测试环境
- OS：Ubuntu 24.04.1 LTS
- 显卡：NVIDIA GeForce RTX 4090
- CUDA Version：12.4
安装conda环境
python版本是3.8，PyTorch版本2.1.0，对应CUDA版本用的是11.8。(不要用仓库中说的PyTorch 1.12.0，在运行mamba的时候torch会找不到某个模块，应该是torch 2.x才加进去的)
conda create -n wfmamba python=3.8
conda activate wfmamba
conda install pytorch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 pytorch-cuda=11.8 -c pytorch -c nvidia
clone github仓库
git clone https://github.com/gwenzhang/Voxel-Mamba.git
安装spconv
pip install spconv-cu118
安装torch_scatter
从网上找到对应CUDA和Python版本的torch_scatter的whl文件，可能发布了多个版本号，选最新的就可以。我用的是torch_scatter-2.1.2+pt21cu118-cp38-cp38-linux_x86_64.whl。下载，然后到下载到的路径下，使用pip安装：pip install torch_scatter-2.1.2+pt21cu118-cp38-cp38-linux_x86_64.whl
安装OpenPCDet
到Voxel-Mamba的主目录下，运行setup.py：python setup.py develop
出现报错：error: Couldn't find a setup script in /tmp/easy_install-hh_h_cgj/scikit_image-0.25.0rc1.tar.gz
先下载scikit-image（指定阿里源，我用清华源下得很慢）：pip install scikit-image -i https://mirrors.aliyun.com/pypi/simple
又报错：
ERROR: pip’s dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
pcdet 0.6.0+ad8172a requires easydict, which is not installed.
pcdet 0.6.0+ad8172a requires llvmlite, which is not installed.
pcdet 0.6.0+ad8172a requires numba, which is not installed.
pcdet 0.6.0+ad8172a requires tensorboardX, which is not installed.
依次安装上面这些库：pip install easydict llvmlite numba tensorboardX，之后再安装scikit-image，显示已经安装。
之后再运行：python setup.py develop，显示：
Using /home/icdm/anaconda/envs/wfmamba/lib/python3.8/site-packages
Finished processing dependencies for pcdet==0.6.0+ad8172a`
说明OpenPCDet安装成功了。
安装Mamba
进入到mamba的目录下：cd mamba
安装：pip install -e .
到此，Voxel Mamba就配置好了。在实际跑代码的时候可能会出现缺一些库的情况，再pip装上就好了。