train.py

import os
import torch
import cv2
import kornia
from random import randint
from lib.utils.loss_utils import l1_loss, l2_loss, psnr, ssim
from lib.utils.img_utils import save_img_torch, visualize_depth_numpy
from lib.models.street_gaussian_renderer import StreetGaussianRenderer
from lib.models.street_gaussian_model import StreetGaussianModel
from lib.utils.general_utils import safe_state
from lib.utils.camera_utils import Camera
from lib.utils.cfg_utils import save_cfg
from lib.models.scene import Scene
from lib.datasets.dataset import Dataset
from lib.config import cfg
from tqdm import tqdm
from argparse import ArgumentParser, Namespace
from lib.utils.system_utils import searchForMaxIteration


try:
    from torch.utils.tensorboard import SummaryWriter
    TENSORBOARD_FOUND = True
except ImportError:
    TENSORBOARD_FOUND = False


def training():
    training_args = cfg.train
    optim_args = cfg.optim
    data_args = cfg.data

    start_iter = 0
    tb_writer = prepare_output_and_logger()
    dataset = Dataset()
    gaussians = StreetGaussianModel(dataset.scene_info.metadata)
    scene = Scene(gaussians=gaussians, dataset=dataset)

    gaussians.training_setup()
    try:
        if cfg.loaded_iter == -1:
            loaded_iter = searchForMaxIteration(cfg.trained_model_dir)
        else:
            loaded_iter = cfg.loaded_iter
        ckpt_path = os.path.join(cfg.trained_model_dir, f'iteration_{loaded_iter}.pth')
        state_dict = torch.load(ckpt_path)
        start_iter = state_dict['iter']
        print(f'Loading model from {ckpt_path}')
        gaussians.load_state_dict(state_dict)
    except:
        pass

    print(f'Starting from {start_iter}')
    save_cfg(cfg, cfg.model_path, epoch=start_iter)

    gaussians_renderer = StreetGaussianRenderer()

    iter_start = torch.cuda.Event(enable_timing = True)
    iter_end = torch.cuda.Event(enable_timing = True)

    ema_loss_for_log = 0.0
    ema_psnr_for_log = 0.0
    psnr_dict = {}
    progress_bar = tqdm(range(start_iter, training_args.iterations))
    start_iter += 1

    viewpoint_stack = None
    for iteration in range(start_iter, training_args.iterations + 1):
    
        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()

        # Every 1000 iterations upsample
        # if iteration % 1000 == 0:
        #     if resolution_scales:  
        #         scale = resolution_scales.pop()

        # if iteration >= cfg.optim.densify_from_iter and iteration <= cfg.optim.densify_until_iter:
        gaussians.set_flip() 

        # Pick a random Camera
        if not viewpoint_stack:
            viewpoint_stack = scene.getTrainCameras().copy()
        
        viewpoint_cam: Camera = viewpoint_stack.pop(randint(0, len(viewpoint_stack) - 1))

        gt_image = viewpoint_cam.original_image.cuda()
        if hasattr(viewpoint_cam, 'original_mask'):
            mask = viewpoint_cam.original_mask.cuda().bool()
        else:
            mask = torch.ones_like(gt_image[0:1]).bool()
        
        if hasattr(viewpoint_cam, 'original_sky_mask'):
            sky_mask = viewpoint_cam.original_sky_mask.cuda()
        else:
            sky_mask = None
            
        if hasattr(viewpoint_cam, 'original_ego_mask'):
            ego_mask = viewpoint_cam.original_ego_mask.cuda()
        else:
            ego_mask = None
            
        if ego_mask is not None:
            mask = mask & (~ego_mask.bool())
            
        if hasattr(viewpoint_cam, 'original_obj_bound'):
            obj_bound = viewpoint_cam.original_obj_bound.cuda().bool()
        else:
            obj_bound = torch.zeros_like(gt_image[0:1]).bool()
            
        if hasattr(viewpoint_cam, 'original_car_mask'):
            car_mask = viewpoint_cam.original_car_mask.cuda().bool()
        else:
            car_mask = None

        # 增加一个逻辑，如果我们不渲染 obj，那么在 mask 上加上 obj_bound（即不让 obj 影响 depth mask 和 l1 mask）
        # if not gaussians.include_obj:
        #     mask = mask & (~obj_bound.bool())
        # 现在的逻辑是读取 car loss，然后添加到 mask 上
        if not gaussians.include_obj:
            mask = mask & (~car_mask.bool())
        
        if iteration % 1000 == 0:  # 每1000次迭代保存一次
            os.makedirs(f"{cfg.model_path}/mask_debug", exist_ok=True)
            
            # 创建可视化图像
            mask_vis = torch.zeros(3, mask.shape[1], mask.shape[2], device=mask.device)
            if mask is not None:
                mask_vis[0] = mask.float().squeeze(0)  # 红色通道显示普通mask
            if sky_mask is not None:
                mask_vis[1] = sky_mask.float().squeeze(0)  # 绿色通道显示sky_mask
            if ego_mask is not None:
                mask_vis[2] = ego_mask.float().squeeze(0)  # 蓝色通道显示ego_mask
                
            # 保存mask可视化结果
            save_img_torch(mask_vis, f"{cfg.model_path}/mask_debug/mask_{iteration}.jpg")
            
            # 保存obj_bound可视化结果
            if obj_bound is not None:
                obj_bound_vis = obj_bound.float().repeat(3, 1, 1)
                save_img_torch(obj_bound_vis, f"{cfg.model_path}/mask_debug/obj_bound_{iteration}.jpg")
            
            # 在特定迭代次数时插入断点
            # if iteration == training_args.debug_from:
            # if iteration == 2000:
            #     import pdb; pdb.set_trace()
                
        if (iteration - 1) == training_args.debug_from:
            cfg.render.debug = True

        render_pkg = gaussians_renderer.render(viewpoint_cam, gaussians)
        image, acc, viewspace_point_tensor, visibility_filter, radii = render_pkg["rgb"], render_pkg['acc'], render_pkg["viewspace_points"], render_pkg["visibility_filter"], render_pkg["radii"]
        depth = render_pkg['depth'] # [1, H, W]
        
        scalar_dict = dict()
        # rgb loss
        Ll1 = l1_loss(image, gt_image, mask)
        scalar_dict['l1_loss'] = Ll1.item()
        loss = (1.0 - optim_args.lambda_dssim) * optim_args.lambda_l1 * Ll1 + optim_args.lambda_dssim * (1.0 - ssim(image, gt_image, mask=mask))

        # sky loss
        if optim_args.lambda_sky > 0 and gaussians.include_sky and sky_mask is not None:
            acc = torch.clamp(acc, min=1e-6, max=1.-1e-6)
            sky_loss = torch.where(sky_mask, -torch.log(1 - acc), -torch.log(acc)).mean()
            if len(optim_args.lambda_sky_scale) > 0:
                sky_loss *= optim_args.lambda_sky_scale[viewpoint_cam.meta['cam']]
            scalar_dict['sky_loss'] = sky_loss.item()
            loss += optim_args.lambda_sky * sky_loss

        # # ego car loss 该方案和原本的 ego mask 方案冲突，且未测试效果，之后尝试
        # if optim_args.lambda_ego > 0 and ego_mask is not None:
        #     acc = torch.clamp(acc, min=1e-6, max=1.-1e-6)
        #     ego_loss = torch.where(ego_mask, -torch.log(1 - acc), -torch.log(acc)).mean()
        #     scalar_dict['ego_loss'] = ego_loss.item()
        #     loss += optim_args.lambda_ego * ego_loss
            
        #     # 如果有ego车的原始图像，可以直接用来替换
        #     if hasattr(viewpoint_cam, 'original_ego_image'):
        #         ego_image = viewpoint_cam.original_ego_image.cuda()
        #         image = torch.where(ego_mask.repeat(3, 1, 1), ego_image, image)

        # semantic loss
        if optim_args.lambda_semantic > 0 and data_args.get('use_semantic', False) and 'semantic' in viewpoint_cam.meta:
            gt_semantic = viewpoint_cam.meta['semantic'].cuda().long() # [1, H, W]
            if torch.all(gt_semantic == -1):
                semantic_loss = torch.zeros_like(Ll1)
            else:
                semantic = render_pkg['semantic'].unsqueeze(0) # [1, S, H, W]
                semantic_loss = torch.nn.functional.cross_entropy(
                    input=semantic, 
                    target=gt_semantic,
                    ignore_index=-1, 
                    reduction='mean'
                )
            scalar_dict['semantic_loss'] = semantic_loss.item()
            loss += optim_args.lambda_semantic * semantic_loss
        
        if optim_args.lambda_reg > 0 and gaussians.include_obj and iteration >= optim_args.densify_until_iter:
            render_pkg_obj = gaussians_renderer.render_object(viewpoint_cam, gaussians)
            image_obj, acc_obj = render_pkg_obj["rgb"], render_pkg_obj['acc']
            acc_obj = torch.clamp(acc_obj, min=1e-6, max=1.-1e-6)

            # box_reg_loss = gaussians.get_box_reg_loss()
            # scalar_dict['box_reg_loss'] = box_reg_loss.item()
            # loss += optim_args.lambda_reg * box_reg_loss
            # 暂时关闭obj acc loss 不过好像不包含 obj也就不会开启这个
            obj_acc_loss = torch.where(obj_bound, 
                -(acc_obj * torch.log(acc_obj) +  (1. - acc_obj) * torch.log(1. - acc_obj)), 
                -torch.log(1. - acc_obj)).mean()
            scalar_dict['obj_acc_loss'] = obj_acc_loss.item()
            loss += optim_args.lambda_reg * obj_acc_loss
            # obj_acc_loss = -((acc_obj * torch.log(acc_obj) +  (1. - acc_obj) * torch.log(1. - acc_obj))).mean()
            # scalar_dict['obj_acc_loss'] = obj_acc_loss.item()
            # loss += optim_args.lambda_reg * obj_acc_loss
        
        # lidar depth loss
        if optim_args.lambda_depth_lidar > 0 and 'lidar_depth' in viewpoint_cam.meta:            
            lidar_depth = viewpoint_cam.meta['lidar_depth'].cuda() # [1, H, W]
            depth_mask = torch.logical_and((lidar_depth > 0.), mask)
            # depth_mask[obj_bound] = False
            if torch.nonzero(depth_mask).any():
                expected_depth = depth / (render_pkg['acc'] + 1e-10)  
                depth_error = torch.abs((expected_depth[depth_mask] - lidar_depth[depth_mask]))
                depth_error, _ = torch.topk(depth_error, int(0.95 * depth_error.size(0)), largest=False)
                lidar_depth_loss = depth_error.mean()
                scalar_dict['lidar_depth_loss'] = lidar_depth_loss
            else:
                lidar_depth_loss = torch.zeros_like(Ll1)  
            loss += optim_args.lambda_depth_lidar * lidar_depth_loss
                    
        # color correction loss
        if optim_args.lambda_color_correction > 0 and gaussians.use_color_correction:
            color_correction_reg_loss = gaussians.color_correction.regularization_loss(viewpoint_cam)
            scalar_dict['color_correction_reg_loss'] = color_correction_reg_loss.item()
            loss += optim_args.lambda_color_correction * color_correction_reg_loss
        
        # pose correction loss
        if optim_args.lambda_pose_correction > 0 and gaussians.use_pose_correction:
            pose_correction_reg_loss = gaussians.pose_correction.regularization_loss()
            scalar_dict['pose_correction_reg_loss'] = pose_correction_reg_loss.item()
            loss += optim_args.lambda_pose_correction * pose_correction_reg_loss
                    
        # scale flatten loss
        if optim_args.lambda_scale_flatten > 0:
            scale_flatten_loss = gaussians.background.scale_flatten_loss()
            scalar_dict['scale_flatten_loss'] = scale_flatten_loss.item()
            loss += optim_args.lambda_scale_flatten * scale_flatten_loss
        
        # opacity sparse loss
        if optim_args.lambda_opacity_sparse > 0:
            opacity = gaussians.get_opacity
            opacity = opacity.clamp(1e-6, 1-1e-6)
            log_opacity = opacity * torch.log(opacity)
            log_one_minus_opacity = (1-opacity) * torch.log(1 - opacity)
            sparse_loss = -1 * (log_opacity + log_one_minus_opacity)[visibility_filter].mean()
            scalar_dict['opacity_sparse_loss'] = sparse_loss.item()
            loss += optim_args.lambda_opacity_sparse * sparse_loss
                
        # normal loss
        if optim_args.lambda_normal_mono > 0 and 'mono_normal' in viewpoint_cam.meta and 'normals' in render_pkg:
            if sky_mask is None:
                normal_mask = mask
            else:
                normal_mask = torch.logical_and(mask, ~sky_mask)
                normal_mask = normal_mask.squeeze(0)
                normal_mask[:50] = False
                
            normal_gt = viewpoint_cam.meta['mono_normal'].permute(1, 2, 0).cuda() # [H, W, 3]
            R_c2w = viewpoint_cam.world_view_transform[:3, :3]
            normal_gt = torch.matmul(normal_gt, R_c2w.T) # to world space
            normal_pred = render_pkg['normals'].permute(1, 2, 0) # [H, W, 3]    
            
            normal_l1_loss = torch.abs(normal_pred[normal_mask] - normal_gt[normal_mask]).mean()
            normal_cos_loss = (1. - torch.sum(normal_pred[normal_mask] * normal_gt[normal_mask], dim=-1)).mean()
            scalar_dict['normal_l1_loss'] = normal_l1_loss.item()
            scalar_dict['normal_cos_loss'] = normal_cos_loss.item()
            normal_loss = normal_l1_loss + normal_cos_loss
            loss += optim_args.lambda_normal_mono * normal_loss
            
        scalar_dict['loss'] = loss.item()

        loss.backward()
        
        iter_end.record()
                
        is_save_images = True
        if is_save_images and (iteration % 1000 == 0):
            # row0: gt_image, image, depth
            # row1: acc, image_obj, acc_obj
            depth_colored, _ = visualize_depth_numpy(depth.detach().cpu().numpy().squeeze(0))
            depth_colored = depth_colored[..., [2, 1, 0]] / 255.
            depth_colored = torch.from_numpy(depth_colored).permute(2, 0, 1).float().cuda()
            row0 = torch.cat([gt_image, image, depth_colored], dim=2)
            acc = acc.repeat(3, 1, 1)
            with torch.no_grad():
                render_pkg_obj = gaussians_renderer.render_object(viewpoint_cam, gaussians)
                image_obj, acc_obj = render_pkg_obj["rgb"], render_pkg_obj['acc']
            acc_obj = acc_obj.repeat(3, 1, 1)
            row1 = torch.cat([acc, image_obj, acc_obj], dim=2)
            image_to_show = torch.cat([row0, row1], dim=1)
            image_to_show = torch.clamp(image_to_show, 0.0, 1.0)
            os.makedirs(f"{cfg.model_path}/log_images", exist_ok = True)
            save_img_torch(image_to_show, f"{cfg.model_path}/log_images/{iteration}.jpg")
        
        with torch.no_grad():
            # Log
            tensor_dict = dict()

            # Progress bar
            ema_loss_for_log = 0.4 * loss.item() + 0.6 * ema_loss_for_log
            ema_psnr_for_log = 0.4 * psnr(image, gt_image, mask).mean().float() + 0.6 * ema_psnr_for_log
            if viewpoint_cam.id not in psnr_dict:
                psnr_dict[viewpoint_cam.id] = psnr(image, gt_image, mask).mean().float()
            else:
                psnr_dict[viewpoint_cam.id] = 0.4 * psnr(image, gt_image, mask).mean().float() + 0.6 * psnr_dict[viewpoint_cam.id]
            if iteration % 10 == 0:
                progress_bar.set_postfix({"Exp": f"{cfg.task}-{cfg.exp_name}", 
                                          "Loss": f"{ema_loss_for_log:.{7}f},", 
                                          "PSNR": f"{ema_psnr_for_log:.{4}f}"})
                progress_bar.update(10)
            if iteration == training_args.iterations:
                progress_bar.close()

            # Log and save
            if (iteration in training_args.save_iterations):
                print("\n[ITER {}] Saving Gaussians".format(iteration))
                scene.save(iteration)

            # Densification
            if iteration < optim_args.densify_until_iter:
                gaussians.set_visibility(include_list=list(set(gaussians.model_name_id.keys()) - set(['sky'])))
                gaussians.parse_camera(viewpoint_cam)   
                gaussians.set_max_radii2D(radii, visibility_filter)
                gaussians.add_densification_stats(viewspace_point_tensor, visibility_filter)
                
                prune_big_points = iteration > optim_args.opacity_reset_interval  # TODO 这个步骤会报错

                if iteration > optim_args.densify_from_iter:
                    if iteration % optim_args.densification_interval == 0:
                        scalars, tensors = gaussians.densify_and_prune(
                            max_grad=optim_args.densify_grad_threshold,
                            min_opacity=optim_args.min_opacity,
                            prune_big_points=prune_big_points,
                        )

                        scalar_dict.update(scalars)
                        tensor_dict.update(tensors)
                        
            # Reset opacity
            if iteration < optim_args.densify_until_iter:
                if iteration % optim_args.opacity_reset_interval == 0:
                    gaussians.reset_opacity()
                if data_args.white_background and iteration == optim_args.densify_from_iter:
                    gaussians.reset_opacity()

            training_report(tb_writer, iteration, scalar_dict, tensor_dict, training_args.test_iterations, scene, gaussians_renderer) # Forbiden due to limited spcae

            # Optimizer step
            if iteration < training_args.iterations:
                gaussians.update_optimizer()

            if (iteration in training_args.checkpoint_iterations):
                print("\n[ITER {}] Saving Checkpoint".format(iteration))
                state_dict = gaussians.save_state_dict(is_final=(iteration == training_args.iterations))
                state_dict['iter'] = iteration
                ckpt_path = os.path.join(cfg.trained_model_dir, f'iteration_{iteration}.pth')
                torch.save(state_dict, ckpt_path)



def prepare_output_and_logger():
    
    # if cfg.model_path == '':
    #     if os.getenv('OAR_JOB_ID'):
    #         unique_str = os.getenv('OAR_JOB_ID')
    #     else:
    #         unique_str = str(uuid.uuid4())
    #     cfg.model_path = os.path.join("./output/", unique_str[0:10])

    # Set up output folder
    print("Output folder: {}".format(cfg.model_path))

    os.makedirs(cfg.model_path, exist_ok=True)
    os.makedirs(cfg.trained_model_dir, exist_ok=True)
    os.makedirs(cfg.record_dir, exist_ok=True)
    if not cfg.resume:
        os.system('rm -rf {}/*'.format(cfg.record_dir))
        os.system('rm -rf {}/*'.format(cfg.trained_model_dir))

    with open(os.path.join(cfg.model_path, "cfg_args"), 'w') as cfg_log_f:
        viewer_arg = dict()
        viewer_arg['sh_degree'] = cfg.model.gaussian.sh_degree
        viewer_arg['white_background'] = cfg.data.white_background
        viewer_arg['source_path'] = cfg.source_path
        viewer_arg['model_path']= cfg.model_path
        cfg_log_f.write(str(Namespace(**viewer_arg)))

    # Create Tensorboard writer
    tb_writer = None
    if TENSORBOARD_FOUND:
        tb_writer = SummaryWriter(cfg.record_dir)
    else:
        print("Tensorboard not available: not logging progress")
    return tb_writer

def training_report(tb_writer, iteration, scalar_stats, tensor_stats, testing_iterations, scene: Scene, renderer: StreetGaussianRenderer):
    if tb_writer:
        try:
            for key, value in scalar_stats.items():
                tb_writer.add_scalar('train/' + key, value, iteration)
            for key, value in tensor_stats.items():
                tb_writer.add_histogram('train/' + key, value, iteration)
        except:
            print('Failed to write to tensorboard')
            
            
    # Report test and samples of training set
    if iteration in testing_iterations:
        torch.cuda.empty_cache()
        validation_configs = ({'name': 'test/test_view', 'cameras' : scene.getTestCameras()},
                              {'name': 'test/train_view', '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
                for idx, viewpoint in enumerate(config['cameras']):
                    image = torch.clamp(renderer.render(viewpoint, scene.gaussians)["rgb"], 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(config['name'] + "_{}/render".format(viewpoint.image_name), image[None], global_step=iteration)
                        if iteration == testing_iterations[0]:
                            tb_writer.add_images(config['name'] + "_{}/ground_truth".format(viewpoint.image_name), gt_image[None], global_step=iteration)
                    
                    if hasattr(viewpoint, 'original_mask'):
                        mask = viewpoint.original_mask.cuda().bool()
                    else:
                        mask = torch.ones_like(gt_image[0]).bool()
                    
                    if hasattr(viewpoint, 'original_ego_mask'):
                        ego_mask = viewpoint.original_ego_mask.cuda()
                    else:
                        ego_mask = None    
                        
                    if ego_mask is not None:
                        mask = mask & (~ego_mask.bool())
                        
                    l1_test += l1_loss(image, gt_image, mask).mean().double()
                    psnr_test += psnr(image, gt_image, mask).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(config['name'] + '/loss_viewpoint - l1_loss', l1_test, iteration)
                    tb_writer.add_scalar(config['name'] + '/loss_viewpoint - psnr', psnr_test, iteration)

        if tb_writer:
            tb_writer.add_histogram("test/opacity_histogram", scene.gaussians.get_opacity, iteration)
            tb_writer.add_scalar('test/points_total', scene.gaussians.get_xyz.shape[0], iteration)
        torch.cuda.empty_cache()

if __name__ == "__main__":
    print("Optimizing " + cfg.model_path)

    # Initialize system state (RNG)
    safe_state(cfg.train.quiet)

    # Start GUI server, configure and run training
    torch.autograd.set_detect_anomaly(cfg.train.detect_anomaly)
    training()

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