fps.py

#
# Copyright (C) 2023, Inria
# GRAPHDECO research group, https://team.inria.fr/graphdeco
# All rights reserved.
#
# This software is free for non-commercial, research and evaluation use 
# under the terms of the LICENSE.md file.
#
# For inquiries contact  george.drettakis@inria.fr
#
# Copyright (c) 2024-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.


import torch
from scene import Scene
import os
import yaml
import socket
import sys
import hashlib
from collections import defaultdict
from tqdm import tqdm
from os import makedirs
from gaussian_renderer import render
import torchvision
import torch.utils.benchmark as benchmark
from utils.general_utils import safe_state
from utils.system_utils import searchForMaxIteration
from argparse import ArgumentParser
from gaussian_renderer import GaussianModel
from utils.loader_utils import MultiViewVideoDataset, SequentialMultiviewSampler
from scene.decoders import DecoderIdentity
from arguments import ModelParams, PipelineParams, OptimizationParams, QuantizeParams, OptimizationParamsInitial, OptimizationParamsRest, get_combined_args
try:
    import wandb
    if not ('SLURM_PROCID' in os.environ and os.environ['SLURM_PROCID']!='0'):
        WANDB_FOUND = True
    else:
        WANDB_FOUND = False
except ImportError:
    WANDB_FOUND = False


def render_set(views, gaussians, pipeline, background):

    for idx, view in enumerate(tqdm(views, desc="Rendering progress")):
        rendering = render(view, gaussians, pipeline, background)["render"]
        # gt = view.original_image[0:3, :, :]
        # torchvision.utils.save_image(rendering, os.path.join(render_path, '{0:04d}'.format(idx) + ".png"))
        # torchvision.utils.save_image(gt, os.path.join(gts_path, '{0:04d}'.format(idx) + ".png"))

    # training(lp_args, op, pp_args, qp_args, args, args.skip_train, args.skip_test)

def render_fn(views, gaussians, pipeline, background, use_amp):
    with torch.autocast(device_type='cuda', dtype=torch.float16, enabled=False):
        for view in views:
            render(view, gaussians, pipeline, background)

def measure_fps(scene, gaussians, pipeline, background, use_amp=False):
    with torch.no_grad():
        views = scene.getTrainCameras() + scene.getTestCameras()
        t0 = benchmark.Timer(stmt='render_fn(views, gaussians, pipeline, background, use_amp)',
                            setup='from __main__ import render_fn',
                            globals={'views': views, 'gaussians': gaussians, 'pipeline': pipeline, 
                                    'background': background, 'use_amp': use_amp},
                            )
        time = t0.timeit(100)
        fps = len(views)/time.median
        print("Rendering FPS: ", fps)
    return fps
        

def measure_fps_decode(gaussians):
    with torch.no_grad():
        t0 = benchmark.Timer(stmt='gaussians.get_decoded_atts',
                            globals={'gaussians': gaussians},
                            )
        time = t0.timeit(100)
        fps = 1/time.median
        print("Decoding FPS: ", fps)
    return fps

def render_sets(dataset: ModelParams, opt: OptimizationParams, pipeline: PipelineParams, qp:QuantizeParams, args,
             skip_train: bool, skip_test: bool):
    
    with torch.no_grad():

        if not skip_train:
            # Create dataset and loader for training and testing at each time instance
            train_image_dataset = MultiViewVideoDataset(dataset.source_path, split='train', test_indices=dataset.test_indices,
                                                        max_frames=dataset.max_frames, start_idx=0)
            train_sampler = SequentialMultiviewSampler(train_image_dataset)
            train_loader = iter(torch.utils.data.DataLoader(train_image_dataset, batch_size=train_image_dataset.n_cams, 
                                                            sampler=train_sampler, num_workers=4))
        
        if not skip_test:
            test_image_dataset = MultiViewVideoDataset(dataset.source_path, split='test', test_indices=dataset.test_indices, 
                                                    max_frames=dataset.max_frames, start_idx=0)
            test_sampler = SequentialMultiviewSampler(test_image_dataset)
            test_loader = iter(torch.utils.data.DataLoader(test_image_dataset, batch_size=test_image_dataset.n_cams, 
                                                            sampler=test_sampler, num_workers=4))
        

        start_frame_idx = dataset.start_idx + 1
        # Fast forward data loading
        for frame_ff in range(0, start_frame_idx):
            if not skip_train:
                train_data = next(train_loader)
                train_images, train_paths = train_data
            if not skip_test:
                try:
                    test_data = next(test_loader)
                    test_images, test_paths = test_data
                except StopIteration:
                    print('No test cameras found, disabling testing.')
                    test_images, test_paths = None, None

        if not skip_train:
            train_image_data = {'image':train_images.cuda(),'path':train_paths,'frame_idx':0}
        else:
            train_image_data = None
        if not skip_test:
            test_image_data = {'image':test_images.cuda(),'path':test_paths,'frame_idx':0}
        else:
            test_image_data = None

        # Create the gaussian model and scene, initialized with frame 1 images from dataset
        qp.seed = dataset.seed
        gaussians = GaussianModel(dataset.sh_degree, qp)
        opt.set_params(start_frame_idx)
        # Setup training arguments
        scene = Scene(dataset, gaussians, 
                      train_image_data= train_image_data, test_image_data=test_image_data)
        
        checkpoint_path = os.path.join(args.model_path,'frames',str(start_frame_idx).zfill(4), 'ckpt.pth')
        print('Loading checkpoint at ', checkpoint_path)
        model_params, iteration, start_frame_idx, training_metrics = torch.load(checkpoint_path)

        gaussians.restore_fps(model_params, opt, start_frame_idx)
        gaussians.frame_idx = start_frame_idx

        for param_name in gaussians.param_names:
            if gaussians.gate_params[param_name]:
                att = gaussians.get_decoded_atts[param_name]
                latents = gaussians.get_atts[param_name]
                latents.data = att.data
                gaussians.latent_decoders[param_name] = DecoderIdentity()
                gaussians.gate_params[param_name] = False
        gaussians.gate_atts = None
                
        scene.model_path = os.path.join(args.model_path,'frames',str(start_frame_idx).zfill(4))
        scene.updateCameraImages(args, train_image_data, test_image_data, start_frame_idx, resolution_scales=[1.0])

        bg_color = [1,1,1] if dataset.white_background else [0, 0, 0]
        background = torch.tensor(bg_color, dtype=torch.float32, device="cuda")

        fps = measure_fps(scene, gaussians, pipeline, background, use_amp=False)
        decode_fps = measure_fps_decode(gaussians)
        return fps, decode_fps

def render_sets_decoded(dataset: ModelParams, opt: OptimizationParams, pipeline: PipelineParams, qp:QuantizeParams, args,
             skip_train: bool, skip_test: bool):
    
    with torch.no_grad():

        if not skip_train:
            # Create dataset and loader for training and testing at each time instance
            train_image_dataset = MultiViewVideoDataset(dataset.source_path, split='train', test_indices=dataset.test_indices,
                                                        max_frames=dataset.max_frames, start_idx=0)
            train_sampler = SequentialMultiviewSampler(train_image_dataset)
            train_loader = iter(torch.utils.data.DataLoader(train_image_dataset, batch_size=train_image_dataset.n_cams, 
                                                            sampler=train_sampler, num_workers=4))
        
        if not skip_test:
            test_image_dataset = MultiViewVideoDataset(dataset.source_path, split='test', test_indices=dataset.test_indices, 
                                                    max_frames=dataset.max_frames, start_idx=0)
            test_sampler = SequentialMultiviewSampler(test_image_dataset)
            test_loader = iter(torch.utils.data.DataLoader(test_image_dataset, batch_size=test_image_dataset.n_cams, 
                                                            sampler=test_sampler, num_workers=4))
        


        start_frame_idx = dataset.start_idx + 1
        # Fast forward data loading
        for frame_ff in range(0, start_frame_idx):
            if not skip_train:
                train_data = next(train_loader)
                train_images, train_paths = train_data
            if not skip_test:
                try:
                    test_data = next(test_loader)
                    test_images, test_paths = test_data
                except StopIteration:
                    print('No test cameras found, disabling testing.')
                    test_images, test_paths = None, None

        if not skip_train:
            train_image_data = {'image':train_images.cuda(),'path':train_paths,'frame_idx':0}
        else:
            train_image_data = None
        if not skip_test:
            test_image_data = {'image':test_images.cuda(),'path':test_paths,'frame_idx':0}
        else:
            test_image_data = None

        # Create the gaussian model and scene, initialized with frame 1 images from dataset
        qp.seed = dataset.seed
        gaussians = GaussianModel(dataset.sh_degree, qp)
        scene = Scene(dataset, gaussians, 
                      train_image_data= train_image_data, test_image_data=test_image_data)
        opt.set_params(start_frame_idx)
        # Setup training arguments
        gaussians.training_setup(opt)
        
        gaussians.frame_idx = start_frame_idx
        scene.model_path = os.path.join(args.model_path,'frames',str(start_frame_idx).zfill(4))
        scene.updateCameraImages(args, train_image_data, test_image_data, start_frame_idx, resolution_scales=[1.0])
        scene.loaded_iter = searchForMaxIteration(os.path.join(scene.model_path, "point_cloud"))
        print("Loading trained model at iteration {}".format(scene.loaded_iter))
        scene.gaussians.load_ply(os.path.join(scene.model_path,
                                                        "point_cloud",
                                                        "iteration_" + str(scene.loaded_iter),
                                                        "point_cloud.ply"))

        bg_color = [1,1,1] if dataset.white_background else [0, 0, 0]
        background = torch.tensor(bg_color, dtype=torch.float32, device="cuda")

        fps = measure_fps(scene, gaussians, pipeline, background, use_amp=False)

        return fps
        
if __name__ == "__main__":

    print('Running on ', socket.gethostname())
    # Config file is used for argument defaults. Command line arguments override config file.
    # testing
    config_path = sys.argv[sys.argv.index("--config")+1] if "--config" in sys.argv else None
    if config_path:
        with open(config_path, "r") as f:
            config = yaml.load(f, Loader=yaml.FullLoader)
    else:
        config = {}
    config = defaultdict(lambda: {}, config)

    # Set up command line argument parser
    parser = ArgumentParser(description="Training script parameters")

    lp = ModelParams(parser, config['model_params'])
    op_i = OptimizationParamsInitial(parser, config['opt_params_initial'])
    op_r = OptimizationParamsRest(parser, config['opt_params_rest'])
    pp = PipelineParams(parser, config['pipe_params'])
    qp = QuantizeParams(parser, config['quantize_params'])

    parser.add_argument('--config', type=str, default=None)
    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")
    parser.add_argument("--save_iterations", nargs="+", type=int, default=[])
    parser.add_argument("--checkpoint_iterations", nargs="+", type=int, default=[])
    args = parser.parse_args(sys.argv[1:])
    # args = get_combined_args(parser)
    # args.save_iterations.append(args.iterations)
    
    # Merge optimization args for initial and rest and change accordingly
    op = OptimizationParams(op_i.extract(args), op_r.extract(args))

    print("Rendering " + args.model_path)
    safe_state(args.quiet)

    lp_args = lp.extract(args)
    pp_args = pp.extract(args)
    qp_args = qp.extract(args)

    # render_sets(model.extract(args), args.iteration, pipeline.extract(args), args.skip_train, args.skip_test)
    fps, decode_fps = render_sets(lp_args, op, pp_args, qp_args, args, args.skip_train, args.skip_test)
    # vanilla_fps = render_sets_decoded(lp_args, op, pp_args, qp_args, args, args.skip_train, args.skip_test)

    wandb_enabled = WANDB_FOUND and lp_args.use_wandb
    
    if wandb_enabled:
        wandb_run_name = args.wandb_run_name
        wandb_entity = args.wandb_entity
        id = hashlib.md5(wandb_run_name.encode('utf-8')).hexdigest()
        name = wandb_run_name

        api = wandb.Api()

        run = api.from_path(os.path.join(wandb_entity, args.wandb_project, "runs", id))
        run.summary['FPS'] = fps
        run.summary['FPS_Decoding'] = decode_fps
        # run.summary['FPS_Vanilla'] = vanilla_fps
        # run.summary['FPS_Overall'] = 1/(1/decode_fps + 1/vanilla_fps)
        # print("Overall FPS: ", 1/(1/decode_fps + 1/vanilla_fps))
        run.summary.update()