Skip to content

README.md

Latest commit

 

History

History
131 lines (103 loc) · 5.97 KB

README.md

File metadata and controls

131 lines (103 loc) · 5.97 KB

OpenSceneFlow Assets

There are two ways to setup the environment: conda in your desktop and docker container isolate environment.

Docker Environment

Build Docker Image

If you want to build docker with compile all things inside, there are some things need setup first in your own desktop environment:

  • NVIDIA-driver: which I believe most of people may already have it. Try nvidia-smi to check if you have it.
  • Docker: 
    # Add Docker's official GPG key:
sudo apt-get update
sudo apt-get install ca-certificates curl
sudo install -m 0755 -d /etc/apt/keyrings
sudo curl -fsSL https://download.docker.com/linux/ubuntu/gpg -o /etc/apt/keyrings/docker.asc
sudo chmod a+r /etc/apt/keyrings/docker.asc

# Add the repository to Apt sources:
echo \
"deb [arch=$(dpkg --print-architecture) signed-by=/etc/apt/keyrings/docker.asc] https://download.docker.com/linux/ubuntu \
$(. /etc/os-release && echo "$VERSION_CODENAME") stable" | \
sudo tee /etc/apt/sources.list.d/docker.list > /dev/null
sudo apt-get update
  • nvidia-container-toolkit 
    sudo apt update && apt install nvidia-container-toolkit

Then follow this stackoverflow answers:

  1. Edit/create the /etc/docker/daemon.json with content:

    {
   "runtimes": {
      "nvidia": {
            "path": "/usr/bin/nvidia-container-runtime",
            "runtimeArgs": []
         } 
   },
   "default-runtime": "nvidia" 
}

  2. Restart docker daemon:

    sudo systemctl restart docker

  3. Then you can build the docker image:

    cd OpenSceneFlow && docker build -f Dockerfile -t zhangkin/opensf .

To Apptainer container

If you want to build a minimal training env for Apptainer container, you can use the following command:

apptainer build opensf.sif assets/opensf.def
# zhangkin/opensf:full is created by Dockerfile

Then run as a Python env with:

PYTHON="apptainer run --nv --writable-tmpfs opensf.sif"
$PYTHON train.py

Installation

We will use conda to manage the environment with mamba for faster package installation.

System

Install conda with mamba for package management and for faster package installation:

curl -L -O "https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-$(uname)-$(uname -m).sh"
bash Miniforge3-$(uname)-$(uname -m).sh

Environment

Create base env: [5~15 minutes based on your network speed and cpu]

git clone https://github.com/KTH-RPL/OpenSceneFlow.git
cd OpenSceneFlow
mamba env create -f assets/environment.yml

Checking important packages in our environment now:

mamba activate opensf
python -c "import torch; print(torch.__version__); print(torch.cuda.is_available()); print(torch.version.cuda)"
python -c "import lightning.pytorch as pl; print('pl version:', pl.__version__)"
python -c "import spconv.pytorch as spconv; print('spconv import successfully')"
python -c "from assets.cuda.mmcv import Voxelization, DynamicScatter;print('successfully import on our lite mmcv package')"
python -c "from assets.cuda.chamfer3D import nnChamferDis;print('successfully import on our chamfer3D package')"
python -c "from av2.utils.io import read_feather; print('av2 package ok') "

Other issues

  1. ImportError: libtorch_cuda.so: undefined symbol: cudaGraphInstantiateWithFlags, version libcudart.so.11.0 The cuda version: pytorch::pytorch-cuda and nvidia::cudatoolkit need be same. Reference link

  2. In cluster have error: pandas ImportError: /lib64/libstdc++.so.6: version 'GLIBCXX_3.4.29' not found Solved by export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/proj/berzelius-2023-154/users/x_qinzh/mambaforge/lib

  3. nvidia channel cannot put into env.yaml file otherwise, the cuda-toolkit will always be the latest one, for me (2025-04-30) I struggling on an hour and get nvcc -V also 12.8 at that time. py=3.10 for cuda >=12.1. (seems it's nvidia cannot be in the channel list???); py<3.10 for cuda <=11.8.0: otherwise 10x, 20x series GPU won't work on cuda compiler. (half precision)

  4. torch_scatter problem: OSError: /home/kin/mambaforge/envs/opensf-v2/lib/python3.10/site-packages/torch_scatter/_version_cpu.so: undefined symbol: _ZN5torch3jit17parseSchemaOrNameERKNSt7__cxx1112basic_stringIcSt11char_traitsIcESaIcEEE Solved by install the torch-cuda version: pip install https://data.pyg.org/whl/torch-2.0.0%2Bcu118/torch_scatter-2.1.2%2Bpt20cu118-cp310-cp310-linux_x86_64.whl

  5. cuda package problem: ValueError(f"Unknown CUDA arch ({arch}) or GPU not supported") Solved by checking GPU compute then manually assign: export TORCH_CUDA_ARCH_LIST=8.6