diff --git a/README.md b/README.md index d0cf13acc527c1a46d3bb7a03c76b8b9fdcc7a49..147a889159012f3638b8685bf761eb379a7c2450 100644 --- a/README.md +++ b/README.md @@ -150,6 +150,7 @@ out, argmax = scatter_max(updates, indices, out) | BEVNeXt | https://gitee.com/ascend/DrivingSDK/tree/master/model_examples/BEVNeXt |N| | MultiPath++ | https://gitee.com/ascend/DrivingSDK/tree/master/model_examples/MultiPath++ |Y| | SalsaNext | https://gitee.com/ascend/DrivingSDK/tree/master/model_examples/SalsaNext |N| +| Panoptic-PolarNet | https://gitee.com/ascend/DrivingSDK/tree/master/model_examples/Panoptic-PolarNet |N| # 支持的产品型号 - Atlas A2 训练系列产品 diff --git a/model_examples/Panoptic-PolarNet/Panoptic-PolarNet.patch b/model_examples/Panoptic-PolarNet/Panoptic-PolarNet.patch new file mode 100644 index 0000000000000000000000000000000000000000..8ece206cde8a35a1e1fe0afd05407ab58eeccfa2 --- /dev/null +++ b/model_examples/Panoptic-PolarNet/Panoptic-PolarNet.patch @@ -0,0 +1,456 @@ +diff --git a/configs/SemanticKITTI_model/Panoptic-PolarNet.yaml b/configs/SemanticKITTI_model/Panoptic-PolarNet.yaml +index 0f3e79d..f4fac33 100644 +--- a/configs/SemanticKITTI_model/Panoptic-PolarNet.yaml ++++ b/configs/SemanticKITTI_model/Panoptic-PolarNet.yaml +@@ -35,7 +35,7 @@ model: + offset_loss: L1 + center_loss_weight: 100 + offset_loss_weight: 10 +- enable_SAP: True ++ enable_SAP: False + SAP: + start_epoch: 30 +- rate: 0.01 +\ No newline at end of file ++ rate: 0.01 +diff --git a/dataloader/dataset.py b/dataloader/dataset.py +index a92921b..28a0d8e 100644 +--- a/dataloader/dataset.py ++++ b/dataloader/dataset.py +@@ -17,6 +17,8 @@ from torch.utils import data + from .process_panoptic import PanopticLabelGenerator + from .instance_augmentation import instance_augmentation + ++from logs import plog ++ + class SemKITTI(data.Dataset): + def __init__(self, data_path, imageset = 'train', return_ref = False, instance_pkl_path ='data'): + self.return_ref = return_ref +@@ -189,7 +191,7 @@ class voxel_dataset(data.Dataset): + cur_grid_size = self.grid_size + + intervals = crop_range/(cur_grid_size-1) +- if (intervals==0).any(): print("Zero interval!") ++ if (intervals==0).any(): plog.logger.info("Zero interval!") + + grid_ind = (np.floor((np.clip(xyz,min_bound,max_bound)-min_bound)/intervals)).astype(np.int) + +@@ -330,7 +332,7 @@ class spherical_dataset(data.Dataset): + cur_grid_size = self.grid_size + intervals = crop_range/(cur_grid_size-1) + +- if (intervals==0).any(): print("Zero interval!") ++ if (intervals==0).any(): plog.logger.info("Zero interval!") + grid_ind = (np.floor((np.clip(xyz_pol,min_bound,max_bound)-min_bound)/intervals)).astype(np.int) + + current_grid = grid_ind[:np.size(labels)] +diff --git a/instance_preprocess.py b/instance_preprocess.py +index cc76a3d..7f0488e 100644 +--- a/instance_preprocess.py ++++ b/instance_preprocess.py +@@ -2,6 +2,7 @@ + # -*- coding: utf-8 -*- + import argparse + from dataloader.dataset import SemKITTI ++from logs import plog + + if __name__ == '__main__': + # instance preprocessing +@@ -13,4 +14,4 @@ if __name__ == '__main__': + + train_pt_dataset = SemKITTI(args.data_path + '/sequences/', imageset = 'train', return_ref = True) + train_pt_dataset.save_instance(args.out_path) +- print('instance preprocessing finished.') +\ No newline at end of file ++ plog.logger.info('instance preprocessing finished.') +\ No newline at end of file +diff --git a/logs/plog.py b/logs/plog.py +new file mode 100644 +index 0000000..e3c9bef +--- /dev/null ++++ b/logs/plog.py +@@ -0,0 +1,22 @@ ++import logging ++ ++# 创建一个日志记录器，默认名称为'root' ++logger = logging.getLogger('my_logger') ++logger.setLevel(logging.DEBUG) # 设置日志记录器的日志级别为DEBUG，意味着DEBUG及以上级别的日志都会被记录 ++ ++# 创建一个handler，用于写入日志文件 ++fh = logging.FileHandler('Train_Panopric-PolarNet.log') ++fh.setLevel(logging.DEBUG) # 可以为handler单独设置级别 ++ ++# 再创建一个handler，用于将日志输出到控制台 ++ch = logging.StreamHandler() ++ch.setLevel(logging.DEBUG) # 控制台输出DEBUG及以上级别的日志 ++ ++# 定义handler的输出格式 ++formatter = logging.Formatter('[%(levelname)s][%(asctime)s][%(filename)s:%(lineno)d] %(funcName)s - %(message)s') ++fh.setFormatter(formatter) ++ch.setFormatter(formatter) ++ ++# 给日志记录器添加handler ++logger.addHandler(fh) ++logger.addHandler(ch) +diff --git a/network/instance_post_processing.py b/network/instance_post_processing.py +index 0346019..4e74fda 100755 +--- a/network/instance_post_processing.py ++++ b/network/instance_post_processing.py +@@ -2,7 +2,9 @@ + # -*- coding: utf-8 -*- + import torch + import torch.nn.functional as F +-import torch_scatter ++import mx_driving ++ ++from logs import plog + + + def find_instance_center(ctr_hmp, threshold=0.1, nms_kernel=5, top_k=None, polar=False): +@@ -153,7 +155,8 @@ def merge_semantic_and_instance(sem_seg, sem, ins_seg, label_divisor, thing_list + ins_seg = torch.unsqueeze(ins_seg,3).expand_as(sem_seg) + thing_mask = (ins_seg > 0) & semantic_thing_seg & thing_seg + if not torch.nonzero(thing_mask).size(0) == 0: +- sem_sum = torch_scatter.scatter_add(sem.permute(0,2,3,4,1)[thing_mask],ins_seg[thing_mask],dim=0) ++ # 选取特定区域特征相加 ++ sem_sum = mx_driving.scatter_add(sem.permute(0,2,3,4,1)[thing_mask], ins_seg[thing_mask].type(torch.int32), None, 0) + class_id = torch.argmax(sem_sum[:,:max(thing_list)],dim=1) + sem_seg[thing_mask] = (ins_seg[thing_mask] * label_divisor) + class_id[ins_seg[thing_mask]]+1 + else: +diff --git a/network/loss.py b/network/loss.py +index a114a94..a031ce9 100644 +--- a/network/loss.py ++++ b/network/loss.py +@@ -4,6 +4,7 @@ + import numpy as np + import torch + from .lovasz_losses import lovasz_softmax ++from logs import plog + + def _neg_loss(pred, gt): + ''' Modified focal loss. Exactly the same as CornerNet. +diff --git a/network/ptBEV.py b/network/ptBEV.py +index 936ed48..0a4e247 100644 +--- a/network/ptBEV.py ++++ b/network/ptBEV.py +@@ -6,7 +6,8 @@ import torch.nn.functional as F + import numpy as np + import numba as nb + import multiprocessing +-import torch_scatter ++from logs import plog ++import mx_driving + + class ptBEVnet(nn.Module): + +@@ -73,6 +74,8 @@ class ptBEVnet(nn.Module): + + cat_pt_fea = torch.cat(pt_fea,dim = 0) + cat_pt_ind = torch.cat(cat_pt_ind,dim = 0) ++ ++ cat_pt_ind = cat_pt_ind.type(torch.int32) + pt_num = cat_pt_ind.shape[0] + + # shuffle the data +@@ -81,10 +84,13 @@ class ptBEVnet(nn.Module): + cat_pt_ind = cat_pt_ind[shuffled_ind,:] + + # unique xy grid index +- unq, unq_inv, unq_cnt = torch.unique(cat_pt_ind,return_inverse=True, return_counts=True, dim=0) +- unq = unq.type(torch.int64) +- ++ unq, unq_inv, unq_cnt = torch.unique(cat_pt_ind.to('cpu'), return_inverse=True, return_counts=True, dim=0) ++ unq = unq.to('npu') ++ unq_inv = unq_inv.to('npu') ++ unq_cnt = unq_cnt.to('npu') ++ + # subsample pts ++ unq_inv = unq_inv.type(torch.int32) + if self.pt_selection == 'random': + grp_ind = grp_range_torch(unq_cnt,cur_dev)[torch.argsort(torch.argsort(unq_inv))] + remain_ind = grp_ind < self.max_pt +@@ -109,6 +115,7 @@ class ptBEVnet(nn.Module): + remain_ind[i_inds[FPS_results[count]]] = True + count += 1 + ++ + cat_pt_fea = cat_pt_fea[remain_ind,:] + cat_pt_ind = cat_pt_ind[remain_ind,:] + unq_inv = unq_inv[remain_ind] +@@ -119,7 +126,7 @@ class ptBEVnet(nn.Module): + processed_cat_pt_fea = self.PPmodel(cat_pt_fea) + + if self.pt_pooling == 'max': +- pooled_data = torch_scatter.scatter_max(processed_cat_pt_fea, unq_inv, dim=0)[0] ++ pooled_data = mx_driving.scatter_max(processed_cat_pt_fea.float(), unq_inv)[0] + else: raise NotImplementedError + + if self.fea_compre: +@@ -139,12 +146,13 @@ class ptBEVnet(nn.Module): + + # run through network + sem_prediction, center, offset = self.BEV_model(out_data) +- ++ + return sem_prediction, center, offset + + def grp_range_torch(a,dev): ++ a = a.type(torch.int32) + idx = torch.cumsum(a,0) +- id_arr = torch.ones(idx[-1],dtype = torch.int64,device=dev) ++ id_arr = torch.ones(idx[-1],dtype = torch.int32,device=dev) + id_arr[0] = 0 + id_arr[idx[:-1]] = -a[:-1]+1 + return torch.cumsum(id_arr,0) +@@ -183,4 +191,5 @@ def nb_greedy_FPS(xyz,K): + candidates_ind[next_ind] = True + remain_ind[next_ind] = False + +- return candidates_ind +\ No newline at end of file ++ return candidates_ind ++ +diff --git a/requirements.txt b/requirements.txt +index e19b023..6448a28 100644 +--- a/requirements.txt ++++ b/requirements.txt +@@ -1,9 +1,7 @@ +-numpy +-torch>=1.7.0 ++numpy==1.23.3 + tqdm + pyyaml + numba>=0.39.0 +-torch_scatter>=1.3.1 + Cython + scipy +-dropblock +\ No newline at end of file ++dropblock +diff --git a/test_pretrain.py b/test_pretrain.py +index 8ad1d63..9c5fcc1 100644 +--- a/test_pretrain.py ++++ b/test_pretrain.py +@@ -10,6 +10,7 @@ import torch + import torch.optim as optim + from tqdm import tqdm + import errno ++from logs import plog + + from network.BEV_Unet import BEV_Unet + from network.ptBEV import ptBEVnet +@@ -57,7 +58,7 @@ def main(args): + if os.path.exists(pretrained_model): + my_model.load_state_dict(torch.load(pretrained_model)) + pytorch_total_params = sum(p.numel() for p in my_model.parameters()) +- print('params: ',pytorch_total_params) ++ plog.logger.info('params: ',pytorch_total_params) + my_model.to(pytorch_device) + my_model.eval() + +@@ -82,9 +83,9 @@ def main(args): + num_workers = 4) + + # validation +- print('*'*80) +- print('Test network performance on validation split') +- print('*'*80) ++ plog.logger.info('*****************************************************') ++ plog.logger.info('Test network performance on validation split') ++ plog.logger.info('*****************************************************') + pbar = tqdm(total=len(val_dataset_loader)) + time_list = [] + pp_time_list = [] +@@ -129,21 +130,21 @@ def main(args): + + class_PQ, class_SQ, class_RQ, class_all_PQ, class_all_SQ, class_all_RQ = evaluator.getPQ() + miou,ious = evaluator.getSemIoU() +- print('Validation per class PQ, SQ, RQ and IoU: ') ++ plog.logger.info('Validation per class PQ, SQ, RQ and IoU: ') + for class_name, class_pq, class_sq, class_rq, class_iou in zip(unique_label_str,class_all_PQ[1:],class_all_SQ[1:],class_all_RQ[1:],ious[1:]): +- print('%15s : %6.2f%% %6.2f%% %6.2f%% %6.2f%%' % (class_name, class_pq*100, class_sq*100, class_rq*100, class_iou*100)) ++ plog.logger.info('%15s : %6.2f%% %6.2f%% %6.2f%% %6.2f%%' % (class_name, class_pq*100, class_sq*100, class_rq*100, class_iou*100)) + pbar.close() +- print('Current val PQ is %.3f' % ++ plog.logger.info('Current val PQ is %.3f' % + (class_PQ*100)) +- print('Current val miou is %.3f'% ++ plog.logger.info('Current val miou is %.3f'% + (miou*100)) +- print('Inference time per %d is %.4f seconds\n, postprocessing time is %.4f seconds per scan' % ++ plog.logger.info('Inference time per %d is %.4f seconds\n, postprocessing time is %.4f seconds per scan' % + (test_batch_size,np.mean(time_list),np.mean(pp_time_list))) + + # test +- print('*'*80) +- print('Generate predictions for test split') +- print('*'*80) ++ plog.logger.info('*'*80) ++ plog.logger.info('Generate predictions for test split') ++ plog.logger.info('*'*80) + pbar = tqdm(total=len(test_dataset_loader)) + with torch.no_grad(): + for i_iter_test,(test_vox_fea,_,_,_,test_grid,_,_,test_pt_fea,test_index) in enumerate(test_dataset_loader): +@@ -180,8 +181,8 @@ def main(args): + del test_pt_fea_ten,test_grid_ten,test_pt_fea,predict_labels,center,offset + pbar.update(1) + pbar.close() +- print('Predicted test labels are saved in %s. Need to be shifted to original label format before submitting to the Competition website.' % output_path) +- print('Remapping script can be found in semantic-kitti-api.') ++ plog.logger.info('Predicted test labels are saved in %s. Need to be shifted to original label format before submitting to the Competition website.' % output_path) ++ plog.logger.info('Remapping script can be found in semantic-kitti-api.') + + if __name__ == '__main__': + # Testing settings +@@ -195,6 +196,6 @@ if __name__ == '__main__': + new_args = yaml.safe_load(s) + args = merge_configs(args,new_args) + +- print(' '.join(sys.argv)) +- print(args) ++ plog.logger.info(' '.join(sys.argv)) ++ plog.logger.info(args) + main(args) +\ No newline at end of file +diff --git a/train.py b/train.py +index bcf0240..c177a83 100644 +--- a/train.py ++++ b/train.py +@@ -8,6 +8,9 @@ import yaml + import torch + import torch.optim as optim + from tqdm import tqdm ++from logs import plog ++import torch_npu ++from torch_npu.contrib import transfer_to_npu + + from network.BEV_Unet import BEV_Unet + from network.ptBEV import ptBEVnet +@@ -46,7 +49,7 @@ def main(args): + compression_model = args['dataset']['grid_size'][2] + grid_size = args['dataset']['grid_size'] + visibility = args['model']['visibility'] +- pytorch_device = torch.device('cuda:0') ++ pytorch_device = torch.device('npu:0') + if args['model']['polar']: + fea_dim = 9 + circular_padding = True +@@ -96,7 +99,9 @@ def main(args): + epoch=0 + best_val_PQ=0 + start_training=False ++ + my_model.train() ++ + global_iter = 0 + exce_counter = 0 + evaluator = PanopticEval(len(unique_label)+1, None, [0], min_points=50) +@@ -106,6 +111,11 @@ def main(args): + for i_iter,(train_vox_fea,train_label_tensor,train_gt_center,train_gt_offset,train_grid,_,_,train_pt_fea) in enumerate(train_dataset_loader): + # validation + if global_iter % check_iter == 0: ++ # 计算 FPS 并打印 ++ if pbar.n > 0: ++ average_time_per_iteration = pbar.format_dict["elapsed"] / pbar.n ++ FPS = train_batch_size * 1 / average_time_per_iteration ++ plog.logger.info(f"Final Performance images/sec (FPS) : {FPS}") + my_model.eval() + evaluator.reset() + with torch.no_grad(): +@@ -135,26 +145,28 @@ def main(args): + panoptic = panoptic_labels[0,val_grid[count][:,0],val_grid[count][:,1],val_grid[count][:,2]] + evaluator.addBatch(panoptic & 0xFFFF,panoptic,np.squeeze(val_pt_labels[count]),np.squeeze(val_pt_ints[count])) + del val_vox_label,val_pt_fea_ten,val_label_tensor,val_grid_ten,val_gt_center,val_gt_center_tensor,val_gt_offset,val_gt_offset_tensor,predict_labels,center,offset,panoptic_labels,center_points ++ + my_model.train() ++ + class_PQ, class_SQ, class_RQ, class_all_PQ, class_all_SQ, class_all_RQ = evaluator.getPQ() + miou,ious = evaluator.getSemIoU() +- print('Validation per class PQ, SQ, RQ and IoU: ') ++ plog.logger.info('Validation per class PQ, SQ, RQ and IoU: ') + for class_name, class_pq, class_sq, class_rq, class_iou in zip(unique_label_str,class_all_PQ[1:],class_all_SQ[1:],class_all_RQ[1:],ious[1:]): +- print('%15s : %6.2f%% %6.2f%% %6.2f%% %6.2f%%' % (class_name, class_pq*100, class_sq*100, class_rq*100, class_iou*100)) ++ plog.logger.info('%15s : %6.2f%% %6.2f%% %6.2f%% %6.2f%%' % (class_name, class_pq*100, class_sq*100, class_rq*100, class_iou*100)) + # save model if performance is improved + if best_val_PQ **说明：** + >该数据集的训练过程脚本只作为一种参考示例。 + +4. 请在 Panoptic-PolarNet 模型根目录，使用如下命令预处理 SemanticKITTI 数据集。 + ``` + # data_root 请替换成 SemanticKITTI 数据集 sequences 目录所在路径 + data_root=/home/datasets/SemanticKITTI/data/sequences + ln -nsf $data_root data/sequences + python instance_preprocess.py + ``` + 处理后的文件夹应该如下。 + ``` + ./ + ├── train.py + ├── ... + └── data/ + ├──instance_path.pkl + ├──sequences + ├── 00/ + │ ├── velodyne/ # Unzip from KITTI Odometry Benchmark Velodyne point clouds. + | | ├── 000000.bin + | | ├── 000001.bin + | | └── ... + │ ├── instance/ + | | ├── 000000_1.bin + | | ├── 000000_2.bin + | | └── ... + │ └── labels/ # Unzip from SemanticKITTI label data. + | ├── 000000.label + | ├── 000001.label + | └── ... + ├── ... + └── 21/ + └── ... + ``` + +# 开始训练 + +## 训练模型 + +1. 运行训练脚本。 + + 该模型仅支持单机单卡训练。 + + - 单机单卡训练 + + ``` + bash ./test/train.sh + ``` + + 训练完成后，权重文件保存在当前路径下，并输出模型训练精度和性能信息。 + +# 训练结果展示 + +**表 2** 训练结果展示表 + +| 芯片 | 卡数 | global batch size | epoch | best PQ | FPS | +| :-----------: | :--: | :---------------: | :---: | :------------: |--------------| +| 竞品A | 1p | 2 | 10 | 52.458 | 1.69 | +| Atlas 800T A2 | 1p | 2 | 10 | 52.159 | 1.28 | + +# 版本说明 + +## 变更 + +2025.03.05: 首次提交。 + +## FAQ + +无。 diff --git a/model_examples/Panoptic-PolarNet/test/env_npu.sh b/model_examples/Panoptic-PolarNet/test/env_npu.sh new file mode 100644 index 0000000000000000000000000000000000000000..4253cd0df2dd6496ead19286d5db280fd67dc68b --- /dev/null +++ b/model_examples/Panoptic-PolarNet/test/env_npu.sh @@ -0,0 +1,43 @@ +#!/bin/bash +msnpureport -g error -d 0 +msnpureport -g error -d 1 +msnpureport -g error -d 2 +msnpureport -g error -d 3 +msnpureport -g error -d 4 +msnpureport -g error -d 5 +msnpureport -g error -d 6 +msnpureport -g error -d 7 + +#将Host日志输出到串口,0-关闭/1-开启 +export ASCEND_SLOG_PRINT_TO_STDOUT=0 +#设置默认日志级别,0-debug/1-info/2-warning/3-error +export ASCEND_GLOBAL_LOG_LEVEL=3 +#设置Host侧Event日志开启标志,0-关闭/1-开启 +export ASCEND_GLOBAL_EVENT_ENABLE=0 +#设置是否开启taskque,0-关闭/1-开启/2-优化 +export TASK_QUEUE_ENABLE=2 +#开启内存池 +export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True +# 绑核 +export CPU_AFFINITY_CONF=1 +#设置hccl超时时间，单位：秒 +export HCCL_EXEC_TIMEOUT=8000 + +path_lib=$(python3 -c """ +import sys +import re +result='' +for index in range(len(sys.path)): + match_sit = re.search('-packages', sys.path[index]) + if match_sit is not None: + match_lib = re.search('lib', sys.path[index]) + + if match_lib is not None: + end=match_lib.span()[1] + result += sys.path[index][0:end] + ':' + + result+=sys.path[index] + '/torch/lib:' +print(result)""" +) + +echo ${path_lib} diff --git a/model_examples/Panoptic-PolarNet/test/train.sh b/model_examples/Panoptic-PolarNet/test/train.sh new file mode 100644 index 0000000000000000000000000000000000000000..2e1c4ae27e13c3cf5954621ff541cd270d64648f --- /dev/null +++ b/model_examples/Panoptic-PolarNet/test/train.sh @@ -0,0 +1,18 @@ +#!/bin/bash + +# test_path_dir 为包含 test 文件夹的路径 +cur_path=`pwd` +cur_path_last_diename=${cur_path##*/} +if [ x"${cur_path_last_diename}" == x"test" ];then + test_path_dir=${cur_path} + # 若当前在 test 目录下，cd 到 test 同级目录 + cd .. + cur_path=`pwd` +else + test_path_dir=${cur_path}/test +fi + +source ${test_path_dir}/env_npu.sh + +python train.py & +wait