render_dynamic.py

#!/usr/bin/env python3
"""
Dynamic Scaffold-GS Rendering Script

[DESIGN] Canonical Scaffold-GS + Deformable Gaussian Architecture:
- Anchors are STATIC canonical scaffolding (time-invariant)
- View-MLPs are purely view-dependent (no time conditioning)  
- Deformation field predicts Gaussian-level temporal offsets: Δx(x_canonical, t)
"""

import os
import torch
import numpy as np
import subprocess
import json
import time
from tqdm import tqdm
import torchvision

# Select GPU with least memory usage
cmd = 'nvidia-smi -q -d Memory |grep -A4 GPU|grep Used'
result = subprocess.run(cmd, shell=True, stdout=subprocess.PIPE).stdout.decode().split('\n')
os.environ['CUDA_VISIBLE_DEVICES'] = str(np.argmin([int(x.split()[2]) for x in result[:-1]]))
os.system('echo $CUDA_VISIBLE_DEVICES')

from scene import Scene, GaussianModel, DeformModel
from gaussian_renderer import render, prefilter_voxel
from utils.general_utils import safe_state
from argparse import ArgumentParser
from arguments import ModelParams, PipelineParams, get_combined_args


def render_set(model_path, name, iteration, views, gaussians, deform, pipeline, background, is_6dof=False):
    """Render a set of views for dynamic scene"""
    render_path = os.path.join(model_path, name, "ours_{}".format(iteration), "renders")
    gts_path = os.path.join(model_path, name, "ours_{}".format(iteration), "gt")
    
    os.makedirs(render_path, exist_ok=True)
    os.makedirs(gts_path, exist_ok=True)
    
    t_list = []
    
    for idx, view in enumerate(tqdm(views, desc="Rendering progress")):
        torch.cuda.synchronize()
        t0 = time.time()
        
        # Get frame ID for temporal deformation
        fid = view.fid if hasattr(view, 'fid') else torch.tensor([0.0]).cuda()
        
        # Get time index for temporal appearance residual
        if hasattr(view, 'frame_idx'):
            time_index = view.frame_idx
        else:
            total_frames = len(views)
            time_index = int(fid.item() * (total_frames - 1)) if total_frames > 1 else 0
        
        # Prefilter visible voxels
        voxel_visible_mask = prefilter_voxel(view, gaussians, pipeline, background)
        
        # [DESIGN v4] Render with pipeline matching architecture diagram
        # Static MLP → Canonical Center, Canonical View-MLP → c_can/α_can
        # DeformNet → Δx/Δr/Δs, Appearance Residual MLP → Δc/Δα
        render_pkg = render(view, gaussians, pipeline, background,
                          visible_mask=voxel_visible_mask,
                          deform_network=deform, time_input=fid,
                          is_6dof=is_6dof,
                          time_index=time_index if gaussians.time_appearance_dim > 0 else None)
        
        torch.cuda.synchronize()
        t1 = time.time()
        t_list.append(t1 - t0)
        
        rendering = render_pkg["render"]
        gt = view.original_image[0:3, :, :]
        
        # Save images
        img_name = view.image_name if hasattr(view, 'image_name') else '{0:05d}'.format(idx)
        torchvision.utils.save_image(rendering, os.path.join(render_path, f'{img_name}.png'))
        torchvision.utils.save_image(gt, os.path.join(gts_path, f'{img_name}.png'))
    
    # Report FPS (skip first 5 for warmup)
    if len(t_list) > 5:
        t = np.array(t_list[5:])
        fps = 1.0 / t.mean()
        print(f'Test FPS: \033[1;35m{fps:.5f}\033[0m')


def render_sets(dataset: ModelParams, iteration: int, pipeline: PipelineParams, 
                skip_train: bool, skip_test: bool):
    """Render train and test sets for dynamic scene"""
    with torch.no_grad():
        # Create Gaussian model (no time embedding - view-MLPs are time-independent)
        gaussians = GaussianModel(
            dataset.feat_dim, 
            dataset.n_offsets, 
            dataset.voxel_size, 
            dataset.update_depth, 
            dataset.update_init_factor, 
            dataset.update_hierachy_factor, 
            dataset.use_feat_bank,
            dataset.appearance_dim, 
            dataset.ratio, 
            dataset.add_opacity_dist, 
            dataset.add_cov_dist, 
            dataset.add_color_dist,
            time_emb_dim=0  # [DESIGN] No time embedding in view-MLPs
        )
        
        # Load scene
        scene = Scene(dataset, gaussians, load_iteration=iteration, shuffle=False)
        gaussians.eval()
        
        # Create and load deformation model
        deform = DeformModel(is_blender=dataset.is_blender, is_6dof=dataset.is_6dof)
        deform.load_weights(dataset.model_path, iteration)
        deform.deform.eval()
        
        # Background color
        bg_color = [1, 1, 1] if dataset.white_background else [0, 0, 0]
        background = torch.tensor(bg_color, dtype=torch.float32, device="cuda")
        
        os.makedirs(dataset.model_path, exist_ok=True)
        
        if not skip_train:
            render_set(dataset.model_path, "train", scene.loaded_iter, 
                      scene.getTrainCameras(), gaussians, deform, pipeline, background,
                      is_6dof=dataset.is_6dof)
        
        if not skip_test:
            render_set(dataset.model_path, "test", scene.loaded_iter, 
                      scene.getTestCameras(), gaussians, deform, pipeline, background,
                      is_6dof=dataset.is_6dof)


if __name__ == "__main__":
    # Set up command line argument parser
    parser = ArgumentParser(description="Dynamic scene rendering script")
    model = ModelParams(parser, sentinel=True)
    pipeline = PipelineParams(parser)
    parser.add_argument("--iteration", default=-1, type=int)
    parser.add_argument("--skip_train", action="store_true")
    parser.add_argument("--skip_test", action="store_true")
    parser.add_argument("--quiet", action="store_true")
    args = get_combined_args(parser)
    
    print("Rendering dynamic scene: " + args.model_path)
    
    # Initialize system state (RNG)
    safe_state(args.quiet)
    
    render_sets(model.extract(args), args.iteration, pipeline.extract(args), 
                args.skip_train, args.skip_test)