环境配置基础步骤

创建conda环境并安装PyTorch

conda create --name bevfusion python=3.8 -y
conda activate bevfusion
conda install pytorch==2.1.0 torchvision==0.16.0 pytorch-cuda=12.1 -c pytorch -c nvidia -y

验证CUDA可用性

import torch
print(torch.cuda.is_available())  # 预期输出True

mmcv-full安装方法

使用预编译wheel文件安装

pip install https://download.openmmlab.com/mmcv/dist/cu121/torch2.1/mmcv_full-1.7.2-cp38-cp38-win_amd64.whl
python -c "import mmcv; print(mmcv.__version__)"  # 预期输出1.7.2

其他依赖安装

安装必要组件

pip install mmdet==2.28.2 mmsegmentation==0.30.0 timm yapf einops torchpack

CUDA Toolkit安装

下载地址

https://developer.nvidia.com/cuda-downloads

环境变量配置要求

• CUDA_PATH设置为C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.6
• Path中添加%CUDA_PATH%\bin并置于首位

项目编译方法

使用VS 2022命令提示符执行

cd C:\项目路径\bevfusion-of-mit
conda activate bevfusion
python setup.py develop

数据集处理流程

解压脚本示例

import tarfile, re
from pathlib import Path
from tqdm import tqdm

p = Path(r"D:\BaiduNetdiskDownload")
tars = sorted(p.glob("v1.0-trainval*_blobs.tar"), 
             key=lambda x: int(re.search(r'trainval(\d+)_blobs', x.name).group(1)))

for tar in tqdm(tars, desc="解压进度"):
    with tarfile.open(tar) as t:
        t.extractall(path=p)
    tar.unlink()

生成数据索引命令

python tools/create_data.py nuscenes --root-path ./data/nuscenes --out-dir ./data/nuscenes --extra-tag nuscenes

模型测试验证

下载预训练模型

mkdir pretrained
cd pretrained
wget https://bevfusion.mit.edu/downloads/pretrained/bevfusion-det.pth
wget https://bevfusion.mit.edu/downloads/pretrained/bevfusion-seg.pth

执行测试命令

python tools/test.py configs/nuscenes/det/transfusion/secfpn/camera+lidar/swint_v0p075/convfuser.yaml pretrained/bevfusion-det.pth --eval bbox

目录结构规范

标准项目结构

bevfusion-of-mit/
├── data/nuscenes/
│   ├── samples/
│   ├── sweeps/
│   ├── v1.0-trainval/
│   ├── maps/
│   ├── nuscenes_infos_train.pkl
│   └── nuscenes_infos_val.pkl
├── pretrained/
│   ├── bevfusion-det.pth
│   └── bevfusion-seg.pth