diff --git a/model_examples/CenterPoint/OpenPCDet_npu.patch b/model_examples/CenterPoint/OpenPCDet_npu.patch index 8ee258d8ef63215185962d1032e7757de2eee4b1..e5c25f4d5d6ab2bf38fd897cb868148a413c5bf9 100644 --- a/model_examples/CenterPoint/OpenPCDet_npu.patch +++ b/model_examples/CenterPoint/OpenPCDet_npu.patch @@ -123,7 +123,7 @@ index 9fdf7d2..e69de29 100644 -if script_version not in __version__: - __version__ = __version__ + '+py%s' % script_version diff --git a/pcdet/datasets/__init__.py b/pcdet/datasets/__init__.py -index 47c3900..226223d 100644 +index 47c3900..7e7b835 100644 --- a/pcdet/datasets/__init__.py +++ b/pcdet/datasets/__init__.py @@ -15,6 +15,8 @@ from .once.once_dataset import ONCEDataset @@ -135,7 +135,7 @@ index 47c3900..226223d 100644 __all__ = { 'DatasetTemplate': DatasetTemplate, 'KittiDataset': KittiDataset, -@@ -52,15 +54,25 @@ class DistributedSampler(_DistributedSampler): +@@ -52,15 +54,26 @@ class DistributedSampler(_DistributedSampler): def build_dataloader(dataset_cfg, class_names, batch_size, dist, root_path=None, workers=4, seed=None, @@ -148,7 +148,7 @@ index 47c3900..226223d 100644 - training=training, - logger=logger, - ) -+ logger=None, training=True, merge_all_iters_to_one_epoch=False, total_epochs=0, use_dynamic_dataloader=False): ++ logger=None, training=True, merge_all_iters_to_one_epoch=False, total_epochs=0, num_replicas=8, use_dynamic_dataloader=False): + + if dataset_cfg.DATASET == 'NuScenesDataset': + dataset = __all__[dataset_cfg.DATASET]( @@ -157,6 +157,7 @@ index 47c3900..226223d 100644 + root_path=root_path, + training=training, + logger=logger, ++ num_replicas=num_replicas, + use_dynamic_dataloader=use_dynamic_dataloader + ) + else: @@ -170,7 +171,7 @@ index 47c3900..226223d 100644 if merge_all_iters_to_one_epoch: assert hasattr(dataset, 'merge_all_iters_to_one_epoch') -@@ -68,14 +80,17 @@ def build_dataloader(dataset_cfg, class_names, batch_size, dist, root_path=None, +@@ -68,14 +81,17 @@ def build_dataloader(dataset_cfg, class_names, batch_size, dist, root_path=None, if dist: if training: @@ -706,16 +707,24 @@ index 543d8f2..3cce4c2 100644 + return dev_iou \ No newline at end of file diff --git a/pcdet/datasets/nuscenes/nuscenes_dataset.py b/pcdet/datasets/nuscenes/nuscenes_dataset.py -index 0f70005..084ec8f 100644 +index 0f70005..a4b1811 100644 --- a/pcdet/datasets/nuscenes/nuscenes_dataset.py +++ b/pcdet/datasets/nuscenes/nuscenes_dataset.py -@@ -1,3 +1,5 @@ +@@ -1,8 +1,13 @@ +# Copyright 2024 Huawei Technologies Co., Ltd + import copy import pickle from pathlib import Path -@@ -10,13 +12,13 @@ from ...utils import common_utils + + import numpy as np ++import torch ++import torch_npu ++from torch_npu.contrib import transfer_to_npu + from tqdm import tqdm + + from ...ops.roiaware_pool3d import roiaware_pool3d_utils +@@ -10,13 +15,13 @@ from ...utils import common_utils from ..dataset import DatasetTemplate from pyquaternion import Quaternion from PIL import Image @@ -725,7 +734,7 @@ index 0f70005..084ec8f 100644 -class NuScenesDataset(DatasetTemplate): - def __init__(self, dataset_cfg, class_names, training=True, root_path=None, logger=None): +class NuScenesDataset(PointCloudDynamicDataset, DatasetTemplate): -+ def __init__(self, dataset_cfg, class_names, training=True, root_path=None, logger=None, use_dynamic_dataloader=False): ++ def __init__(self, dataset_cfg, class_names, training=True, root_path=None, logger=None, num_replicas=8, use_dynamic_dataloader=False): root_path = (root_path if root_path is not None else Path(dataset_cfg.DATA_PATH)) / dataset_cfg.VERSION - super().__init__( - dataset_cfg=dataset_cfg, class_names=class_names, training=training, root_path=root_path, logger=logger @@ -734,19 +743,19 @@ index 0f70005..084ec8f 100644 ) self.infos = [] self.camera_config = self.dataset_cfg.get('CAMERA_CONFIG', None) -@@ -29,6 +31,11 @@ class NuScenesDataset(DatasetTemplate): +@@ -29,6 +34,11 @@ class NuScenesDataset(DatasetTemplate): self.include_nuscenes_data(self.mode) if self.training and self.dataset_cfg.get('BALANCED_RESAMPLING', False): self.infos = self.balanced_infos_resampling(self.infos) + if training and use_dynamic_dataloader: + self.sorting() -+ self.buckets = self.bucketing(bucket_capacity=8) ++ self.buckets = self.bucketing(bucket_capacity=num_replicas) + else: + self.buckets = None def include_nuscenes_data(self, mode): self.logger.info('Loading NuScenes dataset') -@@ -141,17 +148,17 @@ class NuScenesDataset(DatasetTemplate): +@@ -141,17 +151,17 @@ class NuScenesDataset(DatasetTemplate): crop_h = newH - fH crop_w = int(max(0, newW - fW) / 2) crop = (crop_w, crop_h, crop_w + fW, crop_h + fH) @@ -767,7 +776,7 @@ index 0f70005..084ec8f 100644 def load_camera_info(self, input_dict, info): input_dict["image_paths"] = [] input_dict["lidar2camera"] = [] -@@ -200,10 +207,10 @@ class NuScenesDataset(DatasetTemplate): +@@ -200,10 +210,10 @@ class NuScenesDataset(DatasetTemplate): images = [] for name in filename: images.append(Image.open(str(self.root_path / name))) @@ -2417,19 +2426,21 @@ index 51b7178..cd8652d 100644 help='set extra config keys if needed') diff --git a/tools/train.py b/tools/train.py -index 29a88bd..a0d5fc3 100644 +index 29a88bd..c3fd102 100644 --- a/tools/train.py +++ b/tools/train.py -@@ -7,6 +7,8 @@ from pathlib import Path +@@ -7,7 +7,10 @@ from pathlib import Path from test import repeat_eval_ckpt import torch +import torch_npu +from torch_npu.contrib import transfer_to_npu import torch.nn as nn ++import torch.distributed as dist from tensorboardX import SummaryWriter -@@ -33,7 +35,7 @@ def parse_config(): + from pcdet.config import cfg, cfg_from_list, cfg_from_yaml_file, log_config_to_file +@@ -33,7 +36,7 @@ def parse_config(): parser.add_argument('--sync_bn', action='store_true', default=False, help='whether to use sync bn') parser.add_argument('--fix_random_seed', action='store_true', default=False, help='') parser.add_argument('--ckpt_save_interval', type=int, default=1, help='number of training epochs') @@ -2438,7 +2449,7 @@ index 29a88bd..a0d5fc3 100644 parser.add_argument('--max_ckpt_save_num', type=int, default=30, help='max number of saved checkpoint') parser.add_argument('--merge_all_iters_to_one_epoch', action='store_true', default=False, help='') parser.add_argument('--set', dest='set_cfgs', default=None, nargs=argparse.REMAINDER, -@@ -43,20 +45,20 @@ def parse_config(): +@@ -43,20 +46,20 @@ def parse_config(): parser.add_argument('--start_epoch', type=int, default=0, help='') parser.add_argument('--num_epochs_to_eval', type=int, default=0, help='number of checkpoints to be evaluated') parser.add_argument('--save_to_file', action='store_true', default=False, help='') @@ -2462,7 +2473,7 @@ index 29a88bd..a0d5fc3 100644 args.use_amp = args.use_amp or cfg.OPTIMIZATION.get('USE_AMP', False) if args.set_cfgs is not None: -@@ -72,7 +74,7 @@ def main(): +@@ -72,7 +75,7 @@ def main(): total_gpus = 1 else: total_gpus, cfg.LOCAL_RANK = getattr(common_utils, 'init_dist_%s' % args.launcher)( @@ -2471,7 +2482,7 @@ index 29a88bd..a0d5fc3 100644 ) dist_train = True -@@ -104,11 +106,11 @@ def main(): +@@ -104,11 +107,11 @@ def main(): logger.info('Training in distributed mode : total_batch_size: %d' % (total_gpus * args.batch_size)) else: logger.info('Training with a single process') @@ -2485,12 +2496,13 @@ index 29a88bd..a0d5fc3 100644 os.system('cp %s %s' % (args.cfg_file, output_dir)) tb_log = SummaryWriter(log_dir=str(output_dir / 'tensorboard')) if cfg.LOCAL_RANK == 0 else None -@@ -123,13 +125,14 @@ def main(): +@@ -123,13 +126,15 @@ def main(): training=True, merge_all_iters_to_one_epoch=args.merge_all_iters_to_one_epoch, total_epochs=args.epochs, - seed=666 if args.fix_random_seed else None + seed=666 if args.fix_random_seed else None, ++ num_replicas=dist.get_world_size(), + use_dynamic_dataloader=True if args.use_dynamic_dataloader else False ) @@ -2502,7 +2514,7 @@ index 29a88bd..a0d5fc3 100644 optimizer = build_optimizer(model, cfg.OPTIMIZATION) -@@ -144,7 +147,7 @@ def main(): +@@ -144,7 +149,7 @@ def main(): last_epoch = start_epoch + 1 else: ckpt_list = glob.glob(str(ckpt_dir / '*.pth')) @@ -2511,7 +2523,7 @@ index 29a88bd..a0d5fc3 100644 if len(ckpt_list) > 0: ckpt_list.sort(key=os.path.getmtime) while len(ckpt_list) > 0: -@@ -189,11 +192,11 @@ def main(): +@@ -189,11 +194,11 @@ def main(): lr_warmup_scheduler=lr_warmup_scheduler, ckpt_save_interval=args.ckpt_save_interval, max_ckpt_save_num=args.max_ckpt_save_num, diff --git a/model_examples/CenterPoint/README.md b/model_examples/CenterPoint/README.md index bd2b2df18b97f9f009dfe598a6c2e81e097c5582..5f3372b64c68b91fc6838ece016ba5b8edb84a09 100644 --- a/model_examples/CenterPoint/README.md +++ b/model_examples/CenterPoint/README.md @@ -441,4 +441,5 @@ pip install -r 2.5_requirements.txt [2025-03-12] **NEW:** CenterPoint仓CenterHead模块性能优化，并提供高性能内存库安装指导，更新了CenterPoint3d的性能数据 [2025-04-21] **NEW:** CenterPoint2d模型优化fps计算方式，乘以卡数，更新性能指标，readme中添加global batch size [2025-05-29] **NEW:** CenterPoint3d模型优化，更新性能 -[2025-08-21] **NEW:** CenterPoint3d模型性能优化，更新性能 \ No newline at end of file +[2025-08-21] **NEW:** CenterPoint3d模型性能优化，更新性能 +[2025-08-29] **NEW:** CenterPoint3d修复数据处理问题 \ No newline at end of file diff --git a/mx_driving/dataset/agent_dataset.py b/mx_driving/dataset/agent_dataset.py index f0da5d1c5f5cf6c6c2f768ca199fdd7c7a519ab5..fa93f1c8871efe09fb77fea059c098d598e1a3ac 100644 --- a/mx_driving/dataset/agent_dataset.py +++ b/mx_driving/dataset/agent_dataset.py @@ -31,10 +31,6 @@ from torch.utils.data import DataLoader, Sampler, BatchSampler, DistributedSampl from torch.utils.data.dataloader import default_collate from torch.utils.data.distributed import DistributedSampler -from torch_geometric.data import Batch, Dataset -from torch_geometric.data.data import BaseData -from torch_geometric.data.datapipes import DatasetAdapter - from ..dataset.utils.dynamic_dataset import UniformBucketingDynamicDataset from ..dataset.utils.dynamic_sampler import DynamicDistributedSampler @@ -45,6 +41,8 @@ class Collater: self.exclude_keys = exclude_keys def __call__(self, batch): + from torch_geometric.data import Batch + from torch_geometric.data.data import BaseData elem = batch[0] if isinstance(elem, BaseData): return Batch.from_data_list(batch, self.follow_batch, @@ -122,7 +120,7 @@ class AgentDynamicDataset(UniformBucketingDynamicDataset): class DynamicBatchSampler(BatchSampler): def __init__(self, - dataset: Dataset, + dataset, sampler, batch_size: int, drop_last: bool = False) -> None: @@ -144,7 +142,7 @@ class DynamicBatchSampler(BatchSampler): class AgentDynamicBatchSampler(DynamicDistributedSampler): def __init__(self, - dataset: Dataset, + dataset, num_replicas: Optional[int] = None, rank: Optional[int] = None, shuffle: bool = True, @@ -185,7 +183,7 @@ class AgentDynamicBatchSampler(DynamicDistributedSampler): class AgentDynamicBatchDataLoader(DataLoader): def __init__(self, - dataset: Union[Dataset, Sequence[BaseData], DatasetAdapter], + dataset, batch_size: int, train_batch_size: int, shuffle: bool = True,