train.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
#

import os
import torch
from random import randint
from utils.loss_utils import l1_loss, ssim
from gaussian_renderer import render, network_gui
import sys
from scene import Scene, GaussianModel
from utils.general_utils import safe_state, get_expon_lr_func
from utils.graphics_utils import fov2focal
import uuid
from tqdm import tqdm
from utils.image_utils import psnr
from argparse import ArgumentParser, Namespace
from arguments import ModelParams, PipelineParams, OptimizationParams
import lpips
from pytorch_msssim import ssim as ssim_metric
try:
    from torch.utils.tensorboard import SummaryWriter
    TENSORBOARD_FOUND = True
except ImportError:
    TENSORBOARD_FOUND = False

try:
    from fused_ssim import fused_ssim
    FUSED_SSIM_AVAILABLE = True
except:
    FUSED_SSIM_AVAILABLE = False

try:
    from diff_gaussian_rasterization import SparseGaussianAdam
    SPARSE_ADAM_AVAILABLE = True
except:
    SPARSE_ADAM_AVAILABLE = False

# ========== DEBUG MODE: Enable/disable individual losses ==========
# Set these to False to disable specific losses for debugging
DEBUG_ENABLE_DEPTH_LOSS = False
DEBUG_ENABLE_NORMAL_CONSISTENCY_LOSS = False
DEBUG_ENABLE_SCALE_FLATTENING_LOSS = False
DEBUG_ENABLE_NORMAL_ALIGNMENT_LOSS = False
# ====================================================================

def normal_from_invdepth(invdepth, eps=1e-6):
    """
    Approximate per-pixel surface normals from an inverse depth map using image-space gradients.
    Returns a tensor of shape [3, H, W] with unit-length normals.
    """
    if invdepth.dim() == 3:
        invdepth = invdepth.squeeze(0)

    H, W = invdepth.shape[-2], invdepth.shape[-1]
    # Finite differences
    dzdx = torch.zeros_like(invdepth)
    dzdy = torch.zeros_like(invdepth)

    dzdx[:, :-1] = invdepth[:, 1:] - invdepth[:, :-1]
    dzdx[:, -1] = dzdx[:, -2]

    dzdy[:-1, :] = invdepth[1:, :] - invdepth[:-1, :]
    dzdy[-1, :] = dzdy[-2, :]

    # Construct normals in camera/image space: (-dz/dx, -dz/dy, 1)
    nx = -dzdx
    ny = -dzdy
    nz = torch.ones_like(invdepth)

    n = torch.stack([nx, ny, nz], dim=0)  # [3, H, W]
    norm = torch.sqrt((n * n).sum(dim=0, keepdim=True) + eps)
    n = n / norm
    return n

def normal_consistency_loss(render_invdepth, mono_invdepth, depth_mask, eps=1e-6):
    """
    L_nc = mean(1 - n_render · n_depth) over valid pixels,
    where normals are derived from inverse depth maps by image-space gradients.
    """
    # Ensure HxW
    if render_invdepth.dim() == 3:
        render_invdepth = render_invdepth.squeeze(0)
    if mono_invdepth.dim() == 3:
        mono_invdepth = mono_invdepth.squeeze(0)
    if depth_mask.dim() == 3:
        depth_mask = depth_mask.squeeze(0)

    device = render_invdepth.device
    n_render = normal_from_invdepth(render_invdepth)
    n_depth = normal_from_invdepth(mono_invdepth)

    valid = depth_mask > 0
    if not valid.any():
        return torch.zeros((), device=device, dtype=render_invdepth.dtype)

    # Flatten valid positions
    n_render_flat = n_render[:, valid]  # [3, N]
    n_depth_flat = n_depth[:, valid]    # [3, N]

    dots = (n_render_flat * n_depth_flat).sum(dim=0)
    dots = torch.clamp(dots, -1.0, 1.0)

    return (1.0 - dots).mean()

def pcc_patch_loss(render_invdepth, mono_invdepth, depth_mask, num_patches=8, patch_size=32, min_valid_frac=0.5, eps=1e-6):
    """
    Pearson-correlation depth regularization on random (approximately) non-overlapping patches.
    Returns: mean(1 - PCC) over valid patches.
    """
    # Ensure HxW
    if render_invdepth.dim() == 3:
        render_invdepth = render_invdepth.squeeze(0)
    if mono_invdepth.dim() == 3:
        mono_invdepth = mono_invdepth.squeeze(0)
    if depth_mask.dim() == 3:
        depth_mask = depth_mask.squeeze(0)

    H, W = render_invdepth.shape[-2], render_invdepth.shape[-1]
    if H < patch_size or W < patch_size:
        return torch.zeros((), device=render_invdepth.device, dtype=render_invdepth.dtype)

    device = render_invdepth.device
    used = torch.zeros((H, W), dtype=torch.bool, device=device)

    losses = []
    max_tries = num_patches * 50
    tries = 0

    while len(losses) < num_patches and tries < max_tries:
        tries += 1
        y0 = int(torch.randint(0, H - patch_size + 1, (1,), device=device).item())
        x0 = int(torch.randint(0, W - patch_size + 1, (1,), device=device).item())

        if used[y0:y0 + patch_size, x0:x0 + patch_size].any():
            continue

        m = depth_mask[y0:y0 + patch_size, x0:x0 + patch_size] > 0
        valid = int(m.sum().item())
        if valid < int(min_valid_frac * patch_size * patch_size):
            continue

        a = render_invdepth[y0:y0 + patch_size, x0:x0 + patch_size][m]
        b = mono_invdepth[y0:y0 + patch_size, x0:x0 + patch_size][m]

        a = a - a.mean()
        b = b - b.mean()

        if a.std() < 1e-5 or b.std() < 1e-5:
            continue

        var_a = (a * a).mean()
        var_b = (b * b).mean()

        denom = torch.sqrt(var_a * var_b) + eps
        pcc = (a * b).mean() / denom
        pcc = torch.clamp(pcc, -1.0, 1.0)

        losses.append(1.0 - pcc)

        used[y0:y0 + patch_size, x0:x0 + patch_size] = True

    if len(losses) == 0:
        return torch.zeros((), device=device, dtype=render_invdepth.dtype)

    return torch.stack(losses).mean()

def training(dataset, opt, pipe, testing_iterations, saving_iterations, checkpoint_iterations, checkpoint, debug_from):

    if not SPARSE_ADAM_AVAILABLE and opt.optimizer_type == "sparse_adam":
        sys.exit(f"Trying to use sparse adam but it is not installed, please install the correct rasterizer using pip install [3dgs_accel].")

    first_iter = 0
    tb_writer = prepare_output_and_logger(dataset)
    gaussians = GaussianModel(dataset.sh_degree, opt.optimizer_type)
    scene = Scene(dataset, gaussians)
    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)

    use_sparse_adam = opt.optimizer_type == "sparse_adam" and SPARSE_ADAM_AVAILABLE 
    depth_l1_weight = get_expon_lr_func(opt.depth_l1_weight_init, opt.depth_l1_weight_final, max_steps=opt.iterations)

    viewpoint_stack = scene.getTrainCameras().copy()
    viewpoint_indices = list(range(len(viewpoint_stack)))
    ema_loss_for_log = 0.0
    ema_Ll1depth_for_log = 0.0
    ema_Ldnormal_for_log = 0.0
    ema_Lscale_for_log = 0.0
    ema_Lnormal_for_log = 0.0

    progress_bar = tqdm(range(first_iter, opt.iterations), desc="Training progress", ncols=200)

    lpips_fn = lpips.LPIPS(net='vgg').cuda()
    lpips_fn.eval()

    first_iter += 1
    for iteration in range(first_iter, opt.iterations + 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 = network_gui.receive()
                if custom_cam != None:
                    net_image = render(custom_cam, gaussians, pipe, background, scaling_modifier=scaling_modifer, use_trained_exp=dataset.train_test_exp, separate_sh=SPARSE_ADAM_AVAILABLE)["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().copy()
            viewpoint_indices = list(range(len(viewpoint_stack)))
        rand_idx = randint(0, len(viewpoint_indices) - 1)
        viewpoint_cam = viewpoint_stack.pop(rand_idx)
        vind = viewpoint_indices.pop(rand_idx)

        # Render
        if (iteration - 1) == debug_from:
            pipe.debug = True

        bg = torch.rand((3), device="cuda") if opt.random_background else background

        render_pkg = render(viewpoint_cam, gaussians, pipe, bg, use_trained_exp=dataset.train_test_exp, separate_sh=SPARSE_ADAM_AVAILABLE)
        image, viewspace_point_tensor, visibility_filter, radii = render_pkg["render"], render_pkg["viewspace_points"], render_pkg["visibility_filter"], render_pkg["radii"]

        if viewpoint_cam.alpha_mask is not None:
            alpha_mask = viewpoint_cam.alpha_mask.cuda()
            image *= alpha_mask

        # Loss
        gt_image = viewpoint_cam.original_image.cuda()
        Ll1 = l1_loss(image, gt_image)
        if FUSED_SSIM_AVAILABLE:
            ssim_value = fused_ssim(image.unsqueeze(0), gt_image.unsqueeze(0))
        else:
            ssim_value = ssim(image, gt_image)

        loss = (1.0 - opt.lambda_dssim) * Ll1 + opt.lambda_dssim * (1.0 - ssim_value)

        # Depth / normal / scale regularization
        Ll1depth_pure = 0.0
        Ll1depth = torch.tensor(0.0, device="cuda")
        Ldnormal = torch.tensor(0.0, device="cuda")
        Lscale = torch.tensor(0.0, device="cuda")
        Lnormal = torch.tensor(0.0, device="cuda")
        if viewpoint_cam.depth_reliable:
            invDepth = render_pkg["depth"]
            mono_invdepth = 1 - viewpoint_cam.invdepthmap.cuda()
            depth_mask = viewpoint_cam.depth_mask.cuda()
            debug_display = False

            if DEBUG_ENABLE_DEPTH_LOSS:
                Ll1depth_pure = pcc_patch_loss(invDepth, mono_invdepth, depth_mask, num_patches=16, patch_size=64)
                lambda_nc = 0.1
                Ll1depth = lambda_nc * Ll1depth_pure
                # Ll1depth = depth_l1_weight(iteration) * Ll1depth_pure 
                loss += Ll1depth
                if iteration % 1000 == 0 and debug_display:
                    import matplotlib.pyplot as plt
                    plt.subplot(121)
                    plt.imshow(invDepth.squeeze().detach().cpu().numpy())
                    plt.title("render")
                    plt.subplot(122)
                    plt.imshow(mono_invdepth.squeeze().detach().cpu().numpy())
                    plt.title("mono")
                    plt.show()

            # Normal consistency loss between normals from rendered and monocular depth
            if DEBUG_ENABLE_NORMAL_CONSISTENCY_LOSS:
                Lnc = normal_consistency_loss(invDepth, mono_invdepth, depth_mask)
                lambda_nc = 0.1
                Ldnormal = lambda_nc * Lnc
                loss += Ldnormal
                if iteration % 1000 == 0 and debug_display:
                    import matplotlib.pyplot as plt
                    n_render = normal_from_invdepth(invDepth)
                    n_depth = normal_from_invdepth(mono_invdepth)
                    plt.subplot(121)
                    plt.imshow(((n_render.permute(1,2,0)+1)/2).detach().cpu().numpy())
                    plt.title("render normal")
                    plt.subplot(122)
                    plt.imshow(((n_depth.permute(1,2,0)+1)/2).detach().cpu().numpy())
                    plt.title("mono normal")
                    plt.show()
        
        # Scale flattening loss Ls for line Gaussians (type 2)
        # Encourages min(s1, s2, s3) to be small so Gaussians become disc-like.
        if DEBUG_ENABLE_SCALE_FLATTENING_LOSS:
            line_mask = (gaussians._point_types == 2)
            if line_mask.any():
                scales = gaussians.get_scaling[line_mask]  # [N_line, 3]
                min_scale = scales.min(dim=1).values
                Ls = min_scale.abs().mean()
                lambda_s = 100.0
                Lscale = lambda_s * Ls
                loss += Lscale

        # Normal alignment loss L_n: line Gaussian flat normals vs monocular normal map (when -n provided)
        if DEBUG_ENABLE_NORMAL_ALIGNMENT_LOSS and getattr(viewpoint_cam, "mono_normal_map", None) is not None:
            xyz_line, n_flat_world = gaussians.get_flat_normals_line()
            if xyz_line is not None and xyz_line.shape[0] > 0:
                W2C = viewpoint_cam.world_view_transform
                R_w2c = W2C[:3, :3]
                t_w2c = W2C[:3, 3]
                xyz_cam = (R_w2c @ xyz_line.T).T + t_w2c
                z = xyz_cam[:, 2]
                view_cam = xyz_cam / (xyz_cam.norm(dim=1, keepdim=True) + 1e-8)
                n_cam = (R_w2c @ n_flat_world.T).T
                flip = (n_cam * view_cam).sum(dim=1) < 0
                n_cam = n_cam.clone()
                n_cam[flip] = -n_cam[flip]
                W, H = viewpoint_cam.image_width, viewpoint_cam.image_height
                fx = fov2focal(viewpoint_cam.FoVx, W)
                fy = fov2focal(viewpoint_cam.FoVy, H)
                cx, cy = W / 2.0, H / 2.0
                u = (fx * xyz_cam[:, 0] / z + cx).long().clamp(0, W - 1)
                v = (fy * xyz_cam[:, 1] / z + cy).long().clamp(0, H - 1)
                valid = (z > 0.1) & (u >= 0) & (u < W) & (v >= 0) & (v < H)
                if valid.any():
                    nmap = viewpoint_cam.mono_normal_map  # [3, H, W]
                    n_mono = nmap[:, v[valid], u[valid]].T  # [N_valid, 3]
                    n_est = n_cam[valid]
                    n_est = torch.nn.functional.normalize(n_est, dim=1)
                    n_gt = torch.nn.functional.normalize(n_mono, dim=1)
                    Ln = torch.abs(n_est - n_gt).sum(dim=1).mean()
                    Lnormal = 0.5 * Ln
                    loss += Lnormal

        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_Ll1depth_for_log = 0.4 * Ll1depth.item() + 0.6 * ema_Ll1depth_for_log
            ema_Ldnormal_for_log = 0.4 * Ldnormal.item() + 0.6 * ema_Ldnormal_for_log
            ema_Lscale_for_log = 0.4 * Lscale.item() + 0.6 * ema_Lscale_for_log
            ema_Lnormal_for_log = 0.4 * Lnormal.item() + 0.6 * ema_Lnormal_for_log

            if iteration % 10 == 0:
                progress_bar.set_postfix({"Loss": f"{ema_loss_for_log:.{7}f}", "Depth Loss": f"{ema_Ll1depth_for_log:.{7}f}", "Depth Normal Loss": f"{ema_Ldnormal_for_log:.{7}f}", "Scale Loss": f"{ema_Lscale_for_log:.{7}f}", "Mono Normal Loss": f"{ema_Lnormal_for_log:.{7}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, 1., SPARSE_ADAM_AVAILABLE, None, dataset.train_test_exp), dataset.train_test_exp, lpips_fn)
            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.densify_from_iter and iteration % opt.densification_interval == 0:
                    size_threshold = 20 if iteration > opt.opacity_reset_interval else None
                    gaussians.densify_and_prune(
                        opt.densify_grad_threshold, 0.005, scene.cameras_extent, size_threshold, radii, iteration
                    )

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

            # Optimizer step
            if iteration < opt.iterations:
                gaussians.exposure_optimizer.step()
                gaussians.exposure_optimizer.zero_grad(set_to_none = True)
                if use_sparse_adam:
                    visible = radii > 0
                    gaussians.optimizer.step(visible, radii.shape[0])
                    gaussians.optimizer.zero_grad(set_to_none = True)
                else:
                    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 prepare_output_and_logger(args):    
    if not args.model_path:
        if os.getenv('OAR_JOB_ID'):
            unique_str=os.getenv('OAR_JOB_ID')
        else:
            unique_str = str(uuid.uuid4())
        args.model_path = os.path.join("./output/", unique_str[0:10])
        
    # 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, train_test_exp, lpips_fn):
    if tb_writer:
        tb_writer.add_scalar('train_loss_patches/l1_loss', Ll1.item(), iteration)
        tb_writer.add_scalar('train_loss_patches/total_loss', loss.item(), iteration)
        tb_writer.add_scalar('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()},
            {'name': 'train', 'cameras': [scene.getTrainCameras()[idx % len(scene.getTrainCameras())] for idx in range(5, 30, 5)]}
        )

        for config in validation_configs:
            if config['cameras'] and len(config['cameras']) > 0:
                l1_test = 0.0
                psnr_test = 0.0
                ssim_test = 0.0
                lpips_test = 0.0
                for idx, viewpoint in enumerate(config['cameras']):
                    image = torch.clamp(
                        renderFunc(viewpoint, scene.gaussians, *renderArgs)["render"],
                        0.0, 1.0
                    )
                    gt_image = torch.clamp(
                        viewpoint.original_image.to("cuda"),
                        0.0, 1.0
                    )
                    if train_test_exp:
                        image = image[..., image.shape[-1] // 2:]
                        gt_image = gt_image[..., gt_image.shape[-1] // 2:]
                    if tb_writer and (idx < 5):
                        tb_writer.add_images(
                            config['name'] + "_view_{}/render".format(viewpoint.image_name),
                            image[None],
                            global_step=iteration
                        )
                        if iteration == testing_iterations[0]:
                            tb_writer.add_images(
                                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()
                    ssim_val = ssim_metric(
                        image.unsqueeze(0),
                        gt_image.unsqueeze(0),
                        data_range=1.0
                    )
                    ssim_test += ssim_val.mean().double()
                    lpips_val = lpips_fn(
                        image.unsqueeze(0) * 2 - 1,
                        gt_image.unsqueeze(0) * 2 - 1
                    )
                    lpips_test += lpips_val.mean().double()
                num_views = len(config['cameras'])
                l1_test /= num_views
                psnr_test /= num_views
                ssim_test /= num_views
                lpips_test /= num_views
                print(
                    "\n[ITER {}] Evaluating {}: L1 {} PSNR {} SSIM {} LPIPS {}".format(
                        iteration,
                        config['name'],
                        l1_test,
                        psnr_test,
                        ssim_test,
                        lpips_test
                    )
                )
                if tb_writer:
                    tb_writer.add_scalar(
                        config['name'] + '/loss_viewpoint - l1_loss',
                        l1_test,
                        iteration
                    )
                    tb_writer.add_scalar(
                        config['name'] + '/loss_viewpoint - psnr',
                        psnr_test,
                        iteration
                    )
                    tb_writer.add_scalar(
                        config['name'] + '/loss_viewpoint - ssim',
                        ssim_test,
                        iteration
                    )
                    tb_writer.add_scalar(
                        config['name'] + '/loss_viewpoint - lpips',
                        lpips_test,
                        iteration
                    )

        if tb_writer:
            tb_writer.add_histogram(
                "scene/opacity_histogram",
                scene.gaussians.get_opacity,
                iteration
            )

            tb_writer.add_scalar(
                'total_points',
                scene.gaussians.get_xyz.shape[0],
                iteration
            )

        torch.cuda.empty_cache()

if __name__ == "__main__":
    # Set up command line argument parser
    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=[7_000, 30_000])
    parser.add_argument("--save_iterations", nargs="+", type=int, default=[7_000, 30_000])
    parser.add_argument("--quiet", action="store_true")
    parser.add_argument('--disable_viewer', action='store_true', default=False)
    parser.add_argument("--checkpoint_iterations", nargs="+", type=int, default=[])
    parser.add_argument("--start_checkpoint", type=str, default = None)
    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
    if not args.disable_viewer:
        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)

    # All done
    print("\nTraining complete.")