# -------------------------------------------------------- # training code for DUSt3R # -------------------------------------------------------- import os os.environ['OMP_NUM_THREADS'] = '4' # will affect the performance of pairwise prediction import argparse import datetime import json import numpy as np import sys import time import math import wandb from collections import defaultdict from pathlib import Path from typing import Sized import torch import torch.backends.cudnn as cudnn from torch.utils.tensorboard import SummaryWriter torch.backends.cuda.matmul.allow_tf32 = True # for gpu >= Ampere and pytorch >= 1.12 from dust3r.model import AsymmetricCroCo3DStereo, inf # noqa: F401, needed when loading the model from dust3r.datasets import get_data_loader # noqa from dust3r.losses import * # noqa: F401, needed when loading the model from dust3r.inference import loss_of_one_batch, visualize_results, visualize_results_mmask # noqa from dust3r.pose_eval import eval_pose_estimation from dust3r.depth_eval import eval_mono_depth_estimation # from demo import get_3D_model_from_scene import dust3r.utils.path_to_croco # noqa: F401 import croco.utils.misc as misc # noqa from croco.utils.misc import NativeScalerWithGradNormCount as NativeScaler # noqa import PIL.Image as Image from dust3r.cloud_opt.motion_mask_from_raft import get_motion_mask_from_pairs def get_args_parser(): parser = argparse.ArgumentParser('DUST3R training', add_help=False) # model and criterion parser.add_argument('--model', default="AsymmetricCroCo3DStereo(pos_embed='RoPE100', patch_embed_cls='ManyAR_PatchEmbed', \ img_size=(512, 512), head_type='dpt', output_mode='pts3d', depth_mode=('exp', -inf, inf), conf_mode=('exp', 1, inf), \ enc_embed_dim=1024, enc_depth=24, enc_num_heads=16, dec_embed_dim=768, dec_depth=12, dec_num_heads=12, freeze='encoder')", type=str, help="string containing the model to build") parser.add_argument('--pretrained', default=None, help='path of a starting checkpoint') parser.add_argument('--train_criterion', default="ConfLoss(Regr3D(L21, norm_mode='avg_dis'), alpha=0.2)", type=str, help="train criterion") parser.add_argument('--test_criterion', default=None, type=str, help="test criterion") # dataset parser.add_argument('--train_dataset', default='[None]', type=str, help="training set") parser.add_argument('--test_dataset', default='[None]', type=str, help="testing set") # training parser.add_argument('--seed', default=0, type=int, help="Random seed") parser.add_argument('--batch_size', default=64, type=int, help="Batch size per GPU (effective batch size is batch_size * accum_iter * # gpus") parser.add_argument('--test_batch_size', default=64, type=int, help="Batch size per GPU (effective batch size is batch_size * accum_iter * # gpus") parser.add_argument('--accum_iter', default=1, type=int, help="Accumulate gradient iterations (for increasing the effective batch size under memory constraints)") parser.add_argument('--epochs', default=800, type=int, help="Maximum number of epochs for the scheduler") parser.add_argument('--weight_decay', type=float, default=0.05, help="weight decay (default: 0.05)") parser.add_argument('--lr', type=float, default=None, metavar='LR', help='learning rate (absolute lr)') parser.add_argument('--blr', type=float, default=1.5e-4, metavar='LR', help='base learning rate: absolute_lr = base_lr * total_batch_size / 256') parser.add_argument('--min_lr', type=float, default=0., metavar='LR', help='lower lr bound for cyclic schedulers that hit 0') parser.add_argument('--warmup_epochs', type=int, default=40, metavar='N', help='epochs to warmup LR') parser.add_argument('--amp', type=int, default=0, choices=[0, 1], help="Use Automatic Mixed Precision for pretraining") parser.add_argument("--cudnn_benchmark", action='store_true', default=False, help="set cudnn.benchmark = False") parser.add_argument("--eval_only", action='store_true', default=False) parser.add_argument("--fixed_eval_set", action='store_true', default=False) parser.add_argument('--resume', default=None, type=str, help='path to latest checkpoint (default: none)') # others parser.add_argument('--num_workers', default=8, type=int) parser.add_argument('--world_size', default=1, type=int, help='number of distributed processes') parser.add_argument('--local_rank', default=-1, type=int) parser.add_argument('--dist_url', default='env://', help='url used to set up distributed training') parser.add_argument('--eval_freq', type=int, default=5, help='Test loss evaluation frequency') parser.add_argument('--save_freq', default=1, type=int, help='frequence (number of epochs) to save checkpoint in checkpoint-last.pth') parser.add_argument('--keep_freq', default=5, type=int, help='frequence (number of epochs) to save checkpoint in checkpoint-%d.pth') parser.add_argument('--print_freq', default=20, type=int, help='frequence (number of iterations) to print infos while training') parser.add_argument('--wandb', action='store_true', default=False, help='use wandb for logging') parser.add_argument('--num_save_visual', default=4, type=int, help='number of visualizations to save') # switch mode for train / eval pose / eval depth parser.add_argument('--mode', default='train', type=str, help='train / eval_pose / eval_depth') # for pose eval parser.add_argument('--threshold', default=0.5, type=float, help='threshold for motion mask') parser.add_argument('--pose_eval_freq', default=25, type=int, help='pose evaluation frequency') parser.add_argument('--pose_eval_stride', default=1, type=int, help='stride for pose evaluation') parser.add_argument('--scene_graph_type', default='swinstride-5-noncyclic', type=str, help='scene graph window size') parser.add_argument('--save_best_pose', action='store_true', default=False, help='save best pose') parser.add_argument('--n_iter', default=300, type=int, help='number of iterations for pose optimization') parser.add_argument('--save_pose_qualitative', action='store_true', default=False, help='save qualitative pose results') parser.add_argument('--temporal_smoothing_weight', default=0.01, type=float, help='temporal smoothing weight for pose optimization') parser.add_argument('--not_shared_focal', action='store_true', default=False, help='use shared focal length for pose optimization') parser.add_argument('--use_gt_focal', action='store_true', default=False, help='use ground truth focal length for pose optimization') parser.add_argument('--pose_schedule', default='linear', type=str, help='pose optimization schedule') parser.add_argument('--flow_loss_weight', default=0.01, type=float, help='flow loss weight for pose optimization') parser.add_argument('--cananical_space_loss_weight', default=1, type=float, help='cananical_space_loss_weight for pose optimization') parser.add_argument('--flow_loss_fn', default='smooth_l1', type=str, help='flow loss type for pose optimization') parser.add_argument('--use_gt_mask', action='store_true', default=False, help='use gt mask for pose optimization, for sintel/davis') parser.add_argument('--use_pred_mask', action='store_true', default=False, help='use nn predicted mask for pose optimization') parser.add_argument('--evaluate_davis', action='store_true', default=False, help='evaluate davis on first 50 frames') parser.add_argument('--not_batchify', action='store_true', default=False, help='Use non batchify mode for global optimization') parser.add_argument('--dir_path', type=str, help='path to custom dataset for pose evaluation') parser.add_argument('--motion_mask_thre', default=0.35, type=float, help='motion mask threshold for pose optimization') parser.add_argument('--sam2_mask_refine', action='store_true', default=False, help='use sam2 mask refine for the motion for pose optimization') parser.add_argument('--flow_loss_start_epoch', default=0.1, type=float, help='start epoch for flow loss') parser.add_argument('--flow_loss_thre', default=20, type=float, help='threshold for flow loss') parser.add_argument('--pxl_thresh', default=50.0, type=float, help='threshold for flow loss') parser.add_argument('--depth_regularize_weight', default=0.0, type=float, help='depth regularization weight for pose optimization') parser.add_argument('--translation_weight', default=1, type=float, help='translation weight for pose optimization') parser.add_argument('--silent', action='store_true', default=False, help='silent mode for pose evaluation') parser.add_argument('--full_seq', action='store_true', default=False, help='use full sequence for pose evaluation') parser.add_argument('--seq_list', nargs='+', default=None, help='list of sequences for pose evaluation') parser.add_argument('--eval_dataset', type=str, default='sintel', choices=['davis', 'kitti', 'bonn', 'scannet', 'tum', 'nyu', 'sintel'], help='choose dataset for pose evaluation') # for monocular depth eval parser.add_argument('--no_crop', action='store_true', default=False, help='do not crop the image for monocular depth evaluation') # output dir parser.add_argument('--output_dir', default='./results/tmp', type=str, help="path where to save the output") return parser def load_model(args, device): # model print('Loading model: {:s}'.format(args.model)) model = eval(args.model) if args.pretrained and not args.resume: if os.path.isfile(args.pretrained): # load from pth file print('Loading pretrained: ', args.pretrained) ckpt = torch.load(args.pretrained, map_location=device, weights_only=False) print(model.load_state_dict(ckpt['model'], strict=False)) del ckpt # in case it occupies memory else: # load from huggingface print('Loading pretrained from huggingface: ', args.pretrained) model = model.from_pretrained(args.pretrained) model.to(device) model_without_ddp = model if args.distributed: model = torch.nn.parallel.DistributedDataParallel( model, device_ids=[args.gpu], find_unused_parameters=True, static_graph=True) model_without_ddp = model.module return model, model_without_ddp def train(args): misc.init_distributed_mode(args) global_rank = misc.get_rank() world_size = misc.get_world_size() # if main process, init wandb if args.wandb and misc.is_main_process(): wandb.init(name=args.output_dir.split('/')[-1], project='dust3r', config=args, sync_tensorboard=True, dir=args.output_dir) print("output_dir: " + args.output_dir) if args.output_dir: Path(args.output_dir).mkdir(parents=True, exist_ok=True) # auto resume if not specified if args.resume is None: last_ckpt_fname = os.path.join(args.output_dir, f'checkpoint-last.pth') if os.path.isfile(last_ckpt_fname) and (not args.eval_only): args.resume = last_ckpt_fname print('job dir: {}'.format(os.path.dirname(os.path.realpath(__file__)))) print("{}".format(args).replace(', ', ',\n')) device = "cuda" if torch.cuda.is_available() else "cpu" device = torch.device(device) # fix the seed seed = args.seed + misc.get_rank() torch.manual_seed(seed) np.random.seed(seed) cudnn.benchmark = args.cudnn_benchmark model, model_without_ddp = load_model(args, device) if not args.eval_only: # training dataset and loader print('Building train dataset {:s}'.format(args.train_dataset)) # dataset and loader data_loader_train = build_dataset(args.train_dataset, args.batch_size, args.num_workers, test=False) print(f'>> Creating train criterion = {args.train_criterion}') train_criterion = eval(args.train_criterion).to(device) print('Building test dataset {:s}'.format(args.train_dataset)) data_loader_test = {} for dataset in args.test_dataset.split('+'): testset = build_dataset(dataset, args.test_batch_size, args.num_workers, test=True) name_testset = dataset.split('(')[0] if getattr(testset.dataset.dataset, 'strides', None) is not None: name_testset += f'_stride{testset.dataset.dataset.strides}' data_loader_test[name_testset] = testset print(f'>> Creating test criterion = {args.test_criterion or args.train_criterion}') test_criterion = eval(args.test_criterion or args.criterion).to(device) eff_batch_size = args.batch_size * args.accum_iter * misc.get_world_size() if args.lr is None: # only base_lr is specified args.lr = args.blr * eff_batch_size / 256 print("base lr: %.2e" % (args.lr * 256 / eff_batch_size)) print("actual lr: %.2e" % args.lr) print("accumulate grad iterations: %d" % args.accum_iter) print("effective batch size: %d" % eff_batch_size) # following timm: set wd as 0 for bias and norm layers param_groups = misc.get_parameter_groups(model_without_ddp, args.weight_decay) optimizer = torch.optim.AdamW(param_groups, lr=args.lr, betas=(0.9, 0.95)) # print(optimizer) loss_scaler = NativeScaler() def write_log_stats(epoch, train_stats, test_stats): if misc.is_main_process(): if log_writer is not None: log_writer.flush() gathered_test_stats = {} log_stats = dict(epoch=epoch, **{f'train_{k}': v for k, v in train_stats.items()}) for test_name, testset in data_loader_test.items(): if test_name not in test_stats: continue if getattr(testset.dataset.dataset, 'strides', None) is not None: original_test_name = test_name.split('_stride')[0] if original_test_name not in gathered_test_stats.keys(): gathered_test_stats[original_test_name] = [] gathered_test_stats[original_test_name].append(test_stats[test_name]) log_stats.update({test_name + '_' + k: v for k, v in test_stats[test_name].items()}) if len(gathered_test_stats) > 0: for original_test_name, stride_stats in gathered_test_stats.items(): if len(stride_stats) > 1: stride_stats = {k: np.mean([x[k] for x in stride_stats]) for k in stride_stats[0]} log_stats.update({original_test_name + '_stride_mean_' + k: v for k, v in stride_stats.items()}) if args.wandb: log_dict = {original_test_name + '_stride_mean_' + k: v for k, v in stride_stats.items()} log_dict.update({'epoch': epoch}) wandb.log(log_dict) with open(os.path.join(args.output_dir, "log.txt"), mode="a", encoding="utf-8") as f: f.write(json.dumps(log_stats) + "\n") def save_model(epoch, fname, best_so_far, best_pose_ate_sofar=None): misc.save_model(args=args, model_without_ddp=model_without_ddp, optimizer=optimizer, loss_scaler=loss_scaler, epoch=epoch, fname=fname, best_so_far=best_so_far, best_pose_ate_sofar=best_pose_ate_sofar) best_so_far, best_pose_ate_sofar = misc.load_model(args=args, model_without_ddp=model_without_ddp, optimizer=optimizer, loss_scaler=loss_scaler) if best_so_far is None: best_so_far = float('inf') if best_pose_ate_sofar is None: best_pose_ate_sofar = float('inf') if global_rank == 0 and args.output_dir is not None: log_writer = SummaryWriter(log_dir=args.output_dir) else: log_writer = None print(f"Start training for {args.epochs} epochs") start_time = time.time() train_stats = test_stats = {} for epoch in range(args.start_epoch, args.epochs + 1): # Test on multiple datasets new_best = False new_pose_best = False already_saved = False if (epoch > args.start_epoch and args.eval_freq > 0 and epoch % args.eval_freq == 0) or args.eval_only: test_stats = {} for test_name, testset in data_loader_test.items(): print(f'Testing on {test_name}...') stats = test_one_epoch(model, test_criterion, testset, device, epoch, log_writer=log_writer, args=args, prefix=test_name) test_stats[test_name] = stats # Save best of all if stats['loss_med'] < best_so_far: best_so_far = stats['loss_med'] new_best = True # Ensure that eval_pose_estimation is only run on the main process if args.pose_eval_freq>0 and (epoch % args.pose_eval_freq==0 or args.eval_only): ate_mean, rpe_trans_mean, rpe_rot_mean, outfile_list, bug = eval_pose_estimation(args, model, device, save_dir=f'{args.output_dir}/{epoch}') print(f'ATE mean: {ate_mean}, RPE trans mean: {rpe_trans_mean}, RPE rot mean: {rpe_rot_mean}') # Optionally log the results to wandb if args.wandb and misc.is_main_process(): wandb_dict = { 'epoch': epoch, 'ATE mean': ate_mean, 'RPE trans mean': rpe_trans_mean, 'RPE rot mean': rpe_rot_mean, } if args.save_pose_qualitative: for outfile in outfile_list: wandb_dict[outfile.split('/')[-1]] = wandb.Object3D(open(outfile)) wandb.log(wandb_dict) if ate_mean < best_pose_ate_sofar and not bug: # if the pose estimation is better, and w/o any error best_pose_ate_sofar = ate_mean new_pose_best = True # Synchronize all processes to ensure eval_pose_estimation is completed try: torch.distributed.barrier() except: pass # Save more stuff write_log_stats(epoch, train_stats, test_stats) if args.eval_only: exit(0) if epoch > args.start_epoch: if args.keep_freq and epoch % args.keep_freq == 0: save_model(epoch - 1, str(epoch), best_so_far, best_pose_ate_sofar) already_saved = True if new_best: save_model(epoch - 1, 'best', best_so_far, best_pose_ate_sofar) already_saved = True if new_pose_best and args.save_best_pose: save_model(epoch - 1, 'best_pose', best_so_far, best_pose_ate_sofar) already_saved = True # Save immediately the last checkpoint if epoch > args.start_epoch: save_model(epoch - 1, 'last', best_so_far, best_pose_ate_sofar) if epoch >= args.epochs: break # exit after writing last test to disk # Train train_stats = train_one_epoch( model, train_criterion, data_loader_train, optimizer, device, epoch, loss_scaler, log_writer=log_writer, args=args) total_time = time.time() - start_time total_time_str = str(datetime.timedelta(seconds=int(total_time))) print('Training time {}'.format(total_time_str)) save_final_model(args, args.epochs, model_without_ddp, best_so_far=best_so_far) def save_final_model(args, epoch, model_without_ddp, best_so_far=None): output_dir = Path(args.output_dir) checkpoint_path = output_dir / 'checkpoint-final.pth' to_save = { 'args': args, 'model': model_without_ddp if isinstance(model_without_ddp, dict) else model_without_ddp.cpu().state_dict(), 'epoch': epoch } if best_so_far is not None: to_save['best_so_far'] = best_so_far print(f'>> Saving model to {checkpoint_path} ...') misc.save_on_master(to_save, checkpoint_path) def build_dataset(dataset, batch_size, num_workers, test=False): split = ['Train', 'Test'][test] print(f'Building {split} Data loader for dataset: ', dataset) loader = get_data_loader(dataset, batch_size=batch_size, num_workers=num_workers, pin_mem=True, shuffle=not (test), drop_last=not (test)) print(f"{split} dataset length: ", len(loader)) return loader def train_one_epoch(model: torch.nn.Module, criterion: torch.nn.Module, data_loader: Sized, optimizer: torch.optim.Optimizer, device: torch.device, epoch: int, loss_scaler, args, log_writer=None): assert torch.backends.cuda.matmul.allow_tf32 == True model.train(True) metric_logger = misc.MetricLogger(delimiter=" ") metric_logger.add_meter('lr', misc.SmoothedValue(window_size=1, fmt='{value:.6f}')) header = 'Epoch: [{}]'.format(epoch) accum_iter = args.accum_iter if log_writer is not None: print('log_dir: {}'.format(log_writer.log_dir)) if hasattr(data_loader, 'dataset') and hasattr(data_loader.dataset, 'set_epoch'): data_loader.dataset.set_epoch(epoch) if hasattr(data_loader, 'sampler') and hasattr(data_loader.sampler, 'set_epoch'): data_loader.sampler.set_epoch(epoch) optimizer.zero_grad() for data_iter_step, batch in enumerate(metric_logger.log_every(data_loader, args.print_freq, header)): epoch_f = epoch + data_iter_step / len(data_loader) # we use a per iteration (instead of per epoch) lr scheduler if data_iter_step % accum_iter == 0: misc.adjust_learning_rate(optimizer, epoch_f, args) batch_result = loss_of_one_batch(batch, model, criterion, device, symmetrize_batch=True, use_amp=bool(args.amp)) loss, loss_details = batch_result['loss'] # criterion returns two values loss_value = float(loss) if (data_iter_step % max((len(data_loader) // args.num_save_visual), 1) == 0 or data_iter_step == 0) and misc.is_main_process() : print(f'Saving visualizations for data_iter_step {data_iter_step}...') save_dir = f'{args.output_dir}/{epoch}' Path(save_dir).mkdir(parents=True, exist_ok=True) view1, view2, pred1, pred2 = batch_result['view1'], batch_result['view2'], batch_result['pred1'], batch_result['pred2'] gt_rgb_mmask1, gt_rgb_mmask2 = visualize_results_mmask(view1, view2, pred1, pred2, save_dir=save_dir, visualize_type='gt') pred_rgb_mmask1, pred_rgb_mmask2 = visualize_results_mmask(view1, view2, pred1, pred2, save_dir=save_dir, visualize_type='pred') if args.wandb: wandb.log({ 'epoch': epoch, 'train_gt_mmask_1': wandb.Image(gt_rgb_mmask1), 'train_gt_mmask_2': wandb.Image(gt_rgb_mmask2), 'train_pred_mmask_1': wandb.Image(pred_rgb_mmask1), 'train_pred_mmask_2': wandb.Image(pred_rgb_mmask2) }) if not math.isfinite(loss_value): print("Loss is {}, stopping training".format(loss_value), force=True) sys.exit(1) loss /= accum_iter loss_scaler(loss, optimizer, parameters=model.parameters(), update_grad=(data_iter_step + 1) % accum_iter == 0) if (data_iter_step + 1) % accum_iter == 0: optimizer.zero_grad() del loss del batch lr = optimizer.param_groups[0]["lr"] metric_logger.update(epoch=epoch_f) metric_logger.update(lr=lr) metric_logger.update(loss=loss_value, **loss_details) if (data_iter_step + 1) % accum_iter == 0 and ((data_iter_step + 1) % (accum_iter * args.print_freq)) == 0: loss_value_reduce = misc.all_reduce_mean(loss_value) # MUST BE EXECUTED BY ALL NODES if log_writer is None: continue """ We use epoch_1000x as the x-axis in tensorboard. This calibrates different curves when batch size changes. """ epoch_1000x = int(epoch_f * 1000) log_writer.add_scalar('train_loss', loss_value_reduce, epoch_1000x) log_writer.add_scalar('train_lr', lr, epoch_1000x) log_writer.add_scalar('train_iter', epoch_1000x, epoch_1000x) for name, val in loss_details.items(): log_writer.add_scalar('train_' + name, val, epoch_1000x) # gather the stats from all processes metric_logger.synchronize_between_processes() print("Averaged stats:", metric_logger) return {k: meter.global_avg for k, meter in metric_logger.meters.items()} @torch.no_grad() def test_one_epoch(model: torch.nn.Module, criterion: torch.nn.Module, data_loader: Sized, device: torch.device, epoch: int, args, log_writer=None, prefix='test'): model.eval() metric_logger = misc.MetricLogger(delimiter=" ") metric_logger.meters = defaultdict(lambda: misc.SmoothedValue(window_size=9**9)) header = 'Test Epoch: [{}]'.format(epoch) if log_writer is not None: print('log_dir: {}'.format(log_writer.log_dir)) if hasattr(data_loader, 'dataset') and hasattr(data_loader.dataset, 'set_epoch'): data_loader.dataset.set_epoch(epoch) if not args.fixed_eval_set else data_loader.dataset.set_epoch(0) if hasattr(data_loader, 'sampler') and hasattr(data_loader.sampler, 'set_epoch'): data_loader.sampler.set_epoch(epoch) if not args.fixed_eval_set else data_loader.sampler.set_epoch(0) for idx, batch in enumerate(metric_logger.log_every(data_loader, args.print_freq, header)): batch_result = loss_of_one_batch(batch, model, criterion, device, symmetrize_batch=True, use_amp=bool(args.amp)) loss_tuple = batch_result['loss'] loss_value, loss_details = loss_tuple # criterion returns two values metric_logger.update(loss=float(loss_value), **loss_details) if args.num_save_visual>0 and (idx % max((len(data_loader) // args.num_save_visual), 1) == 0) and misc.is_main_process() : # save visualizations save_dir = f'{args.output_dir}/{epoch}' Path(save_dir).mkdir(parents=True, exist_ok=True) view1, view2, pred1, pred2 = batch_result['view1'], batch_result['view2'], batch_result['pred1'], batch_result['pred2'] gt_rgb_mmask1, gt_rgb_mmask2 = visualize_results_mmask(view1, view2, pred1, pred2, save_dir=save_dir, visualize_type='gt') pred_rgb_mmask1, pred_rgb_mmask2 = visualize_results_mmask(view1, view2, pred1, pred2, save_dir=save_dir, visualize_type='pred') if args.wandb: wandb.log({ 'epoch': epoch, f'{prefix}_test_gt_mmask_1': wandb.Image(gt_rgb_mmask1), f'{prefix}_test_gt_mmask_2': wandb.Image(gt_rgb_mmask2), f'{prefix}_test_pred_mmask_1': wandb.Image(pred_rgb_mmask1), f'{prefix}_test_pred_mmask_2': wandb.Image(pred_rgb_mmask2) }) # gather the stats from all processes metric_logger.synchronize_between_processes() print("Averaged stats:", metric_logger) aggs = [('avg', 'global_avg'), ('med', 'median')] results = {f'{k}_{tag}': getattr(meter, attr) for k, meter in metric_logger.meters.items() for tag, attr in aggs} if log_writer is not None: for name, val in results.items(): log_writer.add_scalar(prefix + '_' + name, val, 1000 * epoch) return results