We provide some tips for XRNeRF installation in this file.

• Installation
  • Requirements
  • Prepare environment
    • a. Install development libs.
    • b. Create a conda virtual environment and activate it.
    • c. Install PyTorch and torchvision
    • d. Install Other Needed Python Packages
    • e. Install Extensions
    • d. Download smpl_t_pose to surport GNR
  • Another option: Docker Image
    • a. Build an Image
    • b. Create a Container
  • Verification

• Linux
• Python 3.7
• PyTorch 1.10 (necessary)
• CUDA 11.0 (necessary)
• GCC 7.5
• build-essential: Install by apt-get install -y build-essential git ninja-build ffmpeg libsm6 libxext6 libgl1
• mmcv-full
• Numpy
• ffmpeg (4.2 is preferred)
• opencv-python 3 : Install by pip install opencv-python>=3
• imageio: Install by pip install imageio
• scikit-image: Install by pip install scikit-image
• lpips: Install by pip install lpips
• trimesh: Install by pip install trimesh
• smplx: Install by pip install smplx
• spconv: Install proper vision that matches your cuda-vision, for example pip install spconv-cu113
• pytorch3d: Install by pip install "git https://github.com/facebookresearch/pytorch3d.git@stable"

About hardware requirements: Instant-NGP need GPU-ARCH>=75, which means that at least a RTX 20X0 is required to have a full support.

If you don't need instant-ngp, spconv depends the minimum cuda version. So at least cuda 10.2 is needed.

sudo apt install libgl-dev freeglut3-dev build-essential git ninja-build ffmpeg libsm6 libxext6 libgl1

conda create -n xrnerf python=3.7 -y
conda activate xrnerf

1. check pytorch-cuda vision match table from here or here
2. find a proper torch vision (>=1.10.0 and match your cuda vision) from here, like cu111/torch-1.10.0+cu111-cp37-cp37m-linux_x86_64.whl, download the whl file
3. install your whl file, for example pip install torch-1.10.0 cu111-cp37-cp37m-linux_x86_64.whl
4. check here and install specified vision of torchvision, for example pip install torchvision==0.12.0

• you can use pip install requirements.txt to install most of the needed pkgs. If this step succeeds, you should jump to kilo-cuda and spconv step to install them manually. Or you can skip this step and follow the installation steps below
• pip install 'opencv-python>=3' yapf imageio scikit-image lpips trimesh smplx
• install mmcv-full following their Installation
• install spconv using pip install, for example pip install spconv-cu111. notice that only specified cuda-vision are supported, following their Installation
• install pytorch3d using pip install "git https://github.com/facebookresearch/pytorch3d.git@stable"
• install kilo-cuda following their Installation(optional, only needed for kilo-nerf)
• install tcnn using pip install git https://github.com/NVlabs/tiny-cuda-nn/#subdirectory=bindings/torch, or following their Installation(optional, only needed for instant-ngp)

• build cuda-extension raymarch for instant-ngp supported, following ngp_raymarch
• build cuda-extension mesh_grid for gnr supported, following mesh_grid

• In order to support the GNR algorithm, you need to download the smpl_t_pose folder from GNR, and modify basedata_cfg.t_pose_path in configs/gnr/gnr_genebody.py to the corresponding storage location

You need to set docker daemon, to enable docker-build's gpu support (for cuda extension install).

sudo apt-get install nvidia-container-runtime -f -y
sudo cp -f docker/daemon.json /etc/docker
sudo systemctl restart docker

See here for detail.

We provide a Dockerfile to build an image.

docker build -f ./docker/Dockerfile --rm -t xrnerf .

Important: Make sure you've installed the nvidia-container-toolkit.

Create a container with command:

docker run --gpus all -it xrnerf /bin/bash

Open a teiminal in your host computer, copy project into docker container

# d287273af72e is container id, using 'docker ps -a' to find id
docker cp ProjectPath/xrnerf d287273af72e:/workspace

To verify whether XRNeRF and the required environment are installed correctly, we can run unit-test python codes

coverage run --source xrnerf/models -m pytest -s test/models && coverage report -m

Notice that coverage and pytest need to be installed before that

pip install coverage pytest -i https://pypi.tuna.tsinghua.edu.cn/simple