PERT代码复现

环境

• pycharm 2023.1.2
• wsl + ubuntu20.04
• anaconda3

创建环境

# 创建环境
conda create -n petr python=3.8 -y
conda activate petr

添加依赖

# 添加依赖包
pip install torch==1.10.0+cu113 torchvision==0.11.0+cu113 -f https://download.pytorch.org/whl/torch_stable.html
pip install mmcv-full==1.4.0 -f https://download.openmmlab.com/mmcv/dist/cu113/torch1.10.0/index.html
pip install mmdet==2.14.0
pip install mmsegmentation==0.14.1
pip install ninja
pip install tensorboard==2.13.0
pip install nuscenes-devkit==1.1.10
pip install scikit-image==0.19.0
pip install lyft-dataset-sdk==0.0.8
pip install numpy==1.19.5
pip install pandas==1.4.4
pip install llvmlite==0.31.0
pip install timm
pip install setuptools==59.5.0
pip install einops

# 安装 mmdetection3d
git clone https://github.com/open-mmlab/mmdetection3d.git
cd mmdetection3d
git checkout v0.17.1
python setup.py install

添加数据集

用的 windows 系统，创建链接。数据集中未包含 test 数据集，若需要请自行配置。

单独拷贝文件夹 v1.0-trainval

mklink /j D:\edusoftware\coding\PETR\data\nuscenes\maps E:\datasets\nuscenes\maps
mklink /j D:\edusoftware\coding\PETR\data\nuscenes\samples E:\datasets\nuscenes\samples
mklink /j D:\edusoftware\coding\PETR\data\nuscenes\sweeps E:\datasets\nuscenes\sweeps

目录完整结构如下：

下载预训练模型

github 下载速度慢的话，可从 百度网盘 下载。

cd PETR
mkdir ckpts

cd ckpts & wget https://download.openmmlab.com/pretrain/third_party/resnet50_msra-5891d200.pth

生成数据集信息

生成数据集信息过慢，可从 百度网盘 下载。

# 使用mmdetection3d生成数据集信息
# https://github.com/open-mmlab/mmdetection3d/blob/1.0/docs/en/data_preparation.md
python tools/create_data.py nuscenes --root-path ./data/nuscenes --out-dir ./data/nuscenes --extra-tag nuscenes

ModuleNotFoundError: No module named ‘tools’

export PYTHONPATH=./

ModuleNotFoundError: No module named ‘einops’

pip install einops

FileNotFoundError: [Errno 2] No such file or directory: ‘/data/Dataset/nuScenes/nuscenes_infos_train.pkl’

# petr_r50dcn_gridmask_p4.py 修改地址
data_root = './data/nuScenes/'

断点调试

projects/configs/petr/petr_r50dcn_gridmask_p4.py --work-dir work_dirs/petr_r50dcn_gridmask_p4/