MobileNeRF于2023年提出,源码地址:https://github.com/google-research/jax3d/tree/main/jax3d/projects/mobilenerf ,论文为:《MobileNeRF: Exploiting the Polygon Rasterization Pipeline for Efficient Neural Field Rendering on Mobile Architectures》,一种基于纹理多边形的新NeRF表示,可以通过标准渲染pipeline合成图像。
运行mobilenerf需依赖JAX,这里介绍下在Windows上通过Anaconda配置MobileNeRF的步骤:
1.安装cuda 11.1 + cudnn 8.2.0
(1).从 https://developer.nvidia.com/cuda-11.1.0-download-archive?target_os=Windows&target_arch=x86_64&target_version=10&target_type=exelocal 下载cuda_11.1.0_456.43_win10.exe并安装;
(2).从 https://developer.nvidia.com/rdp/cudnn-archive 下载cudnn-11.3-windows-x64-v8.2.0.53.zip并解压缩:
1).在C:\Program Files\NVIDIA GPU Computing Toolkit目录下,新建CUDNN\8.2.0目录即C:\Program Files\NVIDIA GPU Computing Toolkit\CUDNN\8.2.0;
2).将解压缩后产生的三个文件夹bin, include, lib拷贝到C:\Program Files\NVIDIA GPU Computing Toolkit\CUDNN\8.2.0目录下;
3).将C:\Program Files\NVIDIA GPU Computing Toolkit\CUDNN\8.2.0\bin添加到系统环境变量
2.创建虚拟环境mobilenerf,依次执行如下命令:
conda create -n mobilenerf python=3.8
conda activate mobilenerf
3.clone jax3d:
git clone https://github.com/google-research/jax3d
4.从 https://github.com/cloudhan/jax-windows-builder 下载jaxlib-0.1.76+cuda11.cudnn82-cp38-none-win_amd64.whl,并将此whl拷贝到jax3d/projects/mobilenerf目录下
5.将Anaconda Powershell Prompt定位到jax3d/projects/mobilenerf目录下,依次执行如下命令:
pip install jaxlib-0.1.76+cuda11.cudnn82-cp38-none-win_amd64.whl
pip install opencv-python==4.4.0.46
pip install tqdm==4.62.3
pip install flax==0.3.0
#pip uninstall jax
pip install jax==0.2.28
注意各模块版本,否则会有各种冲突,各个版本如下:
执行mobilenerf要求有8块显卡,否则会报ptxas error: please verify that sufficient filesystem space is provided