文章目录
- 一.安装环境
- 二.运行
一.安装环境
nvcc-V conda create-n sparsedrive python=3.8-y conda activate sparsedrive pip install torch==2.0.0torchvision==0.15.1torchaudio==2.0.1--index-url https://download.pytorch.org/whl/cu118 （必须cuda11.6及以上，我用的cuda11.8）pip install mmcv-full==1.7.2-f https://download.openmmlab.com/mmcv/dist/cu118/torch2.0/index.htmlpip install-r requirement.txt or服务器的是cuda12.X conda create-n sparsedrive python=3.10-y conda activate sparsedrive pip install torch==2.1.0torchvision==0.16.0torchaudio==2.1.0--index-url https://download.pytorch.org/whl/cu121pip install mmcv-full==1.7.2-f https://download.openmmlab.com/mmcv/dist/cu121/torch2.1/index.html --only-binary mmcv-fullpip install-r requirement.txt（个别库版本升级了） （numpy==1.23.5mmcv_full==1.7.2mmdet==2.28.2urllib3==1.26.16pyquaternion==0.9.9nuscenes-devkit==1.1.10yapf==0.33.0tensorboard==2.14.0motmetrics==1.1.3pandas==1.5.3flash-attn==2.3.3opencv-python==4.8.1.78prettytable==3.7.0scikit-learn==1.3.0） 注：flash-attn==2.3.3，建议本地下载安装 wget https://github.com/Dao-AILab/flash-attention/releases/download/v2.3.3/flash_attn-2.3.3+cu122torch2.1cxx11abiFALSE-cp310-cp310-linux_x86_64.whlpip install flash_attn-2.3.3+cu122torch2.1cxx11abiFALSE-cp310-cp310-linux_x86_64.whl--no-build-isolation python-c"import flash_attn; print('安装成功！')"二.运行
1、编译可变形聚合CUDA算子
代码如下（示例）：
cd projects/mmdet3d_plugin/ops export CUDA_HOME=/usr/local/cuda python3 setup.py develop 出现Installed/home/1/SparseDrive-main/projects/mmdet3d_plugin/ops Processing dependenciesfordeformable-aggregation-ext==0.0.0Finished processing dependenciesfordeformable-aggregation-ext==0.0.0则安装成功 cd../../../2、配置nuscenes数据集
在主目录下创建data文件夹，在data创建nuscenes文件夹，可以从 https://www.nuscenes.org/nuscenes 下载数据集，放置下载的数据（实例只下载了v1.0-mini数据集，注意can_bus和maps需要下载全的，不仅仅只有mini内的数据，），对nuscenes中的v1.0-mini复制一份并改名为v1.0-trainval。
（注释：另感兴趣需要OCC数据对于Occupancy预测任务，还需要下载OCC的GT，可以从 https://github.com/CVPR2023-3D-Occupancy-Prediction/CVPR2023-3D-Occupancy-Prediction 下载数据集。 这里建议将下载好的数据集，解压到nuscense数据集文件夹中的 occ3d/gts 下，本文不涉及OCC。）maps文件内容
3、创建数据
sh scripts/create_data.sh在data/infos/文件夹下，生成mini文件夹，文件夹内生成以下两个pkl文件：4、通过K-means生成锚点
采用的是mini，所以修改/tools/kmeans/的代码路径：
#fp='data/infos/nuscenes_infos_train.pkl'fp='data/infos/mini/nuscenes_infos_train.pkl'再运行
sh scripts/kmeans.sh采用mini可正常生成det、map和motion.npy，但kmeans_plan_6.npy报错，对tools/kmeans/kmeans_plan.py代码进行修改，
# 原始的all数据集#clusters=[]#fortrajs in navi_trajs:#trajs=np.concatenate(trajs,axis=0).reshape(-1,12)#cluster=KMeans(n_clusters=K).fit(trajs).cluster_centers_#cluster=cluster.reshape(-1,6,2)#clusters.append(cluster)#forj inrange(K):#plt.scatter(cluster[j,:,0],cluster[j,:,1])#plt.savefig(f'vis/kmeans/plan_{K}',bbox_inches='tight')#plt.close()#clusters=np.stack(clusters,axis=0)#np.save(f'data/kmeans/kmeans_plan_{K}.npy',clusters)# 修改代码适配mini数据集 clusters=[]clusters.append(np.zeros((6,6,2)))fortrajs in navi_trajs[1:]:#fortrajs in navi_trajs:trajs=np.concatenate(trajs,axis=0).reshape(-1,12)cluster=KMeans(n_clusters=K).fit(trajs).cluster_centers_ cluster=cluster.reshape(-1,6,2)clusters.append(cluster)forj inrange(K):plt.scatter(cluster[j,:,0],cluster[j,:,1])plt.savefig(f'vis/kmeans/plan_{K}',bbox_inches='tight')plt.close()clusters=np.stack(clusters,axis=0)np.save(f'data/kmeans/kmeans_plan_{K}.npy',clusters)5、下载预训练权重
mkdir ckpt wget https://download.pytorch.org/models/resnet50-19c8e357.pth -O ckpt/resnet50-19c8e357.pth6、训练和测试（硬件版本太低无法训练）
#trainsh scripts/train.sh 报错如下 warnings.warn(usage:train.py[-h][--work-dir WORK_DIR][--resume-from RESUME_FROM][--no-validate][--gpus GPUS|--gpu-ids GPU_IDS[GPU_IDS...]][--seed SEED][--deterministic][--options OPTIONS[OPTIONS...]][--cfg-options CFG_OPTIONS[CFG_OPTIONS...]][--dist-url DIST_URL][--gpus-per-machine GPUS_PER_MACHINE][--launcher{none,pytorch,slurm,mpi,mpi_nccl}][--local_rank LOCAL_RANK][--autoscale-lr]config train.py:error:unrecognized arguments:--local-rank=0ERROR:torch.distributed.elastic.multiprocessing.api:failed(exitcode:2)local_rank:0(pid:15735)of binary:/home/1/anaconda3/envs/sparsedrive/bin/python3Traceback(most recent call last):File"/home/1/anaconda3/envs/sparsedrive/lib/python3.8/runpy.py",line194,in _run_module_as_mainreturn_run_code(code,main_globals,None,File"/home/1/anaconda3/envs/sparsedrive/lib/python3.8/runpy.py",line87,in _run_codeexec(code,run_globals)File"/home/1/anaconda3/envs/sparsedrive/lib/python3.8/site-packages/torch/distributed/launch.py",line196,in<module>main()File"/home/1/anaconda3/envs/sparsedrive/lib/python3.8/site-packages/torch/distributed/launch.py",line192,in mainlaunch(args)File"/home/1/anaconda3/envs/sparsedrive/lib/python3.8/site-packages/torch/distributed/launch.py",line177,in launchrun(args)File"/home/1/anaconda3/envs/sparsedrive/lib/python3.8/site-packages/torch/distributed/run.py",line785,in runelastic_launch(File"/home/1/anaconda3/envs/sparsedrive/lib/python3.8/site-packages/torch/distributed/launcher/api.py",line134,in __call__returnlaunch_agent(self._config,self._entrypoint,list(args))File"/home/1/anaconda3/envs/sparsedrive/lib/python3.8/site-packages/torch/distributed/launcher/api.py",line250,in launch_agent raiseChildFailedError(torch.distributed.elastic.multiprocessing.errors.ChildFailedError:============================================================这是因为torch2.0（1.8及以下不影响）不再使用torch.distributed.launch 启动，导致自动传入了 --local-rank 参数，而 train.py 无法解析。修改scripts/train.sh内容，启动方式为：
#!/bin/bash # 这是适配你需求的 stage1 训练脚本 # 核心：禁用分布式，用普通 python 命令启动，避免 local-rank 参数 export PYTHONPATH="$(dirname $0)/..":$PYTHONPATH # 定义训练配置和参数 CONFIG_FILE="projects/configs/sparsedrive_small_stage1.py"GPU_NUM=1# 单卡训练 DETERMINISTIC="--deterministic"WORK_DIR="./work_dirs/sparsedrive_stage1"# 可选：指定日志/模型保存目录 # 关键修改：用普通 python 启动，而非 dist_train.sh（避免分布式参数） python tools/train.py \ ${CONFIG_FILE}\--gpus ${GPU_NUM}\ ${DETERMINISTIC}\--launcher none \--work-dir ${WORK_DIR}# 可选，根据需要添加 # 定义训练配置和参数 CONFIG_FILE2="projects/configs/sparsedrive_small_stage2.py"GPU_NUM2=1# 单卡训练 DETERMINISTIC2="--deterministic"WORK_DIR2="./work_dirs/sparsedrive_stage2"# 可选：指定日志/模型保存目录 # 关键修改：用普通 python 启动，而非 dist_train.sh（避免分布式参数） python tools/train.py \ ${CONFIG_FILE2}\--gpus ${GPU_NUM2}\ ${DETERMINISTIC2}\--launcher none \--work-dir ${WORK_DIR2}# 可选，根据需要添加#testsh scripts/test.sh
