methods/wrapper_bags.py

import os
import sys
import math
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.insert(0, '/root/autodl-tmp/BAGS_offy')
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

def tv_loss(grids):
    number_of_grids = grids.shape[0]
    h_tv_count = grids[:, :, 1:, :].shape[1] * grids[:, :, 1:, :].shape[2] * grids[:, :, 1:, :].shape[3]
    w_tv_count = grids[:, :, :, 1:].shape[1] * grids[:, :, :, 1:].shape[2] * grids[:, :, :, 1:].shape[3]
    h_tv = torch.pow((grids[:, :, 1:, :] - grids[:, :, :-1, :]), 2).sum()
    w_tv = torch.pow((grids[:, :, :, 1:] - grids[:, :, :, :-1]), 2).sum()
    return 2 * (h_tv / h_tv_count + w_tv / w_tv_count) / number_of_grids

def get_emb(sin_inp):
    emb = torch.stack((sin_inp.sin(), sin_inp.cos()), dim=-1)
    return torch.flatten(emb, -2, -1)

def get_2d_emb(batch_size, x, y, out_ch, device):
    out_ch = int(np.ceil(out_ch / 4) * 2)
    inv_freq = 1.0 / (10000 ** (torch.arange(0, out_ch, 2).float() / out_ch))
    pos_x = torch.arange(x, device=device).type(inv_freq.type()) * 2 * np.pi / x
    pos_y = torch.arange(y, device=device).type(inv_freq.type()) * 2 * np.pi / y
    sin_inp_x = torch.einsum("i,j->ij", pos_x, inv_freq)
    sin_inp_y = torch.einsum("i,j->ij", pos_y, inv_freq)
    emb_x = get_emb(sin_inp_x).unsqueeze(1)
    emb_y = get_emb(sin_inp_y)
    emb = torch.zeros((x, y, out_ch * 2), device=device)
    emb[:, :, : out_ch] = emb_x
    emb[:, :, out_ch : 2 * out_ch] = emb_y
    return emb[None, :, :, :].repeat(batch_size, 1, 1, 1)

@register_method("bags")
class BAGSWrapper(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.dataset = self.lp.extract(self.args)
        self.opt = self.op.extract(self.args)
        self.pipe = self.pp.extract(self.args)

        self.opt.iterations = 60000
        self.opt.ms_steps = 6000
        self.upsample_iter = [3000, 6000]

        # 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)

        inp_shape = [len(self.scene.getTrainCameras()), int(np.round(self.scene.orig_h/self.args.resolution)), int(np.round(self.scene.orig_w/self.args.resolution))]

        self.gaussians = GaussianModel(
            self.dataset.sh_degree, inp_shape,
            ks1=self.dataset.kernel_size1, ks2=self.dataset.kernel_size2,
            ks3=self.dataset.kernel_size3, ks_ss=self.dataset.kernel_size_ss,
            not_use_rgbd=self.opt.not_use_rgbd, not_use_pe=self.opt.not_use_pe
        )
        
        self.scene.load_gaussian(self.gaussians)
        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.trainCameras_scale4 = self.scene.getTrainCameras(scale=4.0).copy()
        self.trainCameras_scale2 = self.scene.getTrainCameras(scale=2.0).copy()
        self.trainCameras_scale1 = self.scene.getTrainCameras(scale=1.0).copy()
        
        self.viewpoint_stack = self.trainCameras_scale4.copy()
        self.gaussians.compute_3D_filter(cameras=self.trainCameras_scale4)

        self.unfold1 = torch.nn.Unfold(kernel_size=(self.dataset.kernel_size1, self.dataset.kernel_size1), padding=self.dataset.kernel_size1 // 2).cuda()
        self.unfold2 = torch.nn.Unfold(kernel_size=(self.dataset.kernel_size2, self.dataset.kernel_size2), padding=self.dataset.kernel_size2 // 2).cuda()
        self.unfold3 = torch.nn.Unfold(kernel_size=(self.dataset.kernel_size3, self.dataset.kernel_size3), padding=self.dataset.kernel_size3 // 2).cuda()
        if self.dataset.kernel_size_ss != self.dataset.kernel_size3:
            self.opt.use_another_mlp = True
            self.unfold_ss = torch.nn.Unfold(kernel_size=(self.dataset.kernel_size_ss, self.dataset.kernel_size_ss), padding=self.dataset.kernel_size_ss // 2).cuda()
        else:
            self.unfold_ss = self.unfold3

        self.last_n_gaussians = self.gaussians.get_xyz.shape[0]

    def train_iteration(self, step):
        ori_iter = step
        adj_iter = step
        
        if ori_iter == self.upsample_iter[0]:
            self.viewpoint_stack = self.trainCameras_scale2.copy()
            self.gaussians.compute_3D_filter(cameras=self.trainCameras_scale2)
        elif ori_iter == self.upsample_iter[1]:
            self.viewpoint_stack = self.trainCameras_scale1.copy()
            self.gaussians.compute_3D_filter(cameras=self.trainCameras_scale1)

        if ori_iter > self.opt.ms_steps:
            adj_iter = step - self.opt.ms_steps

        self.gaussians.update_learning_rate(adj_iter)
        if adj_iter % 1000 == 0:
            self.gaussians.oneupSHdegree()

        if not self.viewpoint_stack:
            if ori_iter >= self.upsample_iter[0] and ori_iter < self.upsample_iter[1]:
                self.viewpoint_stack = self.trainCameras_scale2.copy()
            elif ori_iter >= self.upsample_iter[1]:
                self.viewpoint_stack = self.trainCameras_scale1.copy()
            else:
                self.viewpoint_stack = self.trainCameras_scale4.copy()

        viewpoint_cam = self.viewpoint_stack.pop(random.randint(0, len(self.viewpoint_stack) - 1))

        subpixel_offset = None
        if self.dataset.ray_jitter:
            subpixel_offset = torch.rand((int(viewpoint_cam.image_height), int(viewpoint_cam.image_width), 2), dtype=torch.float32, device="cuda") - 0.5

        render_pkg = native_render(viewpoint_cam, self.gaussians, self.pipe, self.background, kernel_size=self.dataset.kernel_size, subpixel_offset=subpixel_offset)
        image, depth, viewspace_point_tensor, visibility_filter, radii = render_pkg["render"], render_pkg["depth"], render_pkg["viewspace_points"], render_pkg["visibility_filter"], render_pkg["radii"]

        gt_image = viewpoint_cam.original_image.cuda()
        loss_target = torch.tensor(0.0, device="cuda")
        loss_parasitic = torch.tensor(0.0, device="cuda")

        if adj_iter > 250:
            shuffle_rgb = image.unsqueeze(0)
            shuffle_depth = depth.unsqueeze(0) - depth.min()
            shuffle_depth = shuffle_depth / (shuffle_depth.max() + 1e-7)
            pos_enc = get_2d_emb(1, shuffle_rgb.shape[-2], shuffle_rgb.shape[-1], 16, torch.device("cuda"))

            if ori_iter < 3000:
                kernel_weights, mask = self.gaussians.mlp_rgb_ms(0, pos_enc, torch.cat([shuffle_rgb, shuffle_depth], 1).detach(), ori_iter)
                patches = self.unfold1(shuffle_rgb).view(1, 3, self.dataset.kernel_size1 ** 2, shuffle_rgb.shape[-2], shuffle_rgb.shape[-1])
            elif ori_iter >= 3000 and ori_iter < 6000:
                kernel_weights, mask = self.gaussians.mlp_rgb_ms(0, pos_enc, torch.cat([shuffle_rgb, shuffle_depth], 1).detach(), ori_iter)
                patches = self.unfold2(shuffle_rgb).view(1, 3, self.dataset.kernel_size2 ** 2, shuffle_rgb.shape[-2], shuffle_rgb.shape[-1])
            else:
                if (ori_iter > self.opt.ms_steps) and self.opt.use_another_mlp:
                    kernel_weights, mask = self.gaussians.mlp_rgb_ss(0, pos_enc, torch.cat([shuffle_rgb, shuffle_depth], 1).detach(), adj_iter)
                    patches = self.unfold_ss(shuffle_rgb).view(1, 3, self.dataset.kernel_size_ss ** 2, shuffle_rgb.shape[-2], shuffle_rgb.shape[-1])
                else:
                    kernel_weights, mask = self.gaussians.mlp_rgb_ms(0, pos_enc, torch.cat([shuffle_rgb, shuffle_depth], 1).detach(), ori_iter)
                    patches = self.unfold3(shuffle_rgb).view(1, 3, self.dataset.kernel_size3 ** 2, shuffle_rgb.shape[-2], shuffle_rgb.shape[-1])

            kernel_weights = kernel_weights.unsqueeze(1)
            rgb = torch.sum(patches * kernel_weights, 2)[0]
            mask = mask[0]
            blur_image = mask * rgb + (1 - mask) * image

            depthloss = self.opt.depth_loss_alpha * tv_loss(shuffle_depth) if self.opt.use_depth_loss else torch.tensor(0.0, device="cuda")
            maskloss = self.opt.mask_loss_alpha * mask.mean() if self.opt.use_mask_loss else torch.tensor(0.0, device="cuda")
            tvloss = self.opt.rgbtv_loss_alpha * tv_loss(shuffle_rgb) if self.opt.use_rgbtv_loss else torch.tensor(0.0, device="cuda")

            Ll1 = l1_loss(blur_image, gt_image)
            loss_target = (1.0 - self.opt.lambda_dssim) * Ll1
            loss_parasitic = self.opt.lambda_dssim * (1.0 - ssim(blur_image, gt_image)) + tvloss + maskloss + depthloss
        else:
            Ll1 = l1_loss(image, gt_image)
            loss_target = (1.0 - self.opt.lambda_dssim) * Ll1
            loss_parasitic = self.opt.lambda_dssim * (1.0 - ssim(image, gt_image))

        loss = loss_target + loss_parasitic

        grad_cos_sim = 0.0
        parasitic_ratio = 0.0
        metrics = {}

        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],
                "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.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
                metrics[f"sti_{group_name}"] = ti

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

        with torch.no_grad():
            if adj_iter < self.opt.densify_until_iter:
                self.gaussians.max_radii2D[visibility_filter] = torch.max(self.gaussians.max_radii2D[visibility_filter], radii[visibility_filter])
                self.gaussians.add_densification_stats(viewspace_point_tensor, visibility_filter)
                
                if adj_iter > self.opt.densify_from_iter and adj_iter % self.opt.densification_interval == 0:
                    size_threshold = 20 if adj_iter > self.opt.opacity_reset_interval else None
                    if ori_iter <= self.opt.ms_steps:
                        dgt = self.opt.init_dgt
                        min_op = self.opt.init_opacity if self.opt.init_opacity >= 0 else self.opt.min_opacity
                    else:
                        dgt = self.opt.densify_grad_threshold
                        min_op = self.opt.min_opacity
                    self.gaussians.densify_and_prune(dgt, min_op, self.scene.cameras_extent, size_threshold)
                    
                    if ori_iter >= self.upsample_iter[0] and ori_iter < self.upsample_iter[1]:
                        self.gaussians.compute_3D_filter(cameras=self.trainCameras_scale2)
                    elif ori_iter >= self.upsample_iter[1]:
                        self.gaussians.compute_3D_filter(cameras=self.trainCameras_scale1)
                    else:
                        self.gaussians.compute_3D_filter(cameras=self.trainCameras_scale4)

                if adj_iter % self.opt.opacity_reset_interval == 0 or (self.dataset.white_background and adj_iter == self.opt.densify_from_iter):
                    self.gaussians.reset_opacity()

            if adj_iter % 100 == 0 and adj_iter > self.opt.densify_until_iter:
                if adj_iter < (self.opt.iterations - self.opt.ms_steps) - 100:
                    if ori_iter >= self.upsample_iter[0] and ori_iter < self.upsample_iter[1]:
                        self.gaussians.compute_3D_filter(cameras=self.trainCameras_scale2)
                    elif ori_iter >= self.upsample_iter[1]:
                        self.gaussians.compute_3D_filter(cameras=self.trainCameras_scale1)
                    else:
                        self.gaussians.compute_3D_filter(cameras=self.trainCameras_scale4)

            if ori_iter < self.opt.iterations:
                self.gaussians.optimizer.step()
                self.gaussians.optimizer.zero_grad(set_to_none=True)

        num_gaussians = self.gaussians.get_xyz.shape[0]
        metrics.update({
            "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)
        })
        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):
        self.scene.save(step)

    def load(self, model_path, iteration):
        self.gaussians.load_ply(os.path.join(model_path, 'point_cloud', f'iteration_{iteration}', 'point_cloud.ply'))

    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