The following guidance works well for a machine with 3090 GPU and cuda 11.4, a machine with A100 GPU and cuda 11.7, and more machines.
If you encounter any problems, please feel free to open an issue.
conda remove -n gnfactor --all
conda create -n gnfactor python=3.9
conda activate gnfactor
conda install pytorch==1.10.0 torchvision torchaudio cudatoolkit=11.3 -c pytorch
conda install pytorch==1.10.0 torchvision torchaudio pytorch-cuda=11.7 -c pytorch -c nvidia
cd ..
git clone https://github.com/facebookresearch/pytorch3d.git
cd pytorch3d
conda install -c fvcore -c iopath -c conda-forge fvcore iopath
pip install -e .
cd ../GNFactor
cd ..
git clone https://github.com/openai/CLIP.git
cd CLIP
pip install -e .
cd ..
cd GNFactor
pip install open-clip-torch
wget https://www.coppeliarobotics.com/files/CoppeliaSim_Edu_V4_1_0_Ubuntu18_04.tar.xz --no-check-certificate
tar -xvf CoppeliaSim_Edu_V4_1_0_Ubuntu18_04.tar.xz
rm CoppeliaSim_Edu_V4_1_0_Ubuntu18_04.tar.xz
Remember to source your bashrc (source ~/.bashrc) and reopen a new terminal then.
You should replace the path here with your own path to the coppeliasim installation directory.
export COPPELIASIM_ROOT=EDIT/ME/PATH/TO/COPPELIASIM/INSTALL/DIR
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$COPPELIASIM_ROOT
export QT_QPA_PLATFORM_PLUGIN_PATH=$COPPELIASIM_ROOT
You should open a new terminal here, to make your .bashrc work.
cd third_party/PyRep
pip install -r requirements.txt
pip install .
cd ../..
cd third_party/RLBench
pip install -r requirements.txt
python setup.py develop
cd ../..
cd third_party/YARR
pip install -r requirements.txt
python setup.py develop
cd ../..
cd GNFactor
pip install -r requirements.txt
python setup.py develop
cd ..
pip install packaging==21.3 dotmap pyhocon wandb chardet opencv-python-headless gpustat ipdb visdom sentencepiece
Install xformers (this version is a must to avoid errors from detectron2)
pip install xformers==0.0.18
Install Stable Diffusion
pip install stable-diffusion-sdkit==2.1.3
Install detectron2:
cd ..
git clone https://github.com/facebookresearch/detectron2.git
cd detectron2
pip install -e .
cd ../GNFactor
Install ODISE packages
cd third_party/ODISE
pip install -e .
cd ..
Since a lot of packages are installed, there are some possible bugs. Use these commands first before running the code.
pip install torchvision --upgrade
pip install hydra-core==1.1
pip install opencv-python-headless
pip install numpy==1.23.5
Now, you should be able to run our training and evaluation scripts.
Please make sure you could both train and evaluate algorithms before conducting more experiments.
For possible errors, see ERROR_CATCH.md. Don't hesitate to open an issue if you encounter any hard problems.