303 lines
15 KiB
Python
303 lines
15 KiB
Python
#
|
|
# 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
|
|
#
|
|
|
|
import os
|
|
import torch
|
|
from random import randint
|
|
from utils.loss_utils import l1_loss, ssim, l2_loss, lpips_loss
|
|
from gaussian_renderer import render, network_gui
|
|
import sys
|
|
from scene import Scene, GaussianModel
|
|
from utils.general_utils import safe_state
|
|
import uuid
|
|
from tqdm import tqdm
|
|
from utils.image_utils import psnr
|
|
from argparse import ArgumentParser, Namespace
|
|
from arguments import ModelParams, PipelineParams, OptimizationParams
|
|
from torch.utils.data import DataLoader
|
|
import lpips
|
|
try:
|
|
from torch.utils.tensorboard import SummaryWriter
|
|
TENSORBOARD_FOUND = True
|
|
except ImportError:
|
|
TENSORBOARD_FOUND = False
|
|
def scene_reconstruction(dataset, opt, pipe, testing_iterations, saving_iterations,
|
|
checkpoint_iterations, checkpoint, debug_from,
|
|
gaussians, scene, stage, tb_writer, train_iter):
|
|
first_iter = 0
|
|
gaussians.training_setup(opt)
|
|
if checkpoint:
|
|
(model_params, first_iter) = torch.load(checkpoint)
|
|
gaussians.restore(model_params, opt)
|
|
|
|
bg_color = [1, 1, 1] if dataset.white_background else [0, 0, 0]
|
|
background = torch.tensor(bg_color, dtype=torch.float32, device="cuda")
|
|
|
|
iter_start = torch.cuda.Event(enable_timing = True)
|
|
iter_end = torch.cuda.Event(enable_timing = True)
|
|
|
|
viewpoint_stack = None
|
|
ema_loss_for_log = 0.0
|
|
ema_psnr_for_log = 0.0
|
|
|
|
final_iter = train_iter
|
|
|
|
progress_bar = tqdm(range(first_iter, final_iter), desc="Training progress")
|
|
first_iter += 1
|
|
lpips_model = lpips.LPIPS(net="alex").cuda()
|
|
|
|
for iteration in range(first_iter, final_iter+1):
|
|
if network_gui.conn == None:
|
|
network_gui.try_connect()
|
|
while network_gui.conn != None:
|
|
try:
|
|
net_image_bytes = None
|
|
custom_cam, do_training, pipe.convert_SHs_python, pipe.compute_cov3D_python, keep_alive, scaling_modifer, time = network_gui.receive()
|
|
if custom_cam != None:
|
|
net_image = render(custom_cam, gaussians, pipe, background, scaling_modifer, stage="stage")["render"]
|
|
net_image_bytes = memoryview((torch.clamp(net_image, min=0, max=1.0) * 255).byte().permute(1, 2, 0).contiguous().cpu().numpy())
|
|
network_gui.send(net_image_bytes, dataset.source_path)
|
|
if do_training and ((iteration < int(opt.iterations)) or not keep_alive):
|
|
break
|
|
except Exception as e:
|
|
network_gui.conn = None
|
|
|
|
iter_start.record()
|
|
|
|
gaussians.update_learning_rate(iteration)
|
|
|
|
# Every 1000 its we increase the levels of SH up to a maximum degree
|
|
if iteration % 1000 == 0:
|
|
gaussians.oneupSHdegree()
|
|
|
|
# Pick a random Camera
|
|
if not viewpoint_stack:
|
|
viewpoint_stack = scene.getTrainCameras()
|
|
# if stage == "coarse":
|
|
# batch_size = 1
|
|
# else:
|
|
batch_size = 16
|
|
viewpoint_stack_loader = DataLoader(viewpoint_stack, batch_size=batch_size,shuffle=True,num_workers=32,collate_fn=list)
|
|
loader = iter(viewpoint_stack_loader)
|
|
# idx = randint(0, len(viewpoint_stack)-1)
|
|
# idx = iteration % scene.maxtime
|
|
try:
|
|
viewpoint_cams = next(loader)
|
|
# print(viewpoint_cam.image_name)
|
|
except StopIteration:
|
|
loader = iter(viewpoint_stack_loader)
|
|
# Render
|
|
if (iteration - 1) == debug_from:
|
|
pipe.debug = True
|
|
images = []
|
|
gt_images = []
|
|
radii_list = []
|
|
visibility_filter_list = []
|
|
for viewpoint_cam in viewpoint_cams:
|
|
render_pkg = render(viewpoint_cam, gaussians, pipe, background, stage=stage)
|
|
image, viewspace_point_tensor, visibility_filter, radii = render_pkg["render"], render_pkg["viewspace_points"], render_pkg["visibility_filter"], render_pkg["radii"]
|
|
images.append(image.unsqueeze(0))
|
|
gt_image = viewpoint_cam.original_image.cuda()
|
|
gt_images.append(gt_image.unsqueeze(0))
|
|
radii_list.append(radii.unsqueeze(0))
|
|
visibility_filter_list.append(visibility_filter.unsqueeze(0))
|
|
radii = torch.cat(radii_list,0).max(dim=0).values
|
|
visibility_filter = torch.cat(visibility_filter_list).any(dim=0)
|
|
image_tensor = torch.cat(images,0)
|
|
gt_image_tensor = torch.cat(gt_images,0)
|
|
# Loss
|
|
Ll1 = l1_loss(image_tensor, gt_image_tensor)
|
|
# Ll1 = l2_loss(image, gt_image)
|
|
if opt.weight_constraint_after != 0 or iteration < opt.weight_decay_iteration and stage == "fine":
|
|
position_error, rotation_error = gaussians.standard_constaint()
|
|
else:
|
|
position_error, rotation_error = 0, 0
|
|
psnr_ = psnr(image, gt_image).mean().double()
|
|
# norm
|
|
|
|
if iteration < opt.weight_decay_iteration:
|
|
weight_constraint = opt.weight_constraint_init - iteration*(opt.weight_constraint_init - opt.weight_constraint_after)/(opt.weight_decay_iteration )
|
|
ssim_loss = ssim(image,gt_image)
|
|
lpipsloss = lpips_loss(gt_image,image,lpips_model)
|
|
|
|
loss = (1.0 - opt.lambda_dssim) * Ll1 + opt.lambda_dssim * (1.0-ssim_loss) + weight_constraint * (position_error + rotation_error) + 0.1* lpipsloss
|
|
|
|
loss.backward()
|
|
|
|
iter_end.record()
|
|
|
|
with torch.no_grad():
|
|
# Progress bar
|
|
ema_loss_for_log = 0.4 * loss.item() + 0.6 * ema_loss_for_log
|
|
ema_psnr_for_log = 0.4 * psnr_ + 0.6 * ema_psnr_for_log
|
|
|
|
if iteration % 10 == 0:
|
|
progress_bar.set_postfix({"Loss": f"{ema_loss_for_log:.{7}f}",
|
|
"psnr": f"{psnr_:.{2}f}"})
|
|
progress_bar.update(10)
|
|
if iteration == opt.iterations:
|
|
progress_bar.close()
|
|
|
|
# Log and save
|
|
training_report(tb_writer, iteration, Ll1, loss, l1_loss, iter_start.elapsed_time(iter_end), testing_iterations, scene, render, [pipe, background], stage)
|
|
if (iteration in saving_iterations):
|
|
print("\n[ITER {}] Saving Gaussians".format(iteration))
|
|
scene.save(iteration)
|
|
|
|
# Densification
|
|
if iteration < opt.densify_until_iter :
|
|
# Keep track of max radii in image-space for pruning
|
|
gaussians.max_radii2D[visibility_filter] = torch.max(gaussians.max_radii2D[visibility_filter], radii[visibility_filter])
|
|
gaussians.add_densification_stats(viewspace_point_tensor, visibility_filter)
|
|
# if iteration % opt.densification_interval == 0 and stage == "fine":
|
|
# gaussians.update_deformation_table(0.05)
|
|
# else:
|
|
# gaussians.update_deformation_table(-1)
|
|
opacity_threshold = opt.opacity_threshold - iteration*(opt.opacity_threshold - opt.opacity_threshold_after)/(opt.densify_until_iter )
|
|
|
|
if iteration > opt.densify_from_iter and iteration % opt.densification_interval == 0 and gaussians._xyz.shape[0] < 800000:
|
|
size_threshold = 20 if iteration > opt.opacity_reset_interval else None
|
|
densify_threshold = opt.densify_grad_threshold - iteration*(opt.densify_grad_threshold - opt.densify_grad_threshold_after)/(opt.densify_until_iter )
|
|
gaussians.densify(densify_threshold, 0.05, scene.cameras_extent, size_threshold)
|
|
if iteration > opt.pruning_from_iter and iteration % opt.pruning_interval == 0:
|
|
size_threshold = 20 if iteration > opt.opacity_reset_interval else None
|
|
# pruning_threshold = 0.00005 if stage == "coarse" else 0.05
|
|
|
|
gaussians.prune(opt.densify_grad_threshold, 0.05, scene.cameras_extent, size_threshold)
|
|
# if iteration > opt.pruning_from_iter and iteration % opt.pruning_interval == 0:
|
|
|
|
# gaussians.grow(opt.densify_grad_threshold, 0.005, scene.cameras_extent, size_threshold)
|
|
if iteration % opt.opacity_reset_interval == 0 or (dataset.white_background and iteration == opt.densify_from_iter):
|
|
print("reset opacity")
|
|
gaussians.reset_opacity()
|
|
|
|
|
|
# Optimizer step
|
|
if iteration < opt.iterations:
|
|
gaussians.optimizer.step()
|
|
gaussians.optimizer.zero_grad(set_to_none = True)
|
|
|
|
if (iteration in checkpoint_iterations):
|
|
print("\n[ITER {}] Saving Checkpoint".format(iteration))
|
|
torch.save((gaussians.capture(), iteration), scene.model_path + "/chkpnt" + str(iteration) + ".pth")
|
|
def training(dataset, opt, pipe, testing_iterations, saving_iterations, checkpoint_iterations, checkpoint, debug_from, expname):
|
|
# first_iter = 0
|
|
tb_writer = prepare_output_and_logger(expname)
|
|
gaussians = GaussianModel(dataset.sh_degree)
|
|
dataset.model_path = args.model_path
|
|
|
|
scene = Scene(dataset, gaussians)
|
|
scene_reconstruction(dataset, opt, pipe, testing_iterations, saving_iterations,
|
|
checkpoint_iterations, checkpoint, debug_from,
|
|
gaussians, scene, "coarse", tb_writer, opt.coarse_iterations)
|
|
scene_reconstruction(dataset, opt, pipe, testing_iterations, saving_iterations,
|
|
checkpoint_iterations, checkpoint, debug_from,
|
|
gaussians, scene, "fine", tb_writer, opt.iterations)
|
|
|
|
def prepare_output_and_logger(expname):
|
|
if not args.model_path:
|
|
# if os.getenv('OAR_JOB_ID'):
|
|
# unique_str=os.getenv('OAR_JOB_ID')
|
|
# else:
|
|
# unique_str = str(uuid.uuid4())
|
|
unique_str = expname
|
|
|
|
args.model_path = os.path.join("./output/", unique_str)
|
|
# Set up output folder
|
|
print("Output folder: {}".format(args.model_path))
|
|
os.makedirs(args.model_path, exist_ok = True)
|
|
with open(os.path.join(args.model_path, "cfg_args"), 'w') as cfg_log_f:
|
|
cfg_log_f.write(str(Namespace(**vars(args))))
|
|
|
|
# Create Tensorboard writer
|
|
tb_writer = None
|
|
if TENSORBOARD_FOUND:
|
|
tb_writer = SummaryWriter(args.model_path)
|
|
else:
|
|
print("Tensorboard not available: not logging progress")
|
|
return tb_writer
|
|
|
|
def training_report(tb_writer, iteration, Ll1, loss, l1_loss, elapsed, testing_iterations, scene : Scene, renderFunc, renderArgs, stage):
|
|
if tb_writer:
|
|
tb_writer.add_scalar(f'{stage}/train_loss_patches/l1_loss', Ll1.item(), iteration)
|
|
tb_writer.add_scalar(f'{stage}/train_loss_patchestotal_loss', loss.item(), iteration)
|
|
tb_writer.add_scalar(f'{stage}/iter_time', elapsed, iteration)
|
|
|
|
|
|
# Report test and samples of training set
|
|
if iteration in testing_iterations:
|
|
torch.cuda.empty_cache()
|
|
validation_configs = ({'name': 'test', 'cameras' : [scene.getTestCameras()[idx % len(scene.getTestCameras())] for idx in range(10, 5000, 299)]},
|
|
{'name': 'train', 'cameras' : [scene.getTrainCameras()[idx % len(scene.getTrainCameras())] for idx in range(10, 5000, 299)]})
|
|
|
|
for config in validation_configs:
|
|
if config['cameras'] and len(config['cameras']) > 0:
|
|
l1_test = 0.0
|
|
psnr_test = 0.0
|
|
for idx, viewpoint in enumerate(config['cameras']):
|
|
image = torch.clamp(renderFunc(viewpoint, scene.gaussians,stage=stage, *renderArgs)["render"], 0.0, 1.0)
|
|
gt_image = torch.clamp(viewpoint.original_image.to("cuda"), 0.0, 1.0)
|
|
if tb_writer and (idx < 5):
|
|
tb_writer.add_images(stage + "/"+config['name'] + "_view_{}/render".format(viewpoint.image_name), image[None], global_step=iteration)
|
|
if iteration == testing_iterations[0]:
|
|
tb_writer.add_images(stage + "/"+config['name'] + "_view_{}/ground_truth".format(viewpoint.image_name), gt_image[None], global_step=iteration)
|
|
l1_test += l1_loss(image, gt_image).mean().double()
|
|
psnr_test += psnr(image, gt_image).mean().double()
|
|
psnr_test /= len(config['cameras'])
|
|
l1_test /= len(config['cameras'])
|
|
print("\n[ITER {}] Evaluating {}: L1 {} PSNR {}".format(iteration, config['name'], l1_test, psnr_test))
|
|
if tb_writer:
|
|
tb_writer.add_scalar(stage + "/"+config['name'] + '/loss_viewpoint - l1_loss', l1_test, iteration)
|
|
tb_writer.add_scalar(stage+"/"+config['name'] + '/loss_viewpoint - psnr', psnr_test, iteration)
|
|
|
|
if tb_writer:
|
|
tb_writer.add_histogram(f"{stage}/scene/opacity_histogram", scene.gaussians.get_opacity, iteration)
|
|
|
|
tb_writer.add_scalar(f'{stage}/total_points', scene.gaussians.get_xyz.shape[0], iteration)
|
|
tb_writer.add_scalar(f'{stage}/deformation_rate', scene.gaussians._deformation_table.sum()/scene.gaussians.get_xyz.shape[0], iteration)
|
|
tb_writer.add_histogram(f"{stage}/scene/motion_histogram", scene.gaussians._deformation_accum.mean(dim=-1)/100, iteration,max_bins=500)
|
|
|
|
torch.cuda.empty_cache()
|
|
|
|
if __name__ == "__main__":
|
|
# Set up command line argument parser
|
|
# torch.set_default_tensor_type('torch.FloatTensor')
|
|
|
|
parser = ArgumentParser(description="Training script parameters")
|
|
lp = ModelParams(parser)
|
|
op = OptimizationParams(parser)
|
|
pp = PipelineParams(parser)
|
|
parser.add_argument('--ip', type=str, default="127.0.0.1")
|
|
parser.add_argument('--port', type=int, default=6009)
|
|
parser.add_argument('--debug_from', type=int, default=-1)
|
|
parser.add_argument('--detect_anomaly', action='store_true', default=False)
|
|
parser.add_argument("--test_iterations", nargs="+", type=int, default=[i*500 for i in range(60)])
|
|
parser.add_argument("--save_iterations", nargs="+", type=int, default=[2000, 3000, 7_000, 14000, 21000, 30_000])
|
|
parser.add_argument("--quiet", action="store_true")
|
|
parser.add_argument("--checkpoint_iterations", nargs="+", type=int, default=[])
|
|
parser.add_argument("--start_checkpoint", type=str, default = None)
|
|
parser.add_argument("--expname", type=str, default = "")
|
|
args = parser.parse_args(sys.argv[1:])
|
|
args.save_iterations.append(args.iterations)
|
|
|
|
print("Optimizing " + args.model_path)
|
|
|
|
# Initialize system state (RNG)
|
|
safe_state(args.quiet)
|
|
|
|
# Start GUI server, configure and run training
|
|
network_gui.init(args.ip, args.port)
|
|
torch.autograd.set_detect_anomaly(args.detect_anomaly)
|
|
training(lp.extract(args), op.extract(args), pp.extract(args), args.test_iterations, args.save_iterations, args.checkpoint_iterations, args.start_checkpoint, args.debug_from, args.expname)
|
|
|
|
# All done
|
|
print("\nTraining complete.")
|