MapTR代码复现-nuscenes数据集

本文链接：https://blog.csdn.net/weixin_54678439/article/details/139604515

前言

本节将对环视车道线MapTR算法进行复现，使用nuscenes-mini数据集！

一、环境配置

1、基础环境：

ubuntu20.04，pytorch1.10.0，python3.8，cuda11.3

2、源码下载

下载地址：

git clone https://github.com/hustvl/MapTR.git

3、环境安装

# 1 下载torch
pip install torch==1.10.0+cu113 torchvision==0.11.0+cu113 torchaudio==0.10.0 -f https://download.pytorch.org/whl/torch_stable.html


pip install mmcv-full==1.4.0 mmdet==2.14.0 mmsegmentation==0.14.1 timm==0.9.5

# 2 安装mmdetection3d
cd mmdetection3d
export CUDA_HOME=/usr/local/cuda-11.3
pip install -v e .

# 3 其他功能包
cd ./
pip install -r requirement.txt 

# 4 安装maptr相关的插件
cd ./projects/mmdet3d_plugin/maptr/modules/ops/geometric_kernel_attn
python setup.py build install
'''
 终端显示：Processing dependencies for GeometricKernelAttention==1.0
          Finished processing dependencies for GeometricKernelAttention==1.0 
'''

# 5 更新networkx,numpy,setuptools版本，2.2在maptr-v2运行时会报错
pip install networkx==2.3 numpy==1.22.2 setuptools==58.2.0

# 6 spconv-cu113可以不安装, 本人没有安装
pip install spconv-cu113 -i https://pypi.tuna.tsinghua.edu.cn/simple

（1）在安装pip install mmcv-full==1.4.0出现以下报错

报错：ap/mmcv-full_aab08f9dc5e94beea857df9c1022f8b4/mmcv/ops/csrc/common -I/root/miniconda3/lib/python3.8/site-packages/torch/include -I/root/miniconda3/lib/python3.8/site-packages/torch/include/torch/csrc/api/include -I/root/miniconda3/lib/python3.8/site-packages/torch/include/TH -I/root/miniconda3/lib/python3.8/site-packages/torch/include/THC -I/root/miniconda3/include/python3.8 -c ./mmcv/ops/csrc/pytorch/assign_score_withk.cpp -o build/temp.linux-x86_64-3.8/./mmcv/ops/csrc/pytorch/assign_score_withk.o -std=c++14 -DTORCH_API_INCLUDE_EXTENSION_H -DPYBIND11_COMPILER_TYPE="_gcc" -DPYBIND11_STDLIB="_libstdcpp" -DPYBIND11_BUILD_ABI="_cxxabi1011" -DTORCH_EXTENSION_NAME=_ext -D_GLIBCXX_USE_CXX11_ABI=0
cc1plus: warning: command line option ‘-Wstrict-prototypes’ is valid for C/ObjC but not for C++
gcc: fatal error: Killed signal terminated program cc1plus
compilation terminated.
error: command 'gcc' failed with exit status 1
----------------------------------------
ERROR: Command errored out with exit status 1: /root/miniconda3/bin/python -u -c 'import io, os, sys, setuptools, tokenize; sys.argv[0] = '"'"'/tmp/pip-install-p5i_7fap/mmcv-full_aab08f9dc5e94beea857df9c1022f8b4/setup.py'"'"'; __file__='"'"'/tmp/pip-install-p5i_7fap/mmcv-full_aab08f9dc5e94beea857df9c1022f8b4/setup.py'"'"';f = getattr(tokenize, '"'"'open'"'"', open)(__file__) if os.path.exists(__file__) else io.StringIO('"'"'from setuptools import setup; setup()'"'"');code = f.read().replace('"'"'\r\n'"'"', '"'"'\n'"'"');f.close();exec(compile(code, __file__, '"'"'exec'"'"'))' install --record /tmp/pip-record-mnfvty4w/install-record.txt --single-version-externally-managed --compile --install-headers /root/miniconda3/include/python3.8/mmcv-full Check the logs for full command output.

解决：

可以把pip安装换成mim安装，先来安装mim

pip install -U openmim
pip install mmcv-full==1.4.0

解决：

4、查看版本号

（1）mmlab相关

pip list | grep mm

（ 2）torch相关

pip list | grep torch

二、制作数据集

1、下载can bus

下载地址：nuscenes.org/download

2020 年 2 月，官方发布了 CAN 总线扩展。它包含有关车辆路线、IMU、姿势、转向角反馈、电池、制动器、档位、信号、车轮速度、油门、扭矩、太阳能传感器、里程计等的低级车辆数据。点击US下载：

2、下载nuscenes-mini

文件目录结构如下：

data
├── can_bus
└── nuscenes
    ├── maps
    ├── nuscenes_gt_database
    ├── samples
    ├── sweeps
    └── v1.0-mini

3、生成训练数据集

python tools/create_data.py nuscenes --root-path ./data/nuscenes --out-dir ./data/nuscenes --extra-tag nuscenes --version v1.0-mini --canbus ./data

报错：ModuleNotFoundError: No module named 'tools'

Traceback (most recent call last):
File "tools/create_data.py", line 10, in <module>
from data_converter import indoor_converter as indoor
File "/root/MapTR/tools/data_converter/indoor_converter.py", line 6, in <module>
from tools.data_converter.s3dis_data_utils import S3DISData, S3DISSegData
ModuleNotFoundError: No module named 'tools'

解决：在终端执行

export PYTHONPATH=./

成功运行！

三、训练train

（1）以下两种权重文件皆可使用

mkdir ckpts
cd ckpts
# 1 下载resnet50预训练权重
wget https://download.pytorch.org/models/resnet50-19c8e357.pth

# 2 下载resnet18预训练权重
wget https://download.pytorch.org/models/resnet18-f37072fd.pth

（2）修改projects/configs/maptr/maptr_nano_r18_110e.py配置文件中相关参数

# 1. 预训练权重
pretrained=dict(img='ckpts/resnet18-f37072fd.pth'),

# 2. BN层修改 单卡BN，多卡用SyncBN
# norm_cfg=dict(type='SyncBN', requires_grad=True), # 多卡BN层
norm_cfg=dict(type='BN', requires_grad=True),    # 单卡BN层

# 3. 数据集位置
data_root = 'data/nuscenes/'

# 4. samples_per_gpu和workers_per_gpu
samples_per_gpu=8,
workers_per_gpu=4,

# 训练周期
total_epochs = 100

# 5训练  单卡训练代码最后跟个`1`，八卡最后跟个`8`
./tools/dist_train.sh ./projects/configs/maptr/maptr_nano_r18_110e.py 1

（3）运行train指令

chmod +x ./tools/dist_train.sh
./tools/dist_train.sh ./projects/configs/maptr/maptr_nano_r18_110e.py 1

报错:

TypeError: FormatCode() got an unexpected keyword argument 'verify'
ERROR:torch.distributed.elastic.multiprocessing.api:failed (exitcode: 1) local_rank: 0 (pid: 1401) of binary: /root/miniconda3/bin/python3

解决：

pip install yapf==0.40.1

开始运行：

四、预测

第1步：下载官方的权重文件

将权重文件maptr_tiny_r50_24e.pth放入ckpts目录下，下载地址：Google 云端硬盘 - 病毒扫描警告

第2步：修改配置文件

修改maptr_tiny_r50_24e.py中的_base_参数custom_nus-3d.py与default_runtime.py文件的路径，格式如下：

_base_ = [
    '../projects/configs/datasets/custom_nus-3d.py',
    '../projects/configs/_base_/default_runtime.py'
]

第3步：运行

将MapTR/projects/configs/maptr/maptr_tiny_r50_24e.py文件放入config文件夹下

python tools/maptr/vis_pred.py config/maptr_tiny_r50_24e.py ckpts/maptr_tiny_r50_24e.pth  --show-dir ./vis_dirs

结果保存在./vis_dirs目录下

第4步：生成可视化视频

python tools/maptr/generate_video.py ./vis_dirs

生成视频如下：

MapTR-nuscenes

MapTR代码复现实践

完结撒花~