_filter"]] = torch.max(self.gaussians.max_radii2D[render_pkg["visibility_filter"]], render_pkg["radii"][render_pkg["visibility_filter"]]) self.gaussians.add_densification_stats(render_pkg["viewspace_points"], render_pkg["visibility_filter"]) if step % self.opt.densification_interval == 0: self.gaussians.adaptive_prune(0.1, self.scene.cameras_extent) if self.cap_gaussians is None or len(self.gaussians.get_xyz) < self.cap_gaussians: self.gaussians.adaptive_densify(self.opt.densify_grad_threshold, self.scene.cameras_extent) if step % self.opt.opacity_reset_interval == 0: self.gaussians.reset_opacity(0.12, step) # 统计完致密化状态后，再进行优化更新并清空梯度 self.gaussians.optimizer.step() self.gaussians.optimizer.zero_grad(set_to_none=True) num_gaussians = len(self.gaussians.get_xyz) metrics = { "loss": float(loss), "loss_l1": float(loss_target), "loss_ssim": float(loss_parasitic), "num_gaussians": num_gaussians, "delta_N": num_gaussians - self.last_n_gaussians, "peak_vram_GB": float(torch.cuda.max_memory_allocated() / (1024**3)), "grad_cos_sim": grad_cos_sim, "parasitic_ratio": p_ratio } self.last_n_gaussians = num_gaussians histograms = {} if step % 1000 == 0: histograms["opacity"] = torch.sigmoid(self.gaussians._opacity).detach() scales = torch.exp(self.gaussians._scaling).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): from gaussian_renderer import render as native_render with torch.no_grad(): pkg = native_render(camera, self.gaussians, self.pipe, self.background, [float('inf'), float('inf')]) return {"image": pkg["render"], "depth": pkg["depth"]} def save(self, save_dir, step): self.scene.save(step) def load(self, model_path, iteration): self.gaussians.load_ply(os.path.join(model_path, f"point_cloud/iteration_{iteration}/point_cloud.ply")) def get_spatial_centers(self): return self.gaussians._xyz.detach() def compute_physical_metrics(self, cameras=None): scales = self.gaussians.get_scaling.detach() return { "gamma_median": float(torch.median(scales[:, 0] / (scales[:, 1] + 1e-7))), "alpha_mean": float(self.gaussians.get_opacity.detach().mean()), "z_scale_mean": float(scales[:, 2].mean()) } def evaluate_spatial_field(self, query_points, cameras=None): from pytorch3d.ops import knn_points xyz = self.gaussians.get_xyz.detach() opac = self.gaussians.get_opacity.detach() dist_sq, _, _ = knn_points(query_points.unsqueeze(0), xyz.unsqueeze(0), K=1) dist = torch.sqrt(dist_sq.squeeze(0)) weights = torch.exp(-0.5 * (dist / 0.01)**2) return (weights * opac.T).sum(dim=1)