import os import sys import random import torch 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 sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), '../../hac_plus_official'))) from utils.loss_utils import l1_loss, ssim from gaussian_renderer import prefilter_voxel, render as native_render, generate_neural_gaussians from scene import Scene, GaussianModel from arguments import ModelParams, PipelineParams, OptimizationParams from utils.encodings import get_binary_vxl_size @register_method("hac_plus") class HACPlusWrapper(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.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.args.n_features = 4 self.args.log2 = 13 self.args.log2_2D = 15 self.args.lmbda = 0.001 self.dataset = self.lp.extract(self.args) self.opt = self.op.extract(self.args) self.pipe = self.pp.extract(self.args) is_synthetic_nerf = os.path.exists(os.path.join(self.args.source_path, "transforms_train.json")) self.gaussians = GaussianModel( self.dataset.feat_dim, self.dataset.n_offsets, self.dataset.voxel_size, self.dataset.update_depth, self.dataset.update_init_factor, self.dataset.update_hierachy_factor, self.dataset.use_feat_bank, n_features_per_level=self.args.n_features, log2_hashmap_size=self.args.log2, log2_hashmap_size_2D=self.args.log2_2D, is_synthetic_nerf=is_synthetic_nerf, ) # 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.gaussians.update_anchor_bound() 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 = self.gaussians.get_anchor.shape[0] * self.gaussians.n_offsets def train_iteration(self, step): self.gaussians.update_learning_rate(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)) voxel_visible_mask = prefilter_voxel(viewpoint_cam, self.gaussians, self.pipe, self.background) retain_grad = (step < self.opt.update_until and step >= 0) render_pkg = native_render(viewpoint_cam, self.gaussians, self.pipe, self.background, visible_mask=voxel_visible_mask, retain_grad=retain_grad, step=step) image = render_pkg["render"] gt_image = viewpoint_cam.original_image.cuda() Ll1 = l1_loss(image, gt_image) ssim_value = ssim(image, gt_image) scaling_reg = render_pkg["scaling"].prod(dim=1).mean() loss_target = (1.0 - self.opt.lambda_dssim) * Ll1 loss_parasitic = self.opt.lambda_dssim * (1.0 - ssim_value) + 0.01 * scaling_reg bit_per_param = render_pkg.get("bit_per_param", None) bit_hash_grid_val = 0.0 if bit_per_param is not None: _, bit_hash_grid, _, _ = get_binary_vxl_size((self.gaussians.get_encoding_params() + 1) / 2) bit_hash_grid_val = float(bit_hash_grid) denom = self.gaussians._anchor.shape[0] * (self.gaussians.feat_dim + 6 + 3 * self.gaussians.n_offsets) loss_parasitic = loss_parasitic + self.args.lmbda * (bit_per_param + bit_hash_grid / denom) loss = loss_target + loss_parasitic grad_cos_sim = 0.0 parasitic_ratio = 0.0 if self.track_decoupling and step % 100 == 0: self.gaussians.optimizer.zero_grad(set_to_none=True) loss_target.backward(retain_graph=True) grad_t = self.gaussians._anchor.grad.clone() if self.gaussians._anchor.grad is not None else torch.zeros_like(self.gaussians._anchor) self.gaussians.optimizer.zero_grad(set_to_none=True) loss_parasitic.backward(retain_graph=True) grad_p = self.gaussians._anchor.grad.clone() if self.gaussians._anchor.grad is not None else torch.zeros_like(self.gaussians._anchor) state = self.gaussians.optimizer.state.get(self.gaussians._anchor, {}) v = state.get("exp_avg_sq", torch.ones_like(grad_t) * 1e-8) lr = 0.0 for pg in self.gaussians.optimizer.param_groups: if pg["name"] == "anchor": lr = pg["lr"] break u_t = (lr / (torch.sqrt(v) + 1e-8)) * grad_t u_p = (lr / (torch.sqrt(v) + 1e-8)) * grad_p valid_mask = (torch.norm(u_t, dim=1) > 0) & (torch.norm(u_p, dim=1) > 0) if valid_mask.any(): grad_cos_sim = float(F.cosine_similarity(u_t[valid_mask], u_p[valid_mask], dim=1).mean()) parasitic_ratio = float(torch.norm(u_p, dim=1).mean() / (torch.norm(u_t, dim=1).mean() + 1e-7)) self.gaussians.optimizer.zero_grad(set_to_none=True) loss.backward() else: loss.backward() with torch.no_grad(): if step < self.opt.update_until and step > self.opt.start_stat: self.gaussians.training_statis(render_pkg["viewspace_points"], render_pkg["neural_opacity"], render_pkg["visibility_filter"], render_pkg["selection_mask"], voxel_visible_mask) if step not in range(3000, 4000): if step > self.opt.update_from and step % self.opt.update_interval == 0: self.gaussians.adjust_anchor(check_interval=self.opt.update_interval, success_threshold=self.opt.success_threshold, grad_threshold=self.opt.densify_grad_threshold, min_opacity=self.opt.min_opacity) elif step == self.opt.update_until: if hasattr(self.gaussians, "opacity_accum"): del self.gaussians.opacity_accum if hasattr(self.gaussians, "offset_gradient_accum"): del self.gaussians.offset_gradient_accum if hasattr(self.gaussians, "offset_denom"): del self.gaussians.offset_denom torch.cuda.empty_cache() if step < self.opt.iterations: self.gaussians.optimizer.step() self.gaussians.optimizer.zero_grad(set_to_none=True) num_gaussians = self.gaussians.get_anchor.shape[0] * self.gaussians.n_offsets 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), "bit_per_param": float(bit_per_param) if bit_per_param is not None else 0.0, "bit_hash_grid": float(bit_hash_grid_val) } self.last_n_gaussians = num_gaussians return metrics, {} def render(self, camera): with torch.no_grad(): voxel_visible_mask = prefilter_voxel(camera, self.gaussians, self.pipe, self.background) render_pkg = native_render(camera, self.gaussians, self.pipe, self.background, visible_mask=voxel_visible_mask) return {"image": render_pkg["render"], "depth": render_pkg.get("depth", None)} def save(self, save_dir, step): self.scene.save(step) def load(self, model_path, iteration): self.scene = Scene(self.dataset, self.gaussians, load_iteration=iteration, shuffle=False) self.gaussians.eval() def get_spatial_centers(self): return self.gaussians.get_anchor def compute_physical_metrics(self, cameras=None): metrics = {} with torch.no_grad(): cam = cameras[0] if cameras and len(cameras) > 0 else self.viewpoint_stack[0] xyz, color, opacity, scaling, rot, _ = generate_neural_gaussians(cam, self.gaussians, visible_mask=None, is_training=False, step=30000) scales_2d = scaling[:, :2] if scaling.dim() > 1 and scaling.shape[1] >= 2 else scaling.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(opacity)) 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(rot.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") cam = cameras[0] if cameras and len(cameras) > 0 else self.viewpoint_stack[0] xyz, color, opacity, scaling, rot, _ = generate_neural_gaussians(cam, self.gaussians, visible_mask=None, is_training=False, step=30000) opacities = opacity.squeeze() sigma_sq = (scaling[:, :2].max(dim=1)[0].pow(2)) if scaling.shape[1] >= 2 else scaling.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